Microwave Imaging and Electromagnetic Inverse Scattering Problems

Microwave
Imaging and
Electromagnetic
Inverse Scattering
Problems
Edited by
Loreto Di Donato and Andrea F. Morabito
Printed Edition of the Special Issue Published in Journal of Imaging
www.mdpi.com/journal/jimaging
Microwave Imaging and Electromagnetic
Inverse Scattering Problems
Microwave Imaging and Electromagnetic
Inverse Scattering Problems
Special Issue Editors

Loreto Di Donato
Andrea F. Morabito
MDPI • Basel • Beijing • Wuhan • Barcelona • Belgrade

Loreto Di Donato
University of Catania
Italy
Andrea F. Morabito
Università degli Studi Mediterranea di Reggio Calabria
Italy
Editorial Office
MDPI
St. Alban-Anlage 66
4052 Basel, Switzerland
This is a reprint of articles from the Special Issue published online in the open access
journal Journal of Imaging (ISSN 2313-433X) in 2019 (available at: https://www.mdpi.com/journal/
jimaging/special issues/Microwave Imaging).
For citation purposes, cite each article independently as indicated on the article page online and as
indicated below:
LastName, A.A.; LastName, B.B.; LastName, C.C. Article Title. Journal Name Year, Article Number,
Page Range.
ISBN 978-3-03921-950-6 (Pbk)

ISBN 978-3-03921-951-3 (PDF)

c 2020 by the authors. Articles in this book are Open Access and distributed under the Creative
Commons Attribution (CC BY) license, which allows users to download, copy and build upon
published articles, as long as the author and publisher are properly credited, which ensures maximum
dissemination and a wider impact of our publications.
The book as a whole is distributed by MDPI under the terms and conditions of the Creative Commons
license CC BY-NC-ND.
Contents
About the Special Issue Editors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii
Preface to ”Microwave Imaging and Electromagnetic Inverse Scattering Problems” . . . . . . ix
Alessandro Fedeli, Manuela Maffongelli, Ricardo Monleone, Claudio Pagnamenta,

Matteo Pastorino, Samuel Poretti, Andrea Randazzo and Andrea Salvadè
A Tomograph Prototype for Quantitative Microwave Imaging: Preliminary
Experimental Results
Reprinted from: J. Imaging 2018, 4, 139, doi:10.3390/jimaging4120139 . . . . . . . . . . . . . . . . 1
Yu Zhong and Kuiwen Xu

Contraction Integral Equation for Three-Dimensional Electromagnetic Inverse Scattering
Problems
Reprinted from: J. Imaging 2019, 5, 27, doi:10.3390/jimaging5020027 . . . . . . . . . . . . . . . . . 10
Marco Salucci, Lorenzo Poli and Giacomo Oliveri

Full-Vectorial 3D Microwave Imaging of Sparse Scatterers through a Multi-Task Bayesian
Compressive Sensing Approach
Tushar Rajvanshi, Angela Dell’Aversano, and Raffaele Solimene

Image-Based RCS Estimation from Near-Field Data
Marco Donald Migliore, Fulvio Schettino, Daniele Pinchera, Mario Lucido and
Gaetano Panariello
A Minimum Rank Approach for Reduction of Environmental Noise in Near-Field Array
Antenna Diagnosis
Martina T. Bevacqua and Roberta Palmeri

Qualitative Methods for the Inverse Obstacle Problem: A Comparison on Experimental Data
Loreto Di Donato, David Mascali, Andrea F. Morabito and Gino Sorbello

A Finite-Difference Approach for Plasma Microwave Imaging Profilometry
Reijer Leijsen, Patrick Fuchs, Wyger Brink, Andrew Webb, Rob Remis
Developments in Electrical-Property Tomography Based on the Contrast-Source Inversion
Method
Cameron Kaye, Ian Jeffrey and Joe LoVetri

Novel Stopping Criteria for Optimization-Based Microwave Breast Imaging Algorithms
Declan O’Loughlin, Bárbara L. Oliveira, Martin Glavin, Edward Jones and

Martin O’Halloran
Comparing Radar-Based Breast Imaging Algorithm Performance with Realistic
Patient-SpecificPermittivity Estimation
v
About the Special Issue Editors
Loreto Di Donato received his B.S. and M.S. Laurea degrees in Biomedical Engineering from the
University of Naples “Federico II” (Italy) in 2006 and 2008, respectively. After a year spent at
the Electromagnetic Diagnostic Research Group of the Institute for the Electromagnetic Sensing
of the Environment—National Research Council (IREA-CNR) of Italy in Naples as a research
assistant, he received his Ph.D. in Information Engineering from the University “Mediterranea” of
Reggio Calabria in 2013. He has been with the Department of Electrical, Electronics and Computer
Engineering (DIEEI) as Assistant Professor of Electromagnetic Fields at the University of Catania
(Italy) since his appointment in 2013, and starting from 2021, he shall serve as Associate Professor
here. His research activities are mainly focused on inverse problems in electromagnetics, with
particular interest in inverse scattering problems and microwave imaging for biomedical, subsurface,
and plasma diagnostics, as well as focusing and antenna synthesis problems. Loreto Di Donato was
a Young Scientist Awarded at the XXX URSI General Assembly in 2011, received the prize “Gaetano
Latmiral” from the Electromagnetic Italian Society in 2016 and obtained the Abilitazione Scientifica
Nazionale (ASN) for the academic position Associate Professor in 2017.
Andrea F. Morabito received his Laurea degree in Telecommunications Engineering (summa cum
laude) and Ph.D. degree in Computer, Biomedical, and Telecommunications Engineering from the
University of Reggio Calabria, Italy, where he has been Assistant Professor in Electromagnetic Fields
since 2010. His research work is mainly focused on models and effective strategies for electromagnetic
forward and inverse scattering problems, as well as on antenna theory, design, optimal synthesis and
therapeutic applications, with applications ranging from biomedical imaging to radar and satellite
telecommunications. Dr. Morabito has been a visiting researcher at the Eledia Research Center of the
University of Trento and is a member of the LEMMA Research Group of the Italian Electromagnetics
Society, and of the European Association on Antennas and Propagation. He is author of
about 100 papers published on peer-reviewed scientific journals or international conference
proceedings. Dr. Morabito has been awarded by the Italian Electromagnetics Society with both the
“Barzilai” and “Latmiral” Prizes, and received the Abilitazione Scientifica Nazionale (ASN) for the
academic position Associate Professor.
vii
Preface to ”Microwave Imaging and Electromagnetic
Inverse Scattering Problems”
Microwave imaging (MWI) is a promising technology able to perform noninvasive diagnostics in
many disparate fields, such as nondestructive testing evaluation, civil and geophysical prospecting,
surveillance and security monitoring, biomedical imaging, and many others. The main advantages
of MWI lies in its noninvasiveness, low cost, as well as the on-field portability of the involved
equipment. For this reason, over the last thirty years, a large effort has been spent in studying
and understanding the ultimate achievable performance of MWI, and hence to develop several
reconstruction strategies able to overcome the main drawbacks concerning ill-posedness, false
solutions, resolution limits, and heavy computational burden.
Nowadays, a plethora of approaches for MWI co-exist in the scientific literature and an
ever-increasing number of experimental tests and facilities, such as clinical trials, are under
development. This notwithstanding, microwave imaging and inverse scattering problems still
deserve great efforts from a methodological point of view, in order to develop effective and efficient
solutions able to provide reconstructions as reliable, accurate, and fast as possible. Indeed, despite the
intrinsic low resolution of microwave diffraction tomography, emerging signal processing techniques
and innovative hardware solutions, such as, for example, compressive sensing and deep/machine
learning, metamaterials and metasurfaces, millimeter and THz waves, as well as the employment of
high-performing computers, may allow these huge drawbacks to be overcome, paving the way for
the full advantages of MWI to be expressed in the above-cited fields of application.
This Special Issue presents ten papers by some of the most prominent international scientists
and researchers in the area of microwave imaging and diagnostics. The topics include experimental,
methodological, and emerging fields of application. The presented papers provide a glimpse into the
current state of the art in microwave imaging and electromagnetic inverse scattering and the different
nature of the involved contexts and applications.
Loreto Di Donato, Andrea F. Morabito

ix
Journal of
Imaging
Article
A Tomograph Prototype for Quantitative Microwave
Imaging: Preliminary Experimental Results
Alessandro Fedeli 1 , Manuela Maffongelli 2 , Ricardo Monleone 2 , Claudio Pagnamenta 2 ,
Matteo Pastorino 1, *, Samuel Poretti 2 , Andrea Randazzo 1 and Andrea Salvadè 2
1 Department of Electrical, Electronic, Telecommunications Engineering, and Naval Architecture,
University of Genoa, 16145 Genoa, Italy; [email protected] (A.F.); [email protected] (A.R.)
2 Department of Technology and Innovation, University of Applied Sciences of Southern Switzerland,
6928 Manno, Switzerland; [email protected] (M.M.); [email protected] (R.M.);
[email protected] (C.P.); [email protected] (S.P.); [email protected] (A.S.)
* Correspondence: [email protected]
Received: 1 October 2018; Accepted: 22 November 2018; Published: 26 November 2018
Abstract: A new prototype of a tomographic system for microwave imaging is presented in this paper.
The target being tested is surrounded by an ad-hoc 3D-printed structure, which supports sixteen custom
antenna elements. The transmission measurements between each pair of antennas are acquired through
a vector network analyzer connected to a modular switching matrix. The collected data are inverted
by a hybrid nonlinear procedure combining qualitative and quantitative reconstruction algorithms.
Preliminary experimental results, showing the capabilities of the developed system, are reported.
Keywords: microwave imaging; tomography; inverse problems
1. Introduction
In the last years, there has been growing interest in the development of microwave imaging
systems in various applicative fields, such as for civil, industrial, and biomedical applications [1–4].
In fact, microwaves have the potential ability to penetrate into dielectric materials, allowing for
directly obtaining information about the internal dielectric properties of the samples under test (SUT).
Despite the potential advantages, there are still some problems to be faced, which motivate the
ever-expanding research activities on this topic. First, in many cases, the targets are composed
of inhomogeneous, lossy, and dispersive media (e.g., when dealing with biological materials).
The electromagnetic waves are thus typically subject to high attenuations during the propagation
through the target, and the information about the material properties are contained in a complex
way in the scattered field. Moreover, the dielectric contrast with respect to the external embedding
may be high, requiring the adoption of matching layers. Consequently, it is necessary to properly
design the antennas, the embedding structures, and the transmission/measurement hardware.
Several experimental apparatuses have been proposed, based on both radar concepts and inverse
scattering methods (see, for example, [5–7]). A recent overview containing several examples of
experimental set ups has been reported in the literature [8], and the reader is referred to that book for a
rather detailed description of these configurations.
Furthermore, appropriate inversion procedures, able to exploit the available a-priori information
about the target, should be used in order to process the measurements made available by the hardware.
In this framework, several interesting approaches have been recently introduced [9–18]. On the one
hand, they have been designed in order to improve the capabilities of the so-called qualitative methods,
which are aimed at retrieving only the presence/shape of targets, as well as some particular features of
interest. For example, in non-destructive evaluations, the knowledge of the position and dimension
of a defect or a crack could in fact represent sufficient information. On the other hand, much more
J. Imaging 2018, 4, 139; doi:10.3390/jimaging4120139 1 www.mdpi.com/journal/jimaging

J. Imaging 2018, 4, 139
sophisticated approaches have been devised for inspecting strong scatterers and producing images
of the dielectric properties of the scene under test. This is of fundamental importance in biomedical
applications, where microwave imaging techniques have been proposed for breast cancer [19–22] and
brain stroke detection [23–26], as well as in several other diagnostic processes, such as the monitoring
of microwave ablation [27], guiding endo-capsules [28], and so on.
In this paper, a new prototype of the microwave imaging system is presented. It allows for collecting
frequency-stepped scattered-field data in a tomographic configuration by using a multi-static setup.
The antennas have been specifically designed in order to minimize the reflections from the external
surface, and to optimize the propagation inside the target. An ad-hoc 3D printed holding structure is used
to keep the radiating elements in contact with the target. The measured data are inverted using a hybrid
inverse-scattering procedure [29], which combines a delay-and-sum (DAS) qualitative algorithm [30]
(providing a qualitative image of the dielectric discontinuities) with a quantitative method based on
an inexact-Newton/Landweber (INLW) scheme [31] (providing a reconstruction of the distribution
of the dielectric properties of the target). In particular, the obtained qualitative image allows for
focusing the INLW method on the regions of the target in which the discontinuities with respect to the
background are found. The capabilities of the developed system are assessed by means of preliminary
experimental results.
The paper is organized as follows. The developed measurement hardware and the related inversion
procedure are described in Section 2. Section 3 reports some of the experimental results aimed at
validating both the system and the reconstruction algorithm. Finally, conclusions are drawn in Section 4.
2. Measurement Setup and Inversion Procedure

A block diagram of the developed prototype is shown in Figure 1. It is composed of a vector
network analyzer (VNA), a radio frequency (RF) switch matrix connected to a set of ad-hoc antennas
(M = 16 antennas are adopted in the current prototype), and a control board used to drive the switches.
The whole system is managed by a personal computer (PC), which is used to select the active antenna pair
(via universal serial bus (USB) connection), control the VNA, collect the measurements (via an Ethernet
connection), and execute the reconstruction algorithms. The PC also takes care of the synchronization
between the various components of the system.
Figure 1. Block diagram of the microwave tomographic system architecture. USB—universal

serial bus; μC—microcontroller; VNA—vector network analyzer; PC—personal computer;
TX—transmission; RX—reception.
The RF switch matrix is composed of two levels, as shown in Figure 2. The first one, which is directly
connected to the transmit and receive ports of the VNA, is composed of two 1:4 switch boards, whereas the
second one is composed of four 2:4 switch boards, whose outputs are connected to the 16 antennas.
For both of the levels, Peregrine Semiconductor’s PE42441 SP4T absorptive RF switches have been used
in the switch boards. The working frequency band of such components ranges from 10 MHz up to 8 GHz.
Each board has been manufactured using a Rogers RO4350B substrate (characterized by complex relative
2
J. Imaging 2018, 4, 139
dielectric permittivity r = 3.66 − j0.011 at 2.5 GHz) in order to reduce the RF signal attenuation and
the changes in the electrical parameters. The switching subsystem allows for independently selecting
a pair of antennas among the possible 240 combinations. One antenna at a time is used in transmit
mode (i.e., it is connected to the transmitting (TX) port of the VNA), whereas the remaining ones are
sequentially connected to the receiving port of the VNA for measuring the field scattered by the target.
In order to reduce the interferences in the measurement, all of the inactive antennas are terminated
using absorptive RF switch components, ensuring an isolation between the adjacent channels at the
board connectors of about 90 dB at 3 GHz. The switch control board is equipped with an ATmega328
microcontroller programmed to transparently pass the USB data to an I2C bus, which is connected to
four 16-bit PCF8575CDBE4 input/output (I/O) expanders used to drive the RF switches.
Folded quasi self-complementary antennas (FQSCA) [32] have been used in the prototype
(whose layout is represented in Figure 3). Such antennas have been specifically developed in order
to work in direct contact with the SUT, by reducing the reflections due to the mismatch with the target
interface and optimizing the radiation inside the inspected structure. The operating bandwidth of the
antennas ranges from 1.5 GHz up to 6 GHz. The radiating elements are kept in contact with the SUT
using a custom 3D printed circular holding structure (shown in Figure 3) made of plastic (VeroWhite),
with relative dielectric permittivity r = 2.98 and loss tangent tan δ = 0.029 at 2.4 GHz [33]. The holding
structure has been produced using a stereolithographical process. The separation between the two
adjacent antennas is 22.5◦ and the radius is 60 mm. In order to compensate for the different paths of
the signal in the function of the selected antenna pairs, a calibration is performed. In particular, the
calibration procedure is aimed at compensating the influence (i.e., phase shift and attenuation) of the
switching electronics and cables for each pair of transmitting/receiving (TX/RX) antennas, thus resulting
in measurements referenced to the antenna connectors. Actually, such a calibration is performed by
subtracting from the amplitude (in decibel) and phase of the measured data (with and without inclusions)
of each TX/RX pair the corresponding through measurements obtained by connecting together the RX
and TX antenna connectors.
Figure 2. Block scheme of the switching board and photograph of the 2:4 switch module.
3
J. Imaging 2018, 4, 139
Figure 3. Schematic representation of the holding structure and layout of the folded quasi
self-complementary antennas (FQSCA) antenna.
The scattered-field data (i.e., the difference between the measurements with and without the
inclusions) collected by the system are processed using a hybrid qualitative–quantitative inversion
scheme [29]. A flow chart of the developed inversion procedure is shown in Figure 4. Starting from
the measured stepped-frequency data, an initial qualitative reconstruction Λ(r), r ∈ D (D being
the inspected area), is obtained using a multistatic DAS method [30]. The time-domain response is
synthesized by applying an inverse fast Fourier transform to the data. Such an image provides a rough
indication of the eventually present inclusions.
Subsequently, a frequency-hopping (FH) inexact-Newton/Landweber method is used to obtain
the quantitative distribution of the dielectric properties. In particular, for each considered frequency
(hereafter, the superscript f denotes the frequency index), a hybrid INLW method is used to invert the
f f
non-linear and ill-posed scattering equation relating the unknown contrast function χ f = r /r,b − 1
f f
(r and r,b being the complex relative dielectric permittivity of SUT and background, respectively)
f
with the z-component of the scattered electric field es , as follows:
−1
f f f f
es (r) = Gd χ f I − Gs χ f e b (r) = F f χ f (r), (1)
f f
where eb is the z-component of the electric field due to the background and Gd/s are linear integral
operators whose kernel is the Green’s function for the background. In such an inversion procedure, the
antennas have been modeled as ideal z-directed line-current sources. Consequently, the acquired data
have been further pre-processed by properly scaling the measured values by a complex scaling factor,
obtained by matching the data measured with the empty imaging chamber (filled by the background
medium only) with the corresponding simulated electric field.
The INLW technique, outlined in the right side of Figure 4, is an iterative nonlinear inversion
method formed by two nested loops. The outer one is based on a Newton scheme, which performs an
iterative linearization of Equation (1) around the currently reconstructed value of the contrast
function

f f f f
χi (i being the outer iteration index). The resulting linearized equation Fi hi (r) = es (r) − F f χi (r),
f
Fi being the Fréchet derivative of the scattering operator F f in Equation (1), is solved in a regularized
sense, using a truncated Landweber algorithm [31,34]. In particular, the regularized solution hi (r) is
obtained by means of the following update formula (initialized with hi,0 = 0):

f∗ f f f
hi,l +1 (r) = hi,l (r) − β i Fi Fi hi,l − es (r) + F f χi (r) , l = 0, 1, . . . , NLW (2)
f∗
where Fi is the adjoint of the operator Fi , βi = Fi −2 is a fixed relaxation coefficient, and NLW
f f
is the maximum number of allowed iterations. In order to exploit the information obtained in the
qualitative step, the Newton update has been modified by multiplying the solution of the linearized
4
J. Imaging 2018, 4, 139
problem by the value of the normalized qualitative map Λ in the same point of the investigation domain
f f
(i.e., χi+1 (r) = χi (r) + Λ(r)hi (r)). This is the key aspect of the combination between the qualitative DAS
method and the INLW algorithm. In fact, since Λ(r) ∈ [0, 1], ∀r ∈ D, the contrast function χ is updated
“faster” in the regions, in which the qualitative method found relevant discontinuities with respect to
the background (i.e., when Λ is close to 1), and “slower” when Λ assumes low values, which means in
the points outside the qualitatively-detected targets. Newton iterations (i.e., linearization, regularized
solution, and update) then continue until a convergence criterion is met or a maximum number of
iterations, NIN , is reached. After that, the FH procedure requires that the obtained contrast function
is used as the starting guess for the inversion at the subsequent frequency, until all of the available
frequencies have been processed.
Figure 4. Flow chart of the inversion algorithm. DAS—delay-and-sum; FH-INLW—frequency-hopping

inexact-Newton/Landweber.
3. Preliminary Experimental Results

The proposed system has been preliminarily tested with a SUT composed of a Plexiglas circular
cylinder, with a diameter of d = 120 mm (Figure 5) filled with a mixture of oil, water, salt, and soy
lecithin (relative dielectric permittivity r,b ≈ 3 − j0.73 at 2.45 GHz). The liquid mixture, which occupies
a vertical depth of 90 mm, has been prepared by adding small quantities of the single components to the
mixture step-by-step, and checking the actual properties with a VNA dielectric probe (Keysight N1501A).
In particular, a strong percentage of vegetable oil has been used, whereas other components, in smaller
quantities, have been added to adjust the values to the desired ones. Salt has also been added in order
to achieve the correct conductivity. Moreover, soy lecithin has been used as a surfactant for obtaining
a stable mixture. The investigation domain, D, is a circular region of the diameter, d, discretized into
1240 square subdomains with a 3.12 mm side. Three plastic tubes with a diameter of di = 5 mm,
filled with a different liquid mixture (relative dielectric permittivity r ≈ 35 − j10 in the considered
frequency band), are approximately located at positions r1 = (−20, −25) mm, r2 = (20, −20) mm,
and r3 = (25, 20) mm. In particular, a 50%/50% mixture of vegetable oil and tap water with 2 g of soy
lecithin has been used. A negligible dispersive behavior of the measured dielectric properties of both of
the liquid mixtures has been observed in the considered frequency range. Therefore, dispersion has not
been considered in the inversion procedure.
5
J. Imaging 2018, 4, 139
Figure 5. Photograph of the experimental set up.
The transmission scattering parameters have been acquired for all of the possible antenna pairs by
considering 585 frequency steps in the range between 1 and 5 GHz. The acquisition time is mainly affected
by the VNA measurement sweep time; the switch setting and I/O expander latency can be neglected.
For the presented results, a sweep time of 1 s has been set for each of the antennas’ combinations,
resulting in a total measurement time of 240 s.
The measured values of S21 with and without the considered inclusions are shown in Figure 6 for a
pair of opposite antennas (antennas #6 and #14). It is worth noting that most of the scattering contributions
are located between 2 and 4 GHz. Consequently, the FH-INLW method has been applied, by considering
the F = 13 equally-spaced frequencies in this range. The maximum number of inner and outer iterations
in each INLW inversion step are NLW = 100 and NIN = 100, and the loops are stopped when the variation
in the residual falls below 1%. The values of such parameters have been empirically selected on the basis
of previous analyses. The results provided by the developed procedure are shown in Figures 7 and 8.
In particular, the obtained qualitative map, showing the presence of the three inclusions, is reported in
Figure 7. The quantitative reconstructions of the relative dielectric permittivity at the final frequency step
are shown in Figure 8a (real part) and Figure 8b (imaginary part). As can be seen, the targets are accurately
localized and the dielectric properties are quite correctly estimated. The root mean squared error (RMSE)
on the reconstruction of the complex dielectric permittivity is equal to 0.48 in this case. The proposed
hybrid inversion algorithm has also been compared with a bare INLW technique, in which the update
of the solution in the Newton iterations is not weighted by the qualitative map [31]. The reconstructed
distributions of the dielectric properties of the target are shown in Figure 9, and the corresponding RMSE
is 0.52. This comparison evidences the effectiveness of the hybridization with the qualitative scheme,
which allows for a significant improvement in both the estimation of the complex dielectric permittivity
(especially, the real part) and the size of the target, as well as the reconstruction of the background medium.
(a) (b)
Figure 6. (a) Magnitude and (b) phase of the measured S21 for a couple of facing antennas with and
without the targets.
6
J. Imaging 2018, 4, 139

yP

Λ

xP
Figure 7. Experimental results. Qualitative indicator map.

yP
yP

εr¶¶

εr¶

xP xP
(a) (b)
Figure 8. Experimental results. Reconstructed distributions of the (a) real part and (b) imaginary part
of the complex relative dielectric permittivity at 3.8 GHz obtained with the proposed hybrid algorithm.

yP
yP

ε¶¶

ε¶r

xP
xP
(a) (b)
Figure 9. Experimental results. Reconstructed distributions of the (a) real part and (b) imaginary
part of the complex relative dielectric permittivity at 3.8 GHz obtained with a bare quantitative
inversion procedure.
4. Conclusions
A microwave system prototype for quantitative tomographic imaging has been presented in
this paper. The system is composed of a set of ad-hoc antennas held in contact with the SUT by a
custom 3D-printed structure, a modular switch matrix, a VNA, and a computer-based control and
7
J. Imaging 2018, 4, 139
processing section. A hybrid qualitative/quantitative inversion method has been developed for
processing the measured scattered-field data. The effectiveness of the imaging setup has been assessed
by means of preliminary experimental results. Future developments will be aimed at performing a
wider experimental validation campaign with more complex targets.
Author Contributions: Conceptualization, A.F., M.M., R.M., C.P., M.P., S.P., A.R., and A.S.; formal analysis, A.F.,
M.R., and A.R.; investigation, M.M., R.M., C.P., S.P., and A.S.; resources, M.M., R.M., C.P., S.P., and A.S.; software,
A.F., M.R., and A.R.
Funding: This research received no external funding.
Conflicts of Interest: The authors declare no conflict of interest.
References
1. Meaney, P.M.; Goodwin, D.; Golnabi, A.H.; Zhou, T.; Pallone, M.; Geimer, S.D.; Burke, G.; Paulsen, K.D.
Clinical Microwave Tomographic Imaging of the Calcaneus: A First-in-Human Case Study of Two Subjects.
IEEE Trans. Biomed. Eng. 2012, 59, 3304–3313. [CrossRef] [PubMed]
2. Ghasr, M.T.; Horst, M.J.; Dvorsky, M.R.; Zoughi, R. Wideband microwave camera for real-time 3-D imaging.
IEEE Trans. Antennas Propag. 2017, 65, 258–268. [CrossRef]
3. Gilmore, C.; Zakaria, A.; Pistorius, S.; LoVetri, J. Microwave Imaging of Human Forearms: Pilot Study and
Image Enhancement. Int. J. Biomed. Imaging 2013, 2013, 673027. [CrossRef] [PubMed]
4. Benedetto, A.; Pajewski, L. Civil Engineering Applications of Ground Penetrating Radar; Springer: Cham,
Switzerland, 2015; ISBN 978-3-319-04813-0.
5. Castorina, G.; Di Donato, L.; Morabito, A.F.; Isernia, T.; Sorbello, G. Analysis and design of a concrete embedded
antenna for wireless monitoring applications. IEEE Antennas Propag. Mag. 2016, 58, 76–93. [CrossRef]
6. Monleone, R.D.; Pastorino, M.; Fortuny-Guasch, J.; Salvade, A.; Bartesaghi, T.; Bozza, G.; Maffongelli, M.;
Massimini, A.; Carbonetti, A.; Randazzo, A. Impact of background noise on dielectric reconstructions obtained by
a prototype of microwave axial tomograph. IEEE Trans. Instrum. Meas. 2012, 61, 140–148. [CrossRef]
7. Fear, E.C.; Bourqui, J.; Curtis, C.; Mew, D.; Docktor, B.; Romano, C. Microwave breast imaging with a monostatic
radar-based system: A study of application to patients. IEEE Trans. Microw. Theory Tech. 2013, 61, 2119–2128.
[CrossRef]
8. Pastorino, M.; Randazzo, A. Microwave Imaging Methods and Applications; Artech House: Boston, MA, USA, 2018;
ISBN 978-1-63081-348-2.
9. Palmeri, R.; Bevacqua, M.T.; Crocco, L.; Isernia, T.; Di Donato, L. Microwave imaging via distorted iterated
virtual experiments. IEEE Trans. Antennas Propag. 2017, 65, 829–838. [CrossRef]
10. Solimene, R.; Buonanno, A.; Pierri, R. Imaging small PEC spheres by a linear δ approach. IEEE Trans. Geosci.
Remote Sens. 2008, 46, 3010–3018. [CrossRef]
11. Gennarelli, G.; Vivone, G.; Braca, P.; Soldovieri, F.; Amin, M.G. Multiple extended target tracking for
through-wall radars. IEEE Trans. Geosci. Remote Sens. 2015, 53, 6482–6494. [CrossRef]
12. Brancaccio, A.; Leone, G. Multimonostatic shape reconstruction of two-dimensional dielectric cylinders by a
Kirchhoff-based approach. IEEE Trans. Geosci. Remote Sens. 2010, 48, 3152–3161. [CrossRef]
13. Anselmi, N.; Salucci, M.; Oliveri, G.; Massa, A. Wavelet-Based Compressive Imaging of Sparse Targets.
14. Oliveri, G.; Randazzo, A.; Pastorino, M.; Massa, A. Electromagnetic imaging within the contrast-source
formulation by means of the multiscaling inexact Newton method. J. Opt. Soc. Am. A 2012, 29, 945–958.
[CrossRef] [PubMed]
15. D’Urso, M.; Isernia, T.; Morabito, A.F. On the solution of 2-D inverse scattering problems via source-type
integral equations. IEEE Trans. Geosci. Remote Sens. 2010, 48, 1186–1198. [CrossRef]
16. Shumakov, D.S.; Nikolova, N.K. Fast quantitative microwave imaging with scattered-power maps. IEEE Trans.
Microw. Theory Tech. 2018, 66, 439–449. [CrossRef]
17. Desmal, A.; Bağcı, H. Shrinkage-thresholding enhanced Born iterative method for solving 2D inverse
electromagnetic scattering problem. IEEE Trans. Antennas Propag. 2014, 62, 3878–3884. [CrossRef]
8
J. Imaging 2018, 4, 139
18. Maaref, N.; Millot, P.; Ferrières, X.; Pichot, C.; Picon, O. Electromagnetic Imaging Method Based on Time
Reversal Processing Applied to Through-the-Wall Target Localization. Prog. Electromagn. Res. 2008, 1, 59–67.
[CrossRef]
19. Bellizzi, G.; Bucci, O.M.; Catapano, I. Microwave cancer imaging exploiting magnetic nanoparticles as
contrast agent. IEEE Trans. Biomed. Eng. 2011, 58, 2528–2536. [CrossRef] [PubMed]
20. Catapano, I.; Di Donato, L.; Crocco, L.; Bucci, O.M.; Morabito, A.F.; Isernia, T.; Massa, R. On quantitative
microwave tomography of female breast. Prog. Electromagn. Res. 2009, 97, 75–93. [CrossRef]
21. Miao, Z.; Kosmas, P. Multiple-frequency DBIM-TwIST algorithm for microwave breast imaging. IEEE Trans.
Antennas Propag. 2017, 65, 2507–2516. [CrossRef]
22. Scapaticci, R.; Kosmas, P.; Crocco, L. Wavelet-based regularization for robust microwave imaging in
medical applications. IEEE Trans. Biomed. Eng. 2015, 62, 1195–1202. [CrossRef] [PubMed]
23. Hopfer, M.; Planas, R.; Hamidipour, A.; Henriksson, T.; Semenov, S. Electromagnetic tomography for
detection, differentiation, and monitoring of brain stroke: A virtual data and human head phantom study.
IEEE Antennas Propag. Mag. 2017, 59, 86–97. [CrossRef]
24. Tournier, P.-H.; Bonazzoli, M.; Dolean, V.; Rapetti, F.; Hecht, F.; Nataf, F.; Aliferis, I.; El Kanfoud, I.;
Migliaccio, C.; de Buhan, M.; et al. Numerical modeling and high-speed parallel computing: New
perspectives on tomographic microwave imaging for brain stroke detection and monitoring. IEEE Antennas
Propag. Mag. 2017, 59, 98–110. [CrossRef]
25. Bisio, I.; Estatico, C.; Fedeli, A.; Lavagetto, F.; Pastorino, M.; Randazzo, A.; Sciarrone, A. Brain stroke
microwave imaging by means of a Newton-conjugate-gradient method in Lp Banach spaces. IEEE Trans.
Microw. Theory Tech. 2018, 66, 3668–3682. [CrossRef]
26. Ireland, D.; Bialkowski, K.; Abbosh, A. Microwave imaging for brain stroke detection using Born iterative method.
Antennas Propag. IET Microw. 2013, 7, 909–915. [CrossRef]
27. Bellizzi, G.G.; Crocco, L.; Cavagnaro, M.; Farina, L.; Lopresto, V.; Scapaticci, R. A full-wave numerical
assessment of microwave tomography for monitoring cancer ablation. In Proceedings of the 11th European
Conference on Antennas and Propagation (EUCAP), Paris, France, 19–24 March 2017; pp. 3722–3725.
28. Chandra, R.; Johansson, A.J.; Gustafsson, M.; Tufvesson, F. A microwave imaging-based technique to localize an
in-body RF source for biomedical applications. IEEE Trans. Biomed. Eng. 2015, 62, 1231–1241. [CrossRef] [PubMed]
29. Boero, F.; Fedeli, A.; Lanini, M.; Maffongelli, M.; Monleone, R.; Pastorino, M.; Randazzo, A.; Salvadè, A.;
Sansalone, A. Microwave tomography for the inspection of wood materials: Imaging system and
experimental results. IEEE Trans. Microw. Theory Tech. 2018, 66, 3497–3510. [CrossRef]
30. Li, X.; Hagness, S.C. A confocal microwave imaging algorithm for breast cancer detection. IEEE Microw.
Compon. Lett. 2001, 11, 130–132. [CrossRef]
31. Bozza, G.; Estatico, C.; Pastorino, M.; Randazzo, A. An inexact Newton method for microwave reconstruction
of strong scatterers. IEEE Antennas Wirel. Propag. Lett. 2006, 5, 61–64. [CrossRef]
32. Lanini, M.; Poretti, S.; Salvade, A.; Monleone, R. Design of a slim wideband-antenna to overcome the strong
reflection of the air-to-sample interface in microwave imaging. In Proceedings of the 2015 International
Conference on Electromagnetics in Advanced Applications, Turin, Italy, 7–11 September 2015; pp. 1020–1023.
33. Whittow, W.G. 3D printing, inkjet printing and embroidery techniques for wearable antennas. In Proceedings of
the 10th European Conference on Antennas and Propagation (EuCAP2016), Davos, Switzerland, 10–15 April 2016;
pp. 1–4.
34. Bertero, M.; Boccacci, P. Introduction to Inverse Problems in Imaging; Institute of Physics Pub: Bristol, UK;
Philadelphia, PA, USA, 1998; ISBN 978-0-7503-0439-9.
© 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).
9
Journal of
Imaging
Article
Contraction Integral Equation for Three-Dimensional
Electromagnetic Inverse Scattering Problems
Yu Zhong 1, * and Kuiwen Xu 2
1 Institute of High Performance Computing, Agency for Science, Technology and Research (A*STAR),
Singapore 138632, Singapore
2 Key Lab of RF Circuits and Systems of Ministry of Education, Hangzhou Dianzi University,
Hangzhou 310018, China; [email protected]
Received: 31 December 2018; Accepted: 31 January 2019; Published: 8 February 2019
Abstract: Inverse scattering problems (ISPs) stand at the center of many important imaging
applications, such as geophysical explorations, industrial non-destructive testing, bio-medical
imaging, etc. Recently, a new type of contraction integral equation for inversion (CIE-I) has been
proposed to tackle the two-dimensional electromagnetic ISPs, in which the usually employed
Lippmann–Schwinger integral equation (LSIE) is transformed into a new form with a modified
medium contrast via a contraction mapping. With the CIE-I, the multiple scattering effects,
i.e., the physical reason for the nonlinearity in the ISPs, is substantially suppressed in estimating
the modified contrast, without compromising physical modeling. In this paper, we firstly propose
to implement this new CIE-I for the three-dimensional ISPs. With the help of the FFT type twofold
subspace-based optimization method (TSOM), when handling the highly nonlinear problems with
strong scatterers, those with higher contrast and/or larger dimensions (in terms of wavelengths),
the performance of the inversions with CIE-I is much better than the ones with the LSIE, wherein
inversions usually converge to local minima that may be far away from the solution. In addition,
when handling the moderate scatterers (those the LSIE modeling can still handle), the convergence
speed of the proposed method with CIE-I is much faster than the one with the LSIE. Secondly,
we propose to relax the contraction mapping condition, i.e., different contraction mappings are used
in updating contrast sources and contrast, and we find that the convergence can be further accelerated.
Several numerical tests illustrate the aforementioned interests.
Keywords: inverse scattering; nonlinear problem; contraction integral equation for inversion
(CIE-I); imaging
1. Introduction
Inverse scattering problems (ISPs) in electromagnetics and acoustics are of great interest in
industries due to important imaging applications in various areas, such as geophysical survey,
non-destructive testing, ground-penetrating radar, bio-medical imaging, etc., as solving the ISPs
provides rich information about the unknown targets, such as locations, shapes, and material
distributions within some structures [1–4]. For instance, microwave imaging has been used to inspect
the abnormalities in human bodies like bleeding in the head and tumours in breasts. On the other
hand, they are also quite important due to the representative difficulties in solving a large group
of inverse problems concerning waves and fields, and thus researchers in mathematics, physics,
and engineering societies have devoted great efforts to improving efficiencies and accuracies of the
numerical solvers [5]. As shown in Figure 1, solving ISPs is to determine the unknown scatterers
within a domain, when the domain is illuminated by several different incidences and we can measure
the scattered fields outside the domain (usually contaminated by noise) for each incidence. It is well

J. Imaging 2019, 5, 27
known that the main difficulties in such ISPs are their two intrinsic properties, i.e., ill-posedness and
nonlinearity [1]. As most of the practical problems are three-dimensional (3-D) ones, which require
much more computational resources than those in two-dimensional cases, they are often more difficult
to solve due to the data deficiency (measurement aperture usually covers a small solid angle), and
therefore also stronger ill-posedness and nonlinearity. Great efforts have been paid to tackle these
demanding problems, for instance [6–9].

Figure 1. Schematics of inverse scattering problems.
Methods for solving ISPs can be catalogued into two types of optimization approaches:
deterministic and stochastic. The deterministic type of the inversion methods has been developed for
decades, such as the contrast source inversion (CSI) method [6,10,11], the Born iterative method
and distorted Born iterative method [12,13], the level set method [14,15], the subspace-based
optimization method (SOM) [8,16–18], etc. The second type, the stochastic type of the inversion
methods, usually employs a group of initial guesses and uses the stochastic optimization scheme to
minimize the objective function, such as the genetic algorithm and the evolutionary optimization.
The stochastic type methods increase the possibilities of finding the global minimum rather than being
trapped in a local minimum as the deterministic optimization techniques [19,20]. Both techniques
have also been applied to solve 3-D inverse scattering problems [6,21–24]. Other than the quantitative
methods mentioned above, some qualitative methods, such as the linear sampling method [25] and
other methods, like [26], are proposed to retrieve the geometric supports of the unknown scatterers.
To tackle the nonlinearity, the major difficulty in efficiently solving the ISPs, different types
of integral equations have been proposed, such as in [27–29]. In particular, motivated by the
contraction integral equation (CIE) in solving the direct scattering problems with highly conductive
background media [30], a new type of contraction integral equation for inversion (CIE-I) has been
proposed [29] to tackle the two-dimensional highly nonlinear ISPs by transforming the usually
employed Lippmann–Schwinger integral equation (LSIE) into a new form with a modified medium
contrast via a contraction mapping. With the CIE-I, the global multiple scattering effects (MSE) are
substantially suppressed in estimating the modified contrast in the CSI type methods. With the FFT
type twofold SOM regularization scheme [8], the inversion solver with CIE-I is capable of effectively
alleviating the nonlinearity of the two-dimensional ISPs, especially those highly nonlinear ones with
strong scatterers with large contrast and/or large dimensions (in terms of wavelengths). In this paper,
this new CIE-I will be implemented in tackling the 3-D ISPs.
In summary, the contributions of the paper include:
1. The CIE-I is firstly implemented to tackle the computationally costly full 3-D ISPs, to address the
highly nonlinear 3-D ISPs, and to accelerate the convergence of the inversions.
2. A relaxed type of inversion scheme based on CIE-I is proposed, with different auxiliary
parameters β (the parameter in CIE-I to control the portion of MSE in estimating the contrast) in
11
J. Imaging 2019, 5, 27
updating the contrast sources and in updating the contrast. This means to further accelerate the
convergence of the inversions.
3. Several numerical tests are provided with details, for the sake of further algorithmic studies.
After the Introduction, the proposed 3-D inversion method is detailed in Section 2. In Section 3,
three numerical examples are given to validate the proposed method. Finally, conclusions are drawn.
2. Inversion with CIE-I

In this paper, the domains of interest (DoI) are chosen to be rectangular cuboid in order to
implement the conjugate gradient fast Fourier transform (CG-FFT) scheme and apply the Fourier
basis in TSOM. For the convenience of reading, in this paper, we denote the one-dimensional tensor
ˆ
as a, two-dimensional tensor as a, three-dimensional tensor as â, and four-dimensional tensor as â.
Unless otherwise specified, the subscript of the tensors denotes the index of the element, such as
am,n denotes the element in a with index {m, n}. We use bold symbols to denote vectorial physical
quantities, such as the positions r and the electric fields E in 3-D cases.
2.1. 3-D Modeling

In 3-D cases, there are Ni incident waves from different angles impinging onto the rectangular
cuboid DoI D (D ⊂ R3 , the background 3-D homogeneous medium with permittivity 0 and
permeability μ0 ), where nonmagnetic scatterers are located, and these incident waves are expressed
l (r ), l = 1, 2, . . . , Ni , r ∈ D . For each incidence, the scattered fields are collected by Nr receivers
as Einc
located at r j , j = 1, 2, . . . , Nr . With all such information, including every incident field inside the
domain of interest and the corresponding scattered fields at the positions of all detectors, we aim at
determining the dielectric profile (r ), r ∈ D .
The scattering models are governed by the following electric field volume integral equation based
on the well-known Lippmann–Schwinger integral equation (LSIE). For the l th incidence, the field
equation in the domain D is expressed as
I l (r ) = χ(r ) Einc
l (r ) + χ (r )( GD I l )(r ),
3D
(1)
where χ(r ) = (r (r ) − 1) is the contrast, r (r ), I l (r ) = χ(r ) Etot l (r ) and El (r ) are the relative
inc
permittivity, the contrast source and the incident electric field at r, respectively, and ( G3D D I l )(r )
is an integral operator with the dyadic Green’s function as the integral kernel, which can be
written as [8,31],

∇∇
( G3D
D I l )( r ) : = k 2
0 I + · g(r, r ) I l (r )dr , (2)
D k20
√
with k0 = ω 0 μ0 as the wave number of the background medium, and g(r, r ) as the 3-D Green’s
function for the background homogeneous medium. Notice that the I l (r ) is the contrast source in this
paper, while, in [8], it is the physically induced current that includes a multiplicative factor −iω0
compared to the contrast source. Otherwise, as we know, the dyadic Green’s function is composed of
nine scalar elements, which represent the mapping from the three components of the contrast source
to the three components of the scattered fields inside the DoI.
The nonlinearity of ISPs comes from the MSE, as shown in Equation (1). To alleviate the
nonlinearity of the model, in [29], by some mathematical manipulation on Equation (1), another
new-type integral equation, which is denoted as CIE-I herein, can be obtained

β(r ) I l (r ) = R(r ) β(r ) I l (r ) + R(r ) Einc
l (r ) + ( GD I l )(r ) ,
3D
(3)

where R(r ) = β(r )χ(r ) [ β(r )χ(r ) + 1] is the modified contrast function, and β(r ) is a chosen auxiliary
parameter to control the portion of the MSE in estimating the contrast [29], which can be a constant
12
J. Imaging 2019, 5, 27
or a variable at the different position in the DoI. For the convenience of discretizing the equation,
we rewrite the dyadic Green’s function as
⎧
⎪
⎪ 1 ∂2
⎨ k20 (1 + 2 2 ) D g(r, r ) Il;v (r )dr , if u = v,
k ∂ι
( GD;uv Il;v )(r ) :=
3D 0 u (4)
⎪
⎪ ∂ 2
⎩ g(r, r ) Il;v (r )dr , if u = v,
∂ι u ∂ι v D
where u, v = 1, 2, 3 represent the x-, y- and z-components of a vector, respectively, and ι1 = x, ι2 = y,

and ι3 = z.
Following the conventions in [8], we can rewrite the vectorial Equation (3) into three coupled
scalar equations in the discrete forms. Thus, we first discretize the rectangular domain of interest
into small cuboid subdomains, whose dimensions are much smaller than the wavelength and the
center of which are located at r m,n,p , with m, n, p integers and m ∈ [1, M1 ], n ∈ [1, M2 ] and p ∈ [1, M3 ].
Here, M1 , M2 , and M3 are the total number of subdomains along x-, y-, and z-directions, respectively,
and we let M = M1 × M2 × M3 be the total number of the subdomains. With such discretization,
we have

3
β̂ m,n,p Îl;u;m,n,p = R̂m,n,p β̂ m,n,p Îl;u;m,n,p + R̂m,n,p Êl;u;m,n,p
inc
+ ∑ ĜD;uv;m,n,p
3D
( Îl;v ) , (5)
v =1

where R̂m,n,p = β̂ m,n,p χ̂m,n,p β̂ m,n,p χ̂m,n,p + 1 , χ̂m,n,p is the contrast at r m,n,p , whereas Êl;u;m,n,p
inc and
Îl;u;m,n,p are the incident electric field and the induced current at r m,n,p , respectively. Subscript u = 1, 2, 3
denotes the x, y, and z components of a vector. Note that the convolution-type operators ĜD;uv 3D are
obtained via the Equation (4). For further details of the discretization of Equation (1) and the finite
difference scheme to generate ( G3D D I l )(r ), please refer to the Appendix of [6].
Similarly, the integral operator relating the contrast sources and the scattered fields could also be
expressed as the summation of the contribution from all the subdomains,
sca 3D
El = GS · I l , (6)
T
sca sca T sca T sca T T sca
where El = El;1 , El;2 , El;3 is a 3Nr dimensional vector with El;κ = El;κ;1
sca , Esca , . . . , Esca
l;κ;2 l;κ;Nr
(κ = 1, 2, 3 denotes the x, y, z component
of the corresponding vector, respectively), I l is a 3M
dimensional vector obtained by I l = vec Îˆl . In a 3-D scenario, the vectorization operation vec {·} is

defined to vectorize a four-dimensional tensor into a vector, i.e., if I l = vec Îˆl , we have I l;ϑ = Îˆl;m,n,p,κ
with ϑ = (κ − 1) × M + ( p − 1) × ( M1 × M2 ) + (n − 1) × M1 + m. In Equation (6), the scattering
operator is defined as
⎡ ⎤
GS;11 GS;12 GS;13
3D ⎢ ⎥
GS = ⎣ GS;21 GS;22 GS;23 ⎦ , (7)
GS;31 GS;32 GS;33
a 3Nr × 3M matrix, with GS;uv , a Nr × M matrix, the mapping from the v component of the induced
current to the u component of scattered fields (the subscripts u, v = 1, 2, 3 are not indexed for tensor
elements). The explicit expression of GS;uv is

3D ik0
GS;uv ( a, b) = k20 + R a,b − 1
R2a,b
δ(u − v)+
(8)
k2
(r a )u − (r m,n,p )u (r a )v − (r m,n,p )v − R20 − 3ik0
R3a,b
+ 3
R4a,b
g(r a , r m,n,p ),
a,b
13
J. Imaging 2019, 5, 27

where δ(y) is 1 when y = 0 and is 0 otherwise, R a,b = r a − r m,n,p in which b = ( p − 1) × ( M1 ×
M2 ) + (n − 1) × M1 + m with a = 1, 2, . . . , Nr , b = 1, 2, . . . , M, m = 1, 2, . . . , M1 , n = 1, 2, . . . , M2 ,
and p = 1, 2, . . . , M3 . Here, (r )u denotes the u component of r.
2.2. Objective Function for Inversions

In this subsection, we build the objective function used in the proposed inversion method,
in which we will use the FFT-TSOM [8] as the regularization to stabilize the inversion, as done in [29].
Emphasize that the twofold subspace constraints have different regularization effects. The first fold,
the original SOM, balances the two mismatches in the objective function, whereas the second fold,
the TSOM, is the key to stabilize the inversions with CIE-I. Details are given in below.
3D
For the original SOM part, as given in [16], with the spectral information of GS (the singular
3D 3D 3D
value decomposition–SVD of GS · vSj = σjS uSj and the complexity of thin SVD of GS is
tells GS
O(27Nr M), assuming that the singular values σjS is a non-increasing sequence), the contrast sources
2
can be decomposed into two parts, deterministic part of the contrast sources (DPCS) and ambiguous
part of the contrast sources (APCS), the former being obtained as
sca
d
L uSj ∗ · El +
Il = ∑ vSj = V S · α+
l , (9)
j =1 σjS
+ T
sca
where V S = vS1 , vS2 , . . . , vSL , α+ + + +
l = αl;1 , αl;2 , . . . , αl;L with α+ S∗
l;j = ( u j · El ) /σj , j = 1, 2, . . . , L,
S
and the superscript ∗ denotes the Hermitian operation while superscript + refers to the dominant
current subspace, the subspace corresponding to the dominant singular values. The value of L is
chosen according to the noise level [16]. Later, we will see how this might work after introducing the
objective function for the inversion.
For the second fold subspace constraint, according to the FFT-TSOM [8], the APCS can be written as

a tmp + +∗ tmp
I l (γ̂ˆ l ) = I l − VS · VS · Il , (10)

tmp tmp tmp tmp
where γ̂ˆ l = γ̂x;l ; γ̂y;l ; γ̂z;l , I l = I x;l ; I y;l ; I z;l with
tmp !!
I u;l = vec IDFT γ̂u;l (11)
and u = x, y, and z, IDFT {} is the inverse discrete Fourier transform operator, the vec{·} is the
vectorization operator. Note that the inverse discrete Fourier transform (IDFT) is performed by the 3-D
FFT algorithm, the computational complexity of which is O( M log2 M ), with M = M1 × M2 × M3 .
By using the low-frequency Fourier components, we are able to constrain the APCS within
a low-dimensional subspace. The reason [17] is to use only the contrast sources components in this
subspace that is influential to the scattered fields within the DoI, such that the stability of the inversions
is substantially increased. To reduce the computational costs, such a subspace can be approximately
spanned by low-frequency Fourier bases [8]. Here, if we use all Fourier bases, the construction of the
APCS becomes the one in the original SOM. Otherwise, we can use the low-frequency components,
and set the coefficients for those high-frequency components as zeros. This can be achieved via using
a mask with eight corners being 1, with size MF (1 ≤ MF ≤ M1 /2, M2 /2, M3 /2), and other positions
being 0. The details can be found in [8].
Having expressed the contrast sources as aforementioned, it is straightforward to define the
objective function. Firstly, it is natural to give the mismatch of the scattered fields by
" "
" a d sca "2
l ( γ̂l ) = "GS · I l + GS · I l − El " ,
Δfie ˆ (12)
14
J. Imaging 2019, 5, 27
d a
where I l and I l are as in Equations (9) and (10), respectively, and · denotes the L2 norm of a tensor.
The current equation in Equation (5) is another key equation to satisfy. Using the APCS construction
Equation (10), we define an operator as

# $ 3
L3D γ̂ˆ l
κ;m,n,p
= β Îl;κ;m,n,p
a
− β R̂m,n,p Îl;κ;m,n,p
a
− R̂m,n,p ∑ ĜD;κv;m,n,p
3D
( Îl;v
a
) . (13)
v =1
With this definition, we could write the mismatch of Equation (5) as

" "
" 3D ˆ ˆ 3D "2
l ( γ̂l , R̂ ) = " L ( γ̂l ) − Γ̂l " ,
Δcur ˆ (14)
3D
where Γ̂ˆ l;κ;m,n,p = − β Îl;κ;m,n,p + β R̂m,n,p Îl;κ;m,n,p + R̂m,n,p ˆ inc
d ) + Ê
( Îl;v
d d
∑3v=1 ĜD;κv;m,n,p
3D
l;κ;m,n,p .
Then, the objective function can be given as
" "
Ni " " " inc "2
" sca "2 cur " "
f (γ̂ˆ 1 , γ̂ˆ 2 , . . . , γ̂ˆ Ni , R̂) = ∑ Δfie
l / "E l " + Δ l / " l " .
E (15)
l =1
The inversion is to minimize this objective function. As in [8], the conjugate gradient (CG) type
algorithm that is used in contrast source inversion (CSI) method is adopted to minimize this nonlinear
problem by alternatively updating the γ̂ˆ l and R̂ at every iteration of the optimization.
2.3. Sketch of the Inversion Method

Following the inversion method in [8,29], we summarize it as follows:
1. Set the background medium and null APCS as the initial guesses and choose an L value such
3D
that the corresponding first L singular values of GS are larger than the noise level (assuming
that the noise is a white Gaussian one).
2. Set proper values for β in CIE-I modeling and proper value for MF to control the number of
Fourier bases being used.
3. Carry out the CG type optimization algorithm to alternatively update the two types of variables,
where the APCS is updated with a one-step Polak–Ribière CG scheme and the contrast is updated
with the least squares method.
4. Stop the optimization if a termination condition is met, which can be a maximum number of
iterations or a pre-defined relative change of APCS coefficients.
5. If the maximum number of rounds of inversion is met, go to Step 6. Otherwise, the obtained
contrast and APCS will be used as the initial guesses for the next round of optimization with
smaller β and larger MF in Step 3.
6. Output the obtained contrast.
As mentioned in [29], the first round inversion with a large β and with a low-dimensional
subspace enables a fast convergence to a meaningful coarse result that could be used as an initial guess
in the second round. By increasing the dimension of the APCS subspace and including more MSE
in estimating the contrast, the inversion gives us a result with better resolution. Since in the second
round a (supposedly) good initial guess is given, the convergence is also very fast. For some difficult
problems, the inversion could be carried out with multi-round optimizations by gradually increasing
the dimensions of the APCS subspace in each round while decreasing the values of β in estimating
the contrast.
15
J. Imaging 2019, 5, 27
2.4. Updating Contrast and Contrast Sources with Different β

As mentioned above, there are two updates at each iteration for the two types of unknowns. In [29],
the same β is used in each round of optimization for both updates. In the first round optimization
in inversions, to tackle the nonlinearity, a large β is needed in updating the modified contrast,
and therefore we use the same β for the update of APCS. In the subsequent rounds (instead of
the first round) of optimization, for the sake of stability, a small β is used in estimating the modified
contrast to cope with a high-dimensional APCS subspace, as discussed in [29]. However, there is not
such a need to use a small β in updating the APCS. Consequently, in the second round of optimization,
we can use a CIE-I model with larger value β 1 for the update of the APCS and another with smaller
value β 2 for the update of the contrast at the same iteration. The same applies to the third round
optimization, if there is such a need. The purpose of doing so is to further accelerate the convergence
of these computationally burdensome 3-D inversions after the first round optimization.
As for the value range for β 2 , we carry out a similar calculation as in [29], which reflects the norm
of the scattering operator that maps the contrast sources to the scattered fields within DoI. The results
shown in Figure 2 indicate that, if we want to suppress the MSE in estimating the contrast when using
the least squares method, we might need to choose a value for β 2 that is larger than 3.5, and this is
the guideline for the first round optimization. For the second or subsequent rounds, as good initial
guesses are provided, we can include more MSE to retrieve the fine features of the unknown scatterers.
3
3
5
10 2.5
2.5
10
2 20 2
15
1.5 1.5
30
20
1 1
25 40
0.5 0.5
30 50
10 20 30 10 20 30 40 50
(a) (b)
Figure 2. The values of |(GD;xx

3D I )(r )| in the plane x = 0 when I = 1 for a DoI with size (a) 3λ × 3λ × 3λ
x x
and (b) 5λ × 5λ × 5λ.
3. Numerical Simulations
In this section, we will test the proposed inversion methods in three examples. We will use the
same physical setup of sources and receivers for the 3-D example in [8], as shown in Figure 3. The DoI
are all the same, a box with size 3λ × 3λ × 3λ. The box is illuminated by 60 electric dipole antennas at
300 MHz (λ = 1 m in air), located at three circles (with 20 dipole antennas evenly distributed on each)
with the same radius 3 m. The three circles are in x − y, y − z and x − z planes, and their centers are at
(0.2, 0, –0.1), (0.1, 0, –0.15), and (–0.05, 0.1, 0), respectively. The direction of the electric dipole sources
in the y − z plane are in the x-direction, while those in the x − z and x − y planes are in the y- and
z-directions, respectively. Scattered fields are measured by 60 receivers, located at the same positions as
the 60 dipole sources. We collect all three components of the fields. Consequently, we have a 60 × 180
synthetic data matrix. In all three tests, the synthetic data are calculated with a 60 × 60 × 60 mesh,
while, in the inversions, a 30 × 30 × 30 mesh is used. All synthetic data are contaminated with additive
Gaussian white noise with level of 10%.
16
J. Imaging 2019, 5, 27
-1
-2
-3
2
0
-2 0 1 2 3
-3 -2 -1
Figure 3. The physical setup for the numerical tests.
The termination condition in each round of optimization is to reach a pre-defined relative changing
rate of the APCS coefficients, given as
%
& (h) ( h −1) 2
& 1 Ni γ̂ˆ l − γ̂ˆ l
δ3D =' ∑ ( h −1) 2
, (16)
Ni l =1 γ̂ˆ l
(h)
where γ̂ˆ l is the APCS Fourier coefficient for the l th incidence at hth iteration. We set δ3D < 10−3 as
the termination condition.
The reconstruction results are quantitatively evaluated by the following error:
%
&
&1 |ˆr;m,n,p
est − ˆr;m,n,p
tr |2
Err = ' ∑ . (17)
M m,n,p |
ˆ tr
r;m,n,p | 2
When running on a workstation with eight threads and 32 Gb RAM, every iteration of the
proposed inversion method costs about 25 seconds CPU time in MATLAB, including one update
on the contrast sources and one update on the contrast. Due to the independence between different
incidences, we use 8 MATLAB workers to update the contrast sources in parallel.
3.1. Example 1
We reuse the 3-D test in [8], where a coated cube is employed. The cube is with an outer layer
(r2 = 1.5 + i0.3) and an inner layer (r1 = 2 + i0.8), where the outer edge length is b = 2λ and inner
edge length is a = 1λ, as shown in Figure 4a. In [8], we observe that the inversion with LSIE and
FFT-TSOM is able to satisfactorily retrieve the cube while the LSIE with SOM fails. Here, we illustrate
how the proposed inversion method with CIE-I and FFT-TSOM performs. Similarly, we carry out
two rounds of optimization for this inversion as done in [8] with MF = 6 and 10, but with CIE-I,
we use β = 6 in the first round and β = 1 for the second round. The reconstruction results are
shown in Figure 5, with {ˆrest } as the real part of the reconstructed relative permittivity and
{ˆrest } as the imaginary part, where we see after two rounds of optimization that the coated cube is
successfully found.
Now, we compare the convergence speeds of inversions with LSIE and CIE-I. We plot the
reconstruction errors for both cases, as shown in Figure 6, in which the inversion with CIE-I appears to
be converging much faster than the one with LSIE, with the same reconstruction quality. This confirms
again the speeding convergence of inversions in two-dimensional cases shown in [29]. We would like
to emphasize that this fast convergence property is extremely important in handling the 3-D ISPs, since
the computational costs in 3-D problems are much higher than in two-dimensional cases.
17
J. Imaging 2019, 5, 27
a
εr1
εr2
b
(a) The coated cube in Example 1.
2.2 1
2
0.8
1.8
0.6
1.6
0.4
1.4
0.2
1.2
0
1
0.8 0.2
(b) Real part of the true profile of the coated cube. (c) Imaginary part of the true profile of the coated cube.
Figure 4. Coated cube used in Example 1.
0.8
2
0.7
1.8
0.6
0.5
1.6
0.4
1.4 0.3
0.2
1.2
0.1
0
1
(a) {ˆrest } after the first round optimization. (b) {ˆrest } after the first round optimization.
2.4 1
2.2
0.8
0.6
1.8
1.6 0.4
1.4
0.2
1.2
0
1
(c) {ˆrest } after the second round optimization. (d) {ˆrest } after the second round optimization.
Figure 5. Reconstruction results of inversions with CIE-I. In the four subfigures, the 30 slices of DoI
with each at z = zq plane, where zq , q = 1, . . . , 30, are the grid points along z-direction, and z p < zq if
p < q. The displaying sequence is with the convention of left to right, and top to down. For instance,
the top left corner one is with z1 , and the top row second column one is with z2 . The same applies to
the figures hereafter. In the first round, β 1 = β 2 = 6. In the second round, β 1 = β 2 = 1.
18
J. Imaging 2019, 5, 27
0.22
LSIE, MF = 6
0.2 LSIE, MF = 10
CIE-I, MF = 6, 1
= 6, 2
=6
0.18 CIE-I, MF = 10, = 1, =1
1 2
CIE-I, MF = 10, 1
= 6, 2
=1
0.16
0.14
0.12
0.1
0.08
0 100 200 300 400 500
Figure 6. Errors of reconstructions obtained by different inversion methods for Example 1.
With this example, we further test the relaxed scheme with different β for updating the APCS and
the contrast. The termination condition remains the same, δ3D < 10−3 . We apply this scheme in the
second round optimization, with β 1 = 6 for the APCS update and β 2 = 1 for the contrast update. For
comparison, the recorded reconstruction error at each iteration is plotted in Figure 6 as well, which
shows that the convergence is further accelerated compared with the case of using β 1 = 1.
3.2. Example 2
In this test, we use a profile with four cubes with increasing contrasts. Having edge length 0.8λ,
they are centered at (−0.7, 0.7, −0.7)λ, (0.7, −0.7, −0.7)λ, (−0.7, −0.7, 0.7)λ, and (0.7, 0.7, 0.7)λ, with
r = 2, 3, 4 and 5, respectively, as shown in Figure 7.
5.5
4.5
3.5
2.5
1.5
0.5
Figure 7. True permittivity profile (real part) for Example 2.
Two inversions are carried out with LSIE and CIE-I, each with two rounds (MF = 6 and 10).
For CIE-I, the choices of β are 6 and 1 in the two rounds. The reconstruction results and errors are
shown in Figures 8–10. From these results, we clearly observe that the inversion with LSIE is only
able to find the bottom two cubes with lower contrasts, whereas the inversion with CIE-I succeeds in
finding all four of four cubes. Reconstruction errors displayed in Figure 9 confirm it, inversion with
LSIE diverging while converging with CIE-I.
19
J. Imaging 2019, 5, 27
2.4 0.3
2.2
0.2
2
0.1
1.8
1.6 0
1.4
-0.1
1.2
-0.2
1
0.8 -0.3
0.6
-0.4
(a) {ˆrest } after the first round optimization (b) {ˆrest } after the first round optimization
3 0.8
0.6
2.5
0.4
2
0.2
1.5 0
-0.2
1
-0.4
0.5
-0.6
(c) {ˆrest } after the second round optimization (d) {ˆrest } after the second round optimization
Figure 8. Reconstruction results of inversions with LSIE for Example 2.
0.32
0.3
0.28
0.26
Err
0.24
0.22
0.2
0.18
0 100 200 300 400 500 600 700
Iteration
In addition, we test the inversion scheme with different β for updating the APCS and for updating
the contrast. Firstly, we choose β 1 = 6 for the update of APCS and β 2 = 1 for the update of the contrast
in the second round of optimization with MF = 10, which however immaturely converges. This might
be due to the reason of using a large β 1 . Then, we choose β 1 = 3, and the optimization converges to
a good solution with a convergence speed faster than when using β 1 = 1.
From this example, the inversion with CIE-I is able to handle the highly nonlinear problems
consisting of strong scatterers, those with either large contrasts or large dimensions (in terms of
wavelength), while the inversion with LSIE may fail.
20
J. Imaging 2019, 5, 27
0.6
3.5
0.4
3
2.5 0.2
2 0
1.5
-0.2
1
-0.4
0.4
5.5
5 0.3
4.5
0.2
4
0.1
3.5
3 0
2.5 -0.1
2
-0.2
1.5
-0.3
1
0.5 -0.4
Figure 10. Reconstruction results of inversions with CIE-I for Example 2. In the first round, β 1 = β 2 = 6.
In the second round, β 1 = β 2 = 1.
3.3. Example 3
We will use another highly nonlinear example to confirm the large difference between the
performances of inversion methods with CIE-I and LSIE. In this example, we use a profile that
is similar to the famous two-dimensional “Austria” that has an annular and two separated disks. Here,
we use a coated cube with a hollow inside and two rods outside the cube, as shown in Figure 11, all of
which are with r = 2.5.
2.5
1.5
0.5
Figure 11. True permittivity profile (real part) for Example 3.
21
J. Imaging 2019, 5, 27
We still carry out inversions with LSIE and CIE-I with two rounds of optimization, the first with
MF = 6 and the second with MF = 10. For the CIE-I, β 1 = β 2 = 6 is used in the first round, and
β 1 = β 2 = 1 in the second round. The reconstructions and errors are shown in Figures 12, 13a and 14.
We see that the inversion with LSIE fails to find the profile while the one with CIE-I succeeds. This can
also be seen from the errors of the reconstructions in Figure 13a, where the error by the inversion with
LSIE diverges.
1.4
0.4
1.2 0.3
0.2
1
0.1
0.8
0.6
-0.1
0.4
-0.2
0.2 -0.3
2 0.8
0.6
1.5 0.4
0.2
1 0
-0.2
0.5 -0.4
-0.6
0 -0.8
-1
Figure 12. Reconstruction results of inversions with LSIE for Example 3.
0.45 0.38
0.36
0.4 0.34
0.32
0.35 0.3
0.28
0.3 0.26
0.24
0.25 0.22
0.2
0.2 0.18
0 200 400 600 800 1000 0 100 200 300 400 500 600
(a) Errors of reconstructions in inversions with MF = 6 (b) Errors of reconstructions in inversions with MF = 4,
and 10. 8, and 10.
22
J. Imaging 2019, 5, 27
2.6
2.4 0.8
2.2
0.6
0.4
1.8
1.6
0.2
1.4
0
1.2
1
-0.2
2.5
0.8
2 0.6
0.4
1.5
0.2
1
0
-0.2
0.5
-0.4
Figure 14. Reconstruction results of inversions with CIE-I for Example 3. In the first round, β 1 = β 2 = 6.
In the second round, β 1 = β 2 = 1.
In the reconstruction results in Figure 14, we see that there is an artefact within the coated cube,
where the medium is supposed to be air. This artefact appears in the reconstructed result after the first
round optimization with MF = 6, meaning that the inversion, though capable of retrieving most of the
main features of the scatterers, is still not stable enough. Therefore, we carry out another inversion with
three rounds of optimization, with MF = 4, 8, and 10 in each round, and β = 6, 1, and 0.5, respectively.
The final reconstruction result is given in Figure 15, and we see that there is no more such an artefact.
23
J. Imaging 2019, 5, 27
2.5 0.2
0.1
2
0
1.5 -0.1
-0.2
1
-0.3
-0.4
(a) {ˆrest } after three rounds of optimization (b) {ˆrest } after three rounds of optimization
Figure 15. Reconstruction results of inversions with CIE-I after three rounds of optimization for Example
3. In the first round, β 1 = β 2 = 6 and MF = 4. In the second round, β 1 = β 2 = 1 and MF = 8. In the
third round, β 1 = β 2 = 0.5 and MF = 10.
At last, we also carry out the inversions with different β values for the updates of the APCS and
the contrast, and the results are shown in Figure 13. We see that with larger β 1 value for the update of
the APCS in the second round optimization, the convergence speed is indeed faster.
From this example, we again confirm the great resolvability of the inversion method with CIE-I
against the high nonlinearity of the ISPs.
4. Conclusions
In this paper, we propose an inversion method with CIE-I to tackle the 3-D electromagnetic
ISPs, especially for highly nonlinear problems with strong scatterers, those with high contrasts or
electrically large dimensions. With the CIE-I modeling, the multiple scattering effects in estimating
the contrast can be suppressed such that the nonlinearity of inversion can be effectively alleviated.
Together with the FFT-TSOM regularization scheme, this largely increases the ability of retrieving the
profile of strong scatterers, and effectively accelerates the convergence when handling the moderate
scatterers. In addition, by relaxing the contraction mapping, i.e., different contraction mappings are
used in updating the contrast sources and in updating the contrast, the convergence of the inversions
can be further accelerated. Through numerical examples, we clearly show that the inversions with
CIE-I outperform the ones with LSIE, in terms of the resolvability against the nonlinearity and the
convergence speed. As shown in [32], the resolvability of the inversion solver with CIE-I against
nonlinearity can be further improved in the two-dimensional case if proper regularization techniques
are implemented, which should be expected in the 3-D case as well.
Author Contributions: Conceptualization, Y.Z.; Methodology, Y.Z.; Software, Y.Z. and K.X.; Validation, Y.Z., and
K.X.; Formal Analysis, Y.Z. and K.X.; Investigation, Y.Z.; Resources, K.X.; Data Curation, Y.Z.; Writing—Original
Draft Preparation, Y.Z. and K.X.; Writing—Review and Editing, Y.Z., and K.X.; Visualization, Y.Z.; Supervision, Y.Z.
Funding: K.X. was partially funded by NSFC Grant No. 61601161.
Acknowledgments: The authors would like to thank Dominique Lesselier at Laboratoire des Signaux et Systèmes
(UMR8506 CNRS-CentraleSupélec-Université Paris-Sud), France, for his very kind help in discussions of 3-D
inversion methods and in writing the manuscript.
24
J. Imaging 2019, 5, 27
Abbreviations
The following abbreviations are used in this manuscript:
CIE-I Contraction integral equation for inversion

LSIE Lippmann–Schwinger integral equation
3-D Three-dimensional
CSI Contrast source inversion
SOM Subspace-based optimization method
FFT-TSOM FFT type twofold subspace-based optimization method
APCS Ambiguous part of the contrast source
DPCS Deterministic part of the contrast source
DoI Domain of interest
MSE Multiple scattering effects
References
1. Colton, D.; Kress, R. Inverse Acoustic and Electromagnetic Scattering Theory; Springer: New York, NY, USA, 2013.
2. Abubakar, A.; van den Berg, P.M.; Mallorqui, J. Imaging of biomedical data using a multiplicative regularized
contrast source inversion method. IEEE Trans. Microw. Theory Tech. 2002, 50, 1761–1771. [CrossRef]
3. Abubakar, A.; van den Berg, P.M. Three-dimensional inverse scattering applied to cross-well induction
sensors. IEEE Trans. Antennas Propag. 2000, 38, 1669–1681. [CrossRef]
4. Massa, A.; Boni, A.; Donelli, M. A classification approach based on SVM for electromagnetic subsurface
sensing. IEEE Trans. Antennas Propag. 2005, 43, 2084–2093. [CrossRef]
5. Sabatier, P.C. Past and future of inverse problems. J. Math. Phys. 2000, 41, 4082–4124. [CrossRef]
6. Abubakar, A.; van den Berg, P.M. Iterative forward and inverse algorithms based on domain integral equations
for three-dimensional electric and magnetic objects. J. Comput. Phys. 2004, 195, 236–262. [CrossRef]
7. Zhong, Y.; Chen, X.; Agarwal, K. An improved subspace-based optimization method and its implementation
in solving three-dimensional inverse problems. IEEE Trans. Geosci. Remote Sens. 2010, 48, 3763–3768.
[CrossRef]
8. Zhong, Y.; Chen, X. An FFT twofold subspace-based optimization method for solving electromagnetic inverse
scattering problems. IEEE Trans. Antennas Propag. 2011, 59, 914–927. [CrossRef]
9. Litman, A.; Lorenzo, C. Special section on testing inversion algorithms against experimental data: 3-D targets.
Inverse Probl. 2009, 25, 020201. [CrossRef]
10. Van den Berg, P.M.; Kleinman, R.E. A contrast source inversion method. Inverse Probl. 1997, 13, 1607–1620.
[CrossRef]
11. Van den Berg, P.M.; van Broekhoven, A.L.; Abubakar, A. Extended constrast source inversion. Inverse Probl.
1999, 15, 1325–1344. [CrossRef]
12. Wang, Y.; Chew, W.C. An iterative solution of two-dimensional electromagnetic inverse scattering problem.
Int. J. Imaging Syst. Technol. 1989, 1, 100–108. [CrossRef]
13. Chew, W.C.; Wang, Y. Reconstruction of two-dimensional permittivity distribution using the distorted Born
iterative method. IEEE Trans. Med. Imaging 1990, 9, 218–225. [CrossRef] [PubMed]
14. Dorn, O.; Lesselier, D. Level set methods for inverse scattering. Inverse Probl. 2006, 22, R67–R131. [CrossRef]
15. Benedetti, M.; Lesselier, D.; Lambert, M.; Massa, A. A multi-resolution technique based on shape optimization
for the reconstruction of homogeneous dielectric objects. Inverse Probl. 2009, 25, 015009. [CrossRef]
16. Chen, X. Subspace-based optimization method for solving inverse scattering problems. IEEE Trans. Geosci.
Remote Sens. 2010, 48, 42–49. [CrossRef]
17. Zhong, Y.; Chen, X. Twofold subspace-based optimization method for solving inverse scattering problems.
18. Agarwal, K.; Pan, L.; Chen, X. Subspace-based optimization method for reconstruction of two-dimensional
complex anisotropic dielectric objects. IEEE Trans. Microw. Theory Tech. 2009, 58, 1065–1074. [CrossRef]
19. Pastorino, M. Stochastic optimization methods applied to microwave imaging: A review. IEEE Trans.
Antenna Propag. 2007, 55, 538–548. [CrossRef]
20. Rocca, P.; Benedetti, M.; Donelli, M.; Franceschini, D.; Massa, A. Evolutionary optimization as applied to
inverse scattering problems. Inverse Probl. 2009, 25, 123003. [CrossRef]
25
J. Imaging 2019, 5, 27
21. De Zaeytijd, J.; Franchois, A.; Geffrin, J.M. A new value picking regularization strategy-Application to the
3-D electromagnetic inverse scattering problem. IEEE Trans. Antenna Propag. 2009, 57, 1133–1149. [CrossRef]
22. Chaumet, P.; Belkebir, K. Three-dimensional reconstruction from real data using a conjugate gradient-coupled
dipole method. Inverse Probl. 2009, 25, 024003. [CrossRef]
23. Yu, C.; Yuan, M.; Liu, Q.H. Reconstruction of 3-D objects from multi-frequency experimental data with a fast
DBIM-BCGS method. Inverse Probl. 2009, 25, 024007. [CrossRef]
24. Donelli, M.; Franceschini, D.; Rocca, P.; Massa, A. Three-dimensional microwave imaging problems solved
through an efficient multiscaling particle swarm optimization. IEEE Trans. Geosci. Remote Sens. 2009, 47,
1467–1481. [CrossRef]
25. Agarwal, K.; Chen, X.; Zhong, Y. A multipole-expansion based linear sampling method for solving inverse
scattering problems. Opt. Express 2010, 18, 6366–6381. [CrossRef] [PubMed]
26. Bevacqua, M.T.; Isernia, T. Boundary Indicator for Aspect Limited Sensing of Hidden Dielectric Objects.
IEEE Geosci. Remote Sens. Lett. 2018, 15, 838–842. [CrossRef]
27. Isernia, T.; Crocco L.; D’Urso M. New tools and series for forward and inverse scattering problems in lossy
media. IEEE Geosci. Remote Sens. Lett. 2004, 1, 327–331. [CrossRef]
28. D’Urso, M.; Isernia T.; Morabito, A.F. On the Solution of 2-D Inverse Scattering Problems via Source-Type
Integral Equations. IEEE Trans. Geosci. Remote Sens. 2010, 48, 1186–1198. [CrossRef]
29. Zhong, Y.; Lambert, M.; Lesselier, D.; Chen, X. A new integral equation method to solve highly nonlinear
inverse scattering problems. IEEE Trans. Antennas Propag. 2016, 64, 1788–1799. [CrossRef]
30. Pankratov, O.V.; Avdeyev, D.B.; Kuvshinov, A.V. Electromagnetic field scattering in a heterogeneous earth:
A solution to the forward problem. Phys. Solid Earth 1995, 31, 201–209.
31. Peterson, A.F.; Ray, S.L.; Mittra, R. Computational Methods for Electromagnetics; IEEE Press: New York, NY,
USA, 1998.
32. Xu, K.; Zhong, Y.; Wang, G. A hybrid regularization technique for solving highly nonlinear inverse scattering
problems. IEEE Trans. Microw. Theory Tech. 2018, 64, 11–21. [CrossRef]
c 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
26
Journal of
Imaging
Article
Full-Vectorial 3D Microwave Imaging of Sparse
Scatterers through a Multi-Task Bayesian
Compressive Sensing Approach
Marco Salucci 1,2,† , Lorenzo Poli 1,2,† and Giacomo Oliveri 1,2, *,†
1 ELEDIA Research Center (ELEDIA@UniTN—University of Trento), Via Sommarive 9, I-38123 Trento, Italy;
[email protected] (M.S.); [email protected] (L.P.)
2 ELEDIA Research Center (ELEDIA@L2S—UMR 8506), 3 rue Joliot Curie, 91192 Gif-sur-Yvette, France
† These authors contributed equally to this work.
Received: 3 December 2018; Accepted: 8 January 2019; Published: 15 January 2019
Abstract: In this paper, the full-vectorial three-dimensional (3D) microwave imaging (MI) of sparse
scatterers is dealt with. Towards this end, the inverse scattering (IS) problem is formulated within
the contrast source inversion (CSI) framework and it is aimed at retrieving the sparsest and most
probable distribution of the contrast source within the imaged volume. A customized multi-task
Bayesian compressive sensing (MT-BCS) method is used to yield regularized solutions of the 3D-IS
problem with a remarkable computational efficiency. Selected numerical results on representative
benchmarks are presented and discussed to assess the effectiveness and the reliability of the proposed
MT-BCS strategy in comparison with other competitive state-of-the-art approaches, as well.
Keywords: microwave imaging; inverse scattering; Bayesian compressive sensing (BCS); contrast
source inversion (CSI); 3D
1. Introduction
Microwave imaging (MI) techniques are aimed at inferring the complex permittivity distribution
within an inaccessible investigation domain from the scattering interactions between the matter and
probing electromagnetic (EM) waves [1]. They have been successfully applied in several diagnostic
scenarios including non-destructive testing and evaluation [2,3], through-wall imaging [4], subsurface
prospecting [5–10], and structural health monitoring [11]. Moreover, they represent a very appealing
technology in many biomedical applications [12] such as, for instance, breast cancer detection [13–22]
thanks to the use of non-ionizing radiations. To date, significant efforts have been mostly devoted
to the development of two-dimensional (2D) MI algorithms, mainly based on transverse-magnetic
(TM) [23–28] or transverse-electric (TE) polarized [29] configurations, rather than fully-vectorial
three-dimensional (3D) ones [30]. As a matter of fact, under the assumption that the EM properties
of the unknown scattering scenario are invariant along a longitudinal direction, the arising inverse
scattering (IS) problem can be recast to the solution of simplified scalar Helmholtz equations [1]. On the
other hand, 2D-MI approaches are prone to errors when finite-volume scatterers are under test [31]
because of the over-simplified modelling of the scattering phenomena. The slower evolution of 3D-MI
techniques has been mainly caused by the higher complexity of both data-collection/storage and
image reconstruction processes with respect to the tomographic (2D) case. Moreover, a significantly
larger number of unknowns has to be retrieved and it becomes very hard to manage when there is
the need of high-resolution images (e.g., a realistic discretization of a human thorax needs millions of
voxels for having a clinical significance [32]). Furthermore, the amount of non-redundant information
on the investigation domain achievable from measurements is upper-bounded and the ratio between

J. Imaging 2019, 5, 19
unknowns and scattering data turns out to be very high [33]. Owing to such limitations, solving
3D-MI problems faces hard challenges and it requires the non-trivial implementation of effective
countermeasures to both the non-linearity and the ill-posedness issues of the arising full-vectorial
IS problem.
Dealing with 3D scenarios, synthetic aperture radar (SAR)-based methodologies such as confocal
MI [34] and synthetic near-field focusing [5] have been proposed. They are based on the emission of
wide-band pulses from multiple transmitting positions and the successive processing of the collected
echoes. Only target detection and localization (i.e., a qualitative imaging of the investigation domain)
is typically yielded, while quantitatively retrieving the distribution of the EM properties needs the
numerical solution of the non-linear scattering equations. Towards this end, effective 3D-MI approaches
have been presented in the scientific literature. They are based on the processing of time-domain [19,35]
or frequency-domain [36,37] data and, due to the extremely-wide dimension of the solution space,
which is usually proportional to the number of discretization domains, state-of-the-art methodologies
are primarily based on deterministic methods (e.g., Gauss-Newton (GN) [36,38], Conjugate-Gradient
(CG) [39], Level Set (LS) [16], and Inexact-Newton (IN) [40] methods, possibly formulated in different
functional spaces [41]), even though they do not a priori guarantee to reach the actual solution (i.e.,
the global optimum of the cost function quantifying the mismatch between measured and estimated
scattering data). To overcome such a drawback, either very efficient forward solvers [42,43] have been
introduced or stochastic multi-agent inversion algorithms, suitably integrated with multi-resolution
strategies [37] for a sustainable customization to the 3D case, have been successfully exploited.
Alternatively, computationally-efficient approaches to the MI problem have been recently explored
within the compressive sensing (CS) framework [23,29,44–49]. Despite the early stage [47,50] of their
implementation in 3D cases, very interesting applicative examples are already available [20].
According to the CS theory, sparseness priors can be enforced to solve the IS problem and to yield
a regularized solution provided that it admits a representation with few non-null coefficients in a
suitably chosen basis [44,45]. However, since available CS solvers generally deal with linear problems,
many sparsity-promoting approaches have been formulated within Born-like approximations, their
success being limited to weak scatterers or to specific applications where a qualitative imaging is
enough [29]. Alternatively, contrast source inversion (CSI)-based formulations of the IS problem can be
successfully employed to yield accurate reconstructions also in the presence of scatterers with high EM
contrast with respect to the surrounding medium [23]. Following such a line of reasoning, this paper is
aimed at presenting a novel computationally-efficient approach, based on a Bayesian CS (BCS) method,
to solve the 3D-IS problem concerned with non-weak scatterers. Towards this end, the full-vectorial IS
problem is formulated within a probabilistic CSI framework and then it is efficiently solved through a
customized multi-task BCS (MT-BCS) strategy.
The outline of the paper is as follows. The formulation of the 3D-CSI MI problem is detailed in
Section 2, while the proposed BCS-based solution strategy is described in Section 3. Selected numerical
results, from representative test cases, are presented and compared with competitive state-of-the-art
alternatives in Section 4. Finally, some concluding remarks are drawn (Section 5).
2. Mathematical Formulation
Let us consider a 3D isotropic non-magnetic [μ (r) = μ0 ] scattering scenario characterized by a
relative permittivity distribution, ε (r), and a conductivity profile, σ (r), r being the position vector
defined as r = xux + yuy + zuz , u p being the unit vector along the p-th (p = { x, y, z}) direction.
The goal of the MI problem at hand is to estimate the contrast function [51]
σ (r) − σ0
τ (r) [ ε (r) − 1] − j (1)
ωε 0
within the investigation domain D (i.e., ∀r ∈ D ), ω = 2π f , ε 0 , and σ0 being the angular frequency
(A time dependency factor exp ( jωt) is assumed and omitted hereinafter to simplify the notation,
28
J. Imaging 2019, 5, 19
but without loss of generality in the mathematical formulation.), the background permittivity and
conductivity (with σ0 = 0 hereinafter), respectively, and standing for “defined as”. Towards this
end, a set of V monochromatic ( f being the working frequency) plane-waves impinging from known
angular directions (θv , ϕv ), v = 1, ..., V, with known electric field
Eiv (r) = ∑ i
Ev,p (r) u p v = 1, ..., V (2)
p={ x, y, z}
is used to successively probe the unaccessible domain D ( Figure 1).
D z Eiv r
*1 *2 Tv
D 1
D 2 r uz
uy y
ux
Mv
*K m 1
:
x L
m 2
Figure 1. Geometry of the 3D-MI microwave imaging problem.
Under such hypotheses and by adopting a CSI formulation [52] for the scattering problem,
the electromagnetic interactions between the scatterers in D and the v-th (v = 1, ..., V) incident wave
can be mathematically described through the following integral data equation
# $ # $
Esv (r) = −ω 2 ε 0 μ0 Jv r · G r, r dr r ∈ Ω (3)
D
where Ω is an observation domain external to D (Ω ∩ D = ∅—Figure 1), Esv (r) [Esv (r) =
s (r) u ] is the v-th (v = 1, ..., V) scattered field defined as the difference between the
∑ p={ x, y, z} Ev,p p
v-th (v = 1, ..., V) electric field with, Ev (r) = ∑ p={ x, y, z} Ev,p (r) u p (i.e., the v-th total electric field),
and without, Eiv (r) (i.e., the v-th incident electric field), the scatterers in the background medium

(Esv (r) Ev (r) − Eiv (r) ), · stands for the scalar product, and
# √ $
# $ 1 1 exp − jω ε 0 μ0 |r − r |
G r, r = ∑ ∑ G pq u p uq = I+ (4)
q={ x, y, z} p={ x, y, z}
4π ω 2 ε 0 μ0 | r − r |
is the dyadic Green’s function for the homogeneous free-space background medium of dielectric
and magnetic properties ε 0 and μ0 , respectively, I being the unit tensor. In (3), Jv [Jv (r) =
29
J. Imaging 2019, 5, 19
∑ p={ x, y, z} Jv,p (r) u p ] is the v-th (v = 1, ..., V) unknown contrast current density induced in the
investigation domain (r ∈ D ) by the v-th probing field (i.e., the v-th illumination of D )
Jv (r) = τ (r ) ∑ Ev,p (r) u p (5)

p={ x, y, z}
that models the scattering profile of D and it is defined as the v-th (v = 1, ..., V) equivalent source
radiating in the background medium an electromagnetic field equal to the v-th (v = 1, ..., V) scattered
field Esv (r). Furthermore, the v-th (v = 1, ..., V) incident field Eiv (r) complies with the so-called
integral state equation
# $ # $
Eiv (r) = Ev (r) + ω 2 ε 0 μ0 Jv r · G r, r dr (6)
D
within the investigation domain (r ∈ D ).

To numerically deal with the inverse problem at hand, a set of N 3D rectangular pulse functions
(
(n) 1 if r ∈ D (n)
Ψ (r) = ; n = 1, ..., N (7)
0 otherwise
defined in a set of N cubic (also called voxels) sub-domains (D = ∪nN=1 D (n) —Figure 1) is adopted
for yielding the following piece-wise constant representation of the v-th (v = 1, ..., V) unknown
contrast source
N
(n)
Jv (r ) = ∑ ∑ Jv,p Ψ(n) (r) u p (8)
p={ x, y, z} n=1
(n)
where Jv,p = Jv,p (rn ) (p = { x, y, z}) and rn denotes the barycentre of the n-th (n = 1, ..., N) voxel D (n) .
By sampling the scattered field in M probing locations of Ω (rm ∈ Ω, m = 1, ..., M), it is possible
to rewrite (3) in the following discrete form [1]
N
(n) (mn)
Esv (rm ) = ∑ ∑ ∑ Jv,q G pq u p , (9)
p={ x, y, z} q={ x, y, z} n=1
(mn)
where G pq = − ω 2 ε 0 μ0 D (n) G pq (rm , r ) dr (p, q = { x, y, z}), or in compact matrix notation
as follows T
T
= G ext J v,x , J v,y , J v,z
Esv,x , Esv,y , Esv,z (10)

where .T indicates the transpose operator, Esv,p Ev,p s (r ) ; m = 1, ..., M (Es 1× M ), J
m v,p ∈ C
v,p
(n) ×
Jv,p ; n = 1, ..., N (J v,p ∈ C 1 N ), p = { x, y, z}, and
⎡ ⎤
G ext
xx
G ext
xy
G ext
xz
⎢ ext ⎥
G ext ⎢
⎣ G yx G ext G ext ⎥ (11)
yy yz ⎦
G ext
zx
G ext
zy
ext
G zz
is the external 3D Green’s operator (G ext ∈ C3M×3N ) where the (m, n)-th (m = 1, ..., M; n = 1, ..., N)
(mn)
entry of the ( p, q)-th (p, q = { x, y, z}) sub-matrix, G ext
pq
∈ C M× N , is equal to G pq .
Under the hypothesis that the contrast distribution in D is sparse with respect to the basis at
hand (7), it also turns out that the v-th (v = 1, ...,
V) contrast source
(8) can benumerically modeled
(o ) (o ) (o )
with just 3 × O (O N) non-zero coefficients Jv,x ; Jv,y ; Jv,z ; o = 1, ..., O , O being the number
30
J. Imaging 2019, 5, 19
of sub-domains D (o) ∈ D , o = 1, ..., O, occupied by scatterers (i.e., dielectric discontinuities of the

background medium within the investigation domain).
3. Inversion Method
The matrix relation in (10) is representative of a severely ill-posed problem being (i) ill-defined
since its solution is not-unique due to the presence of non-radiating sources in D that afford a
null/not-measurable field in Ω and (ii) ill-conditioned since the condition number of G ext , η, is large
(η 1). To counteract such a negative feature for yielding stable reconstructions of the EM properties
of the investigation domain from the solution of (10) when processing noisy data, as well, enforcing the
a-priori knowledge of the sparseness of the unknown equivalent source with respect to a suitable basis
turns out to be an effective regularization strategy. Towards this end, the CS formulation based on the
Bayesian theory is adopted to re-formulate the 3D-CSI problem at hand in a probabilistic sense. Such a
choice is done to avoid the need of numerically checking the fulfillment of the restricted isometry
property (RIP) of the 3D Green’s operator (11) as required by deterministic CS solvers. Indeed, whether
such a compliance test is already very hard from a computational viewpoint for small-scale problems,
it becomes computationally unfeasible in 3D scattering scenarios since exponentially heavier than for
2D cases.
In order to apply computationally-efficient BCS solvers, (10) is first rewritten as a larger real-valued
linear problem
E v,p = G J v (12)
p
T T
where E v,p R Esv,p , I Esv,p and J v R J v,p ; p = { x, y, z} , I J v,p ; p = { x, y, z}
are the p-th (p = { x, y, z}) data vector, E v,p ∈ R2M×1 , and the v-th (v = 1, ..., V) unknown source
vector, J v ∈ R6N ×1 , respectively, while the Green’s matrix operator, G ∈ R2M×6N , is given by
p
⎡ ⎤
R G ext G ext G ext −I G ext G ext G ext
G ⎣ ⎦
px py pz px py pz
(13)
p I G ext
px
G ext
py
G ext
pz R G ext
px
G ext
py
G ext
pz
R ( . ) and I ( . ) being the real and the imaginary part function, respectively. According to such a
description, the original inverse scattering problem can be then formulated as follows
3D-CSI BCS-Based Problem Formulation—Starting from the measurement of E v,p (v =

1, ..., V, p = { x, y, z}), and the knowledge of G (p = { x, y, z}), determine the sparsest
p
(t)
guess of J v , J)v = J)v ; t = 1, ..., 6 × N as the maximum a-posteriori probability (MAP)
estimate ( *

J)v = arg max P J v E v,p (14)
Jv

(t)
provided that the support of J)v , Sv = t ∈ [1, ..., 6 × N ]| J)v = 0 (v = 1, ..., V) is the
same for all V different illuminations (i.e., S1 = S2 = ... = SV ).
In order to enforce the physical correlation existing among the V solutions of (14) as well as
between the unknown contrast sources and the vectorial components of the known scattered field, a
customized version of the MT-BCS solver [53] is adopted by setting the number of parallel “tasks” to
H = (3 × V ) (owing to the 3D nature of the problem). More in detail, (14) is firstly reformulated in the
Bayesian MT framework as [29,53]

⎧ ⎡ # $ ⎤⎫
⎨
P E v,p J v P J v ⎬
J)v = arg max ⎣ ⎦ (15)
⎩ Jv P E v,p ⎭
31
J. Imaging 2019, 5, 19
# $
starting from the definition of a shared prior P J v that statistically links the 3 × V parallel problem
unknowns as [29,53]
# $
P J v = P J v ψ dψ (16)

where ψ (ψ ψ(t) ; t = 1, ..., 6 × N ) is the set of shared—among the V-views and the three Cartesian
scattered
field components—hyper-parameters. By assuming a hierarchical Gaussian model for

P J ψ and substituting (16) in (15) [29,53], the MT-BCS v-th (v = 1, ..., V) source term is then
v
given after simple manipulations [29] by the following expression
−1
1
J)v = ∑ GTG + A G T E v,p (17)
3 p={ x, y, z}
p p p

where A = diag ψ̂ and ψ̂ ψ̂(t) ; t = 1, ..., 6 × N . These latter are determined with a
computationally-efficient relevant vector machine (RVM) [54] by solving the following maximum
marginal likelihood problem
ψ̂ = arg maxψ L ψ |α, β (18)

L ψ |α, β being the logarithmic marginal likelihood function for the fully-vectorial problem given by

L ψ |α, β = − 12 ∑V v=1 ∑ p={ x, y, z} 2 ( M + α ) log E v,p
T
−1 *
(19)

I + G A −1 G T E v,p + 2β + log I + G A−1 G T
p p p p
where I is the identity matrix, while α and β are user-defined control parameters. Once J)v (v = 1, ..., V)
has been estimated through (17), the contrast function, τ) (r) = ∑nN=1 τ)(n) Ψ(n) (r), isretrieved by
computing the corresponding expansion coefficient vector τ) = τ)(n) ∈ C; n = 1, ..., N , whose n-th
(n = 1, ..., N) entry is given by
V )(n)
Jv,p
τ)(n) = ∑ ∑ (20)
)
v=1 3 × V × Ev,p (rn )
p={ x, y, z}
where E )v,p (rn ) is the p-th (p = { x, y, z}) component of the v-th (v = 1, ..., V) total field in the n-th
voxel (rn ∈ D (n) ) (n = 1, ..., N) yielded from the following field relation
N # $
)v,p (rn ) = Ev,p (n)
E i
(rn ) − ω 2 ε 0 μ 0 ∑ ∑ )Jv,q D (n)
G pq rn , r dr , (21)
q={ x, y, z} n=1
(n)
while the coefficient )
Jv,p (n = 1, ..., N; v = 1, ..., V; p = { x, y, z}) is derived from J)v (17) according to
the following mapping
⎧ (n) ( n +3× N )
⎪ J)v + jJ)v
⎨ if p = x
)( n ) (n+ N ) ( n +4× N )
Jv,p = )
Jv + jJ)v if p = y . (22)
⎪
⎩ )(n+2N ) ( n +5× N )
Jv + jJv ) if p = z
4. Numerical Assessment
In this Section, representative results from a wide numerical analysis are presented and discussed
to assess the reconstruction capabilities and the robustness of the proposed 3D-MI approach. Moreover,
some practical guidelines for the optimal setting of its key calibration parameters will be provided for
helping the interested users. Towards this end, the following reference scenario has been considered.
32
J. Imaging 2019, 5, 19
A cubic volume of side L = 1.25 [λ], λ being the free-space wavelength, has been chosen as
investigation domain D and it has been probed by V = 48 plane-waves impinging from the V
# $
angular directions (θv , ϕv ) = π2 , (v − 1) 2π
V , v = 1, ..., V. The scattered field has been collected in
M = 48 locations ⎧
⎪
⎪ xm = ρ cos m + 1 − κm 3 M
M 6π
⎨
ym = ρ sin m + 1 − κm M 6π ; m = 1, ..., M (23)
⎪
⎪ 3 M
⎩
zm = L2 (κm − 1)
. /
3( m −1)
where κm = M (· being the floor operator) and ρ = 3.2 [λ]. To avoid the so-called
inverse crime [1], two different voxel-based discretization of D have been considered for the inverse
(N = 10 × 10 × 10) and the forward (N f wd = 20 × 20 × 20) problem, respectively. As for a quantitative
evaluation of the reconstructions, the following error metric has been computed

(n) (n)
1
NR)
τ − τ
NR n∑
ξR (24)
τ (n) + 1
=1
for the whole investigation domain (R = tot ⇒ Ntot = N), the scatterers support (R = int ⇒ Nint = O),
and its complementary region (It is worthwhile to remark that the scatterer support is not an a-priori
information exploited during the inversion, but it is rather employed in the post-processing phase only
to compute the error metrics (24).) [R = ext ⇒ Next = ( N − O)], τ (n) and τ)(n) being the actual and the
estimated contrast of the n-th (n = 1, ..., N) voxel in D , respectively.
The first test case is concerned with the retrieval of a single (K = 1, K being the number of
disconnected scattering regions lying in D ) off-centered scatterer composed by a single-voxel (O = 1)
of side x = y = z = 0.125 [λ] with contrast τ = 1.0 (Figure 2a). First, a sensitivity analysis for setting
the optimal trade-off values of the control parameters of the MT-BCS solver [i.e., α and β in (19)] has
been carried out. More specifically, the reconstruction error (24) has been computed by varying the

values of the calibration coefficients within the ranges α ∈ 1, 102 and β ∈ 10−5 , 10−2 for different
signal-to-noise ratios (SNRs). Figure 2b shows that the reconstruction “quality” is almost insensitive to
the choice of α when SNR > 5 [dB] and it mainly depends on the noise level.
Differently, significant degradations occur when letting β > 5 × 10−4 whatever the SNR
ς(opt) SNR dSNR
/
(Figure 2c). By computing the optimal trade-off value as ς(opt) SNR dSNR , ς(opt)
SNR SNR
arg {minς ( ξ tot | SNR)} being the best value of ς = {α, β} at a fixed SNR ( · standing for “evaluated
at”), it turned out that α(opt) = 10 and β(opt) = 10−4 . These thresholds have then been used throughout
the whole numerical validation. A pictorial view of the 3D MT-BCS reconstructions when processing
different noisy data is shown in Figure 3.
As it can be observed, the object support is always faithfully detected regardless of the amount
of noise. There are only slight deviations from the actual contrast value (e.g., τ)SNR=50 [dB] =
0.81—Figure 3a; τ)SNR=5 [dB] = 0.77—Figure 3d) as quantitatively indicated by the corresponding
values of the error index ξ tot being equal to ξ tot SNR=50 [dB] = 9.58 × 10−5 (Figure 3a) and
ξ tot SNR=5 [dB] = 1.16 × 10−4 (Figure 3d) at the highest and lowest SNRs, respectively.
In order to prove that such results are not due to a customization of the MT-BCS solver to
the specific scenario at hand, a set of W = 100 inversions has been performed by randomly
changing the position of the same target within the investigation domain. The results of such a
min min
statistical analysis are summarized in Table 1 where the minimum (ξ tot w=1, ..., W { ξ tot,w }),
avg
the maximum (ξ tot maxw=1, ..., W {ξ tot,w }), the average (ξ tot W ∑w=1 ξ tot,w ), and the variance
max 1 W

var 1 avg 2
W ∑w=1 ξ tot,w − ξ tot
W
[ξ tot ] values of the total error among the W experiments are reported.
avg
Whatever the SNR, the scatterer retrieval is very accurate since on average ξ tot < 1.0 × 10−4 (Table 1)
var , is very small (i.e., ξ var < 5.0 × 10−11 —Table 1).
and the variance of the reconstruction error, ξ tot tot
33
J. Imaging 2019, 5, 19
Such a positive outcome holds true when dealing with stronger scatterers, as well. Thanks to the
CSI formulation, which avoids any physical assumption/approximation in the application of the CS to
the 3D scattering equations (Section 2), faithful data inversions are yielded also when increasing the
actual contrast well-beyond the value of the first example. As a proof, the plots of ξ tot (Figure 4a), ξ int
(Figure 4b), and ξ ext (Figure 4c) versus τ and the noise level indicate that D can be carefully imaged
when the scatterer is at least up to τ = 4.0.
As expected, the reconstruction progressively degrades when higher and higher contrasts are at
hand especially when processing highly blurred data. Figure 5 shows the actual and the retrieved
volumetric distributions when τ = 4.0 and SNR ≤ 10 [dB]. As it can be inferred, some artifacts are
present only in very harsh conditions (Figure 5c—SNR = 5 [dB]) when the external error (Figure 4c)
=1.0 τ =4.0
grows from ξ ext τSNR =5 [dB] = 0.0 (Figure 3d) to ξ ext SNR=5 [dB] = 8.96 × 10
−6 (Figure 5c). Nevertheless,
it is worth pointing out that these inaccuracies correspond to voxels with a very low contrast (i.e.,
τ) < 3 × 10−4 —Figure 5c) that can be easily filtered out by a simple (i.e., the result is not so sensitive to
the choice of the threshold value τth ) thresholding (Figure 5d—τth = 10−3 ) and the scatterer is always
correctly localized (Figure 5c as well as Figure 5d vs. Figure 5a).
K=O=1, τ=1.0
R[τ(r)] [Arbitrary Unit]

1
0.6 0.8
0.6
0.3 0.4
0.2
z/λ 0
0
−0.3
0.6
0.3
−0.6 0 y/λ
−0.3
−0.6 −0.3 0 0.3 0.6 −0.6
x/λ
(a)
K=O=1, τ=1.0 - β=β(opt)=10-4 K=O=1, τ=1.0 - α=α(opt)=10
8 12
ξtot [Arbitrary Unit] (×10 )
SNR=50 [dB]
ξtot [Arbitrary Unit] (×10-4)
-4
7 10
SNR=20 [dB]
6 (opt)
α SNR=10 [dB] 8 β(opt)
5
SNR=5 [dB]
4 6
SNR=50 [dB]
3 4 SNR=20 [dB]
2
2 SNR=10 [dB]
1
SNR=5 [dB]
0 0
-5 -4 -3 -2
100 101 102 10 10 10 10
Control Parameter α [Arbitrary Unit] Control Parameter β [Arbitrary Unit]
(b) (c)
Figure 2. Sensitivity Analysis (K = 1, O = 1, τ = 1.0, SNR ∈ [5, 50] [dB])—Actual contrast

function (a). Behavior of the total error, ξ tot , versus the MT-BCS control parameters: (b) α ∈ 1, 102
( ) −
−5 −
( )
(β = β opt = 10 ) and (c) β ∈ 10 , 10
4 2 (α = α opt = 10).
34
J. Imaging 2019, 5, 19
K=O=1, τ=1.0 − SNR=50 [dB] K=O=1, τ=1.0 − SNR=20 [dB]

1 1
0.6 0.8 0.6 0.8
0.6 0.6
0.3 0.4 0.3 0.4
0.2 0.2
z/λ 0 z/λ 0
0 0
^
−0.3 −0.3
0.6 0.6
0.3 0.3
−0.6 0 y/λ −0.6 0 y/λ
−0.3 −0.3
−0.6 −0.3 0 0.3 0.6 −0.6 −0.6 −0.3 0 0.3 0.6 −0.6
x/λ x/λ
(a) (b)

1 1
0.6 0.8 0.6 0.8
0.6 0.6
0.3 0.4 0.3 0.4
0.2 0.2
z/λ 0 z/λ 0
0 0
^
^
−0.3 −0.3
0.6 0.6
0.3 0.3
−0.6 0 y/λ −0.6 0 y/λ
−0.3 −0.3
−0.6 −0.3 0 0.3 0.6 −0.6 −0.6 −0.3 0 0.3 0.6 −0.6
x/λ x/λ
(c) (d)
Figure 3. Numerical Assessment (K = 1, O = 1, τ = 1.0) - MT-BCS reconstructions when processing the

scattering data with (a) SNR = 50 [dB] ()τ = 0.81); (b) SNR = 20 [dB] ()
τ = 0.79); (c) SNR = 10 [dB]
τ = 0.78), and (d) SNR = 5 [dB] ()
() τ = 0.77).
Table 1. Numerical Assessment (K = 1, O = 1, τ = 1.0, W = 100, MT-BCS)—Total error statistics.

avg
SN R [dB] ξ min
tot ξ max
tot ξ tot ξ var
tot
50 7.99 × 10−5 9.69 × 10−5 9.10 × 10−5 1.11 × 10−11
20 8.37 × 10−5 9.75 × 10−5 9.12 × 10−5 1.21 × 10−11
10 8.66 × 10−5 9.86 × 10−5 9.27 × 10−5 1.87 × 10−11
5 8.68 × 10−5 1.21 × 10−4 9.78 × 10−5 4.66 × 10−11
K=O=1 K=O=1 K=O=1

50 10
ξext [Arbitrary Unit] (×10-6)
50 1.6 50 1.6
ξint [Arbitrary Unit] (×10 )

-4
-1
45 45 45
1.5 1.5 8
40 40 40
1.4 1.4 35
SNR [dB]
35 35
SNR [dB]
SNR [dB]
6
30 1.3 30 1.3 30
25 1.2 25 1.2 25
4
20 20 20
1.1 1.1
15 15 15 2
10 1 10 1 10
5 0.9 5 0.9 5 0
1 1.5 2 2.5 3 3.5 4 1 1.5 2 2.5 3 3.5 4 1 1.5 2 2.5 3 3.5 4
τ [Arbitrary Unit] τ [Arbitrary Unit] τ [Arbitrary Unit]
(a) (b) (c)
Figure 4. Numerical Assessment (K = 1, O = 1, τ ∈ [1.0, 4.0], SNR ∈ [5, 50] [dB])—Behavior of the
(a) total (ξ tot ); (b) internal (ξ int ); and (c) external (ξ ext ) reconstruction errors when processing the
scattering data with the MT-BCS.
35
J. Imaging 2019, 5, 19
K=O=1, τ=4.0 K=O=1, τ=4.0 − SNR=10 [dB]

5 5
0.6 4 0.6 4
3 3
0.3 2 0.3 2
1 1
z/λ 0 z/λ 0
0 0
^
−0.3 −0.3
0.6 0.6
0.3 0.3
−0.6 0 y/λ −0.6 0 y/λ
−0.3 −0.3
−0.6 −0.3 0 0.3 0.6 −0.6 −0.6 −0.3 0 0.3 0.6 −0.6
x/λ x/λ
(a) (b)

5 5
0.6 4 0.6 4
3 3
0.3 2 0.3 2
1 1
z/λ 0 z/λ 0
0 0
^
^
−0.3 −0.3
0.6 0.6
0.3 0.3
−0.6 0 y/λ −0.6 0 y/λ
−0.3 −0.3
−0.6 −0.3 0 0.3 0.6 −0.6 −0.6 −0.3 0 0.3 0.6 −0.6
x/λ x/λ
(c) (d)
Figure 5. Numerical Assessment (K = 1, O = 1, τ = 4.0)—(a) Actual contrast function and (b,c)
MT-BCS reconstructions when processing the scattering data with (b) SNR = 10 [dB] () τ = 3.40),
τ = 3.28), and (d) SNR = 5 [dB] (filtered τth = 10−3 ) ()
(c) SNR = 5 [dB] (unfiltered) () τ = 3.28).
The previous examples have dealt with real-valued target contrasts. However, the imaginary part
of τ is not enforced to be zero during the inversion (see Section 3). In order to assess the reliability
of the MT-BCS also when complex contrast values are at hand, the retrieval of a K = 1, O = 1
target with τ = 1.0 − 0.6j (Figure 6a,b) has been considered next. The analysis of the retrieved
profiles (Figure 6) shows that the location, size, and contrast of the scatterer is correctly retrieved
by the proposed imaging strategy both in low (e.g., SNR = 50 [dB]—Figure 6c,d) and in high noise
conditions (e.g., SNR = 5 [dB]—Figure 6e,f), as shown by the corresponding error indexes (e.g.,
ξ tot SNR=50 [dB] = 8.63 × 10−5 —Figure 6c,d; ξ tot SNR=5 [dB] = 1.06 × 10−4 —Figure 6e,f).
Being assessed the effectiveness of the proposed approach in reconstructing the sparsest (O = 1)
actual profile, let us now analyze its performance for scenarios exhibiting a lower sparsity order with
respect to the selected voxel basis (7). Towards this end, a second single-voxel scatterer has been
added to the configuration of the first test case (Figure 2a), but in a different position in D (K = O = 2,
τ = 1.0—Figure 7a).
36
J. Imaging 2019, 5, 19
K=O=1, τ=1.0−j0.6 K=O=1, τ=1.0−j0.6
I[τ(r)] [Arbitrary Unit]

1 0
0.6 0.8 0.6
0.6 −0.2
0.3 0.4 0.3 −0.4
0.2
z/λ 0 z/λ 0
0 −0.6
−0.3 −0.3
0.6 0.6
0.3 0.3
−0.6 0 y/λ −0.6 0 y/λ
−0.3 −0.3
−0.6 −0.3 0 0.3 0.6 −0.6 −0.6 −0.3 0 0.3 0.6 −0.6
x/λ x/λ
(a) (b)
K=O=1, τ=1.0−j0.6 − SNR=50 [dB] K=O=1, τ=1.0−j0.6 − SNR=50 [dB]

1 0
0.6 0.8 0.6
0.6 −0.2
0.3 0.4 0.3 −0.4
0.2
z/λ 0 z/λ 0
0 −0.6
^
^
−0.3 −0.3
0.6 0.6
0.3 0.3
−0.6 0 y/λ −0.6 0 y/λ
−0.3 −0.3
−0.6 −0.3 0 0.3 0.6 −0.6 −0.6 −0.3 0 0.3 0.6 −0.6
x/λ x/λ
(c) (d)
K=O=1, τ=1.0−j0.6 − SNR=5 [dB] K=O=1, τ=1.0−j0.6 − SNR=5 [dB]
1 0
0.6 0.8 0.6
0.6 −0.2
0.3 0.4 0.3 −0.4
0.2
z/λ 0 z/λ 0
0 −0.6
^
−0.3 −0.3
0.6 0.6
0.3 0.3
−0.6 0 y/λ −0.6 0 y/λ
−0.3 −0.3
−0.6 −0.3 0 0.3 0.6 −0.6 −0.6 −0.3 0 0.3 0.6 −0.6
x/λ x/λ
(e) (f)
Figure 6. Numerical Assessment (K = 1, O = 1, τ = 1.0 − 0.6j)—Real (a,c,e) and imaginary parts (b,d,f)
of the (a,b) actual contrast function and of the (b–f) MT-BCS reconstructed profiles when processing
the scattering data with (c,d) SNR = 50 [dB] () τ = 0.84 − 0.49j), (e,f) SNR = 5 [dB] ()
τ = 0.81 − 0.45j).
37
J. Imaging 2019, 5, 19
K=O=2, τ=1.0

1
0.6 0.8
0.6
0.3 0.4
0.2
z/λ 0
0
−0.3
0.6
0.3
−0.6 0 y/λ
−0.3
−0.6 −0.3 0 0.3 0.6 −0.6
x/λ
(a)

1 1
0.6 0.8 0.6 0.8
0.6 0.6
0.3 0.4 0.3 0.4
0.2 0.2
z/λ 0 z/λ 0
0 0
^
^
−0.3 −0.3
0.6 0.6
0.3 0.3
−0.6 0 y/λ −0.6 0 y/λ
−0.3 −0.3
−0.6 −0.3 0 0.3 0.6 −0.6 −0.6 −0.3 0 0.3 0.6 −0.6
x/λ x/λ
(b) (c)

1 1
0.6 0.8 0.6 0.8
0.6 0.6
0.3 0.4 0.3 0.4
0.2 0.2
z/λ 0 z/λ 0
0 0
^
−0.3 −0.3
0.6 0.6
0.3 0.3
−0.6 0 y/λ −0.6 0 y/λ
−0.3 −0.3
−0.6 −0.3 0 0.3 0.6 −0.6 −0.6 −0.3 0 0.3 0.6 −0.6
x/λ x/λ
(d) (e)
Figure 7. Numerical Assessment (K = 2, O = 2, τ = 1.0)—(a) Actual contrast function and

MT-BCS reconstructions when processing the scattering data with (b) SNR = 50 [dB] () τmax = 0.82),
(c) SNR = 20 [dB] ()
τmax = 0.84), (d) SNR = 10 [dB] ()
τmax = 0.84), and (e) SNR = 5 [dB] ()
τmax = 0.83).
As it can be inferred from the plots in Figure 7b–e, the MT-BCS is always able to correctly
retrieve both scatterers even from very low-noise data (e.g., ξ tot SNR=5 [dB] = 9.80 × 10−4 —Figure 7e).
However, the same statistical analysis of the single-scatterer/single-voxel case (i.e., randomly changing
var ≥ 4.5 × 10−7
the locations of the scatterers) here results in wider variations of the integral error (i.e., ξ tot
var ≥ 1.11 × 10−11 (Table 1)). Therefore, to better understand how the inversion accuracy
(Table 2) vs. ξ tot
is affected by the positions of the scatterers within the investigation domain, a further set of experiments
has been performed by deterministically changing the relative locations of the K = 2 scatterers, d being
the distance of each one of them from the origin.
38
J. Imaging 2019, 5, 19
Table 2. Numerical Assessment (K = 2, O = 2, τ = 1.0, W = 100, MT-BCS)—Total error statistics.

avg
SN R [dB] ξ min
tot ξ max
tot ξ tot ξ var
tot
50 1.49 × 10−4 2.57 × 10−3 4.60 × 10−4 4.50 × 10−7
20 1.56 × 10−4 2.59 × 10−3 4.63 × 10−4 4.96 × 10−7
10 1.61 × 10−4 2.62 × 10−3 4.72 × 10−4 5.20 × 10−7
5 1.83 × 10−4 2.63 × 10−3 5.30 × 10−4 5.27 × 10−7
√
Some representative test cases are reported in Figure 8: d = dmin = 3
= 0.11 [λ] (Figure 8a,b),
√ 2
d = 0.54 [λ] (Figure 8c,d), and d = dmax = ( L − ) = 0.97 [λ] (Figure 8e,f). As expected, 2
3
the reconstruction worsens when the two objects get closer (Figure 9) until some artifacts appear
SNR=10 [dB]
ξ tot d=0.11 [λ]
when d = dmin [ SNR=10 [dB] 2.41—Figure 8b vs. Figure 8f].
ξ tot d=0.97 [λ]
K=O=2, τ=1.0, d=0.11 [λ] K=O=2, τ=1.0, d=0.11 [λ] − SNR=10 [dB]

1 1
0.6 0.8 0.6 0.8
d 0.6 0.6
0.3 0.4 0.3 0.4
d
0.2 0.2
z/λ 0 z/λ 0
0 0
^
−0.3 −0.3
0.6 0.6
0.3 0.3
−0.6 0 y/λ −0.6 0 y/λ
−0.3 −0.3
−0.6 −0.3 0 0.3 0.6 −0.6 −0.6 −0.3 0 0.3 0.6 −0.6
x/λ x/λ
(a) (b)
1 1
0.6 0.8 0.6 0.8
d 0.6 0.6
0.3 0.4 0.3 0.4
d 0.2 0.2
z/λ 0 z/λ 0
0 0
^
−0.3 −0.3
0.6 0.6
0.3 0.3
−0.6 0 y/λ −0.6 0 y/λ
−0.3 −0.3
−0.6 −0.3 0 0.3 0.6 −0.6 −0.6 −0.3 0 0.3 0.6 −0.6
x/λ x/λ
(c) (d)
Figure 8. Cont.
39
J. Imaging 2019, 5, 19

1 1
0.6 d 0.8 0.6 0.8
0.6 0.6
0.3 0.4 0.3 0.4
0.2 0.2
z/λ 0 d z/λ 0
0 0
^
−0.3 −0.3
0.6 0.6
0.3 0.3
−0.6 0 y/λ −0.6 0 y/λ
−0.3 −0.3
−0.6 −0.3 0 0.3 0.6 −0.6 −0.6 −0.3 0 0.3 0.6 −0.6
x/λ x/λ
(e) (f)
Figure 8. Numerical Assessment (K = 2, O = 2, τ = 1.0, SNR = 10 [dB])—(a,c,e) Actual contrast
function and (b,d,f) MT-BCS reconstructions when the distance of the scatterers from the origin is equal
to (a,b) d = dmin = 0.11 [λ] ()
τmax = 0.73), (c,d) d = 0.54 [λ] ()
τmax = 0.83), and (e,f) d = dmax = 0.97
τmax = 0.84).
[λ] ()
K=O=2, τ=1.0
Total Error, ξtot [Arbitrary Unit] (×10-4)
6
SNR=50 [dB]
5 SNR=20 [dB]
SNR=10 [dB]
4 SNR=5 [dB]
1
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
d/λ
Figure 9. Numerical Assessment (K = 2, O = 2, τ = 1.0, SNR ∈ [5, 50] [dB])—Behavior of the total error,
ξ tot , versus the distance of the scatterers from the origin, d.
Further lowering the sparseness of the actual profile causes a decrement of the CS performance
as proven by the results of the third test case when O = 4 and the K = 4 single-voxel scatterers are
randomly placed in D (Figure 10a).
40
J. Imaging 2019, 5, 19
K=O=4, τ=1.0

1
0.6 0.8
0.6
0.3 0.4
0.2
z/λ 0
0
−0.3
0.6
0.3
−0.6 0 y/λ
−0.3
−0.6 −0.3 0 0.3 0.6 −0.6
x/λ
(a)

1 1
0.6 0.8 0.6 0.8
0.6 0.6
0.3 0.4 0.3 0.4
0.2 0.2
z/λ 0 z/λ 0
0 0
^
^
−0.3 −0.3
0.6 0.6
0.3 0.3
−0.6 0 y/λ −0.6 0 y/λ
−0.3 −0.3
−0.6 −0.3 0 0.3 0.6 −0.6 −0.6 −0.3 0 0.3 0.6 −0.6
x/λ x/λ
(b) (c)

1 1
0.6 0.8 0.6 0.8
0.6 0.6
0.3 0.4 0.3 0.4
0.2 0.2
z/λ 0 z/λ 0
0 0
^
−0.3 −0.3
0.6 0.6
0.3 0.3
−0.6 0 y/λ −0.6 0 y/λ
−0.3 −0.3
−0.6 −0.3 0 0.3 0.6 −0.6 −0.6 −0.3 0 0.3 0.6 −0.6
x/λ x/λ
(d) (e)
Figure 10. Numerical Assessment (K = 4, O = 4, τ = 1.0)—(a) Actual contrast function and MT-BCS
reconstructions when processing the scattering data characterized by (b) SNR = 50 [dB] ()τmax = 0.84),
(c) SNR = 20 [dB] ()
τmax = 0.84), (d) SNR = 10 [dB] ()τmax = 0.80), and (e) SNR = 5 [dB] ()
τmax = 0.70).
Unavoidably, the MT-BCS gets worse and undesired artifacts occur also when processing
low-noisy data (e.g., ξ tot SNR=50 [dB] = 2.37 × 10−3 —Figure 10b), but still without preventing the
possibility to correctly identify K = 4 separate objects (Figure 10b–e). In addition, in this case, changing
the distance d of the objects with respect to the origin influences the quality of the retrieval as pictorially
shown in Figure 11 (SNR = 10 [dB]).
41
J. Imaging 2019, 5, 19

1 1
0.6 0.8 0.6 0.8
d 0.6 0.6
0.3 0.4 0.3 0.4
d d 0.2 0.2
z/λ 0 z/λ 0
d 0 0
^
−0.3 −0.3
0.6 0.6
0.3 0.3
−0.6 0 y/λ −0.6 0 y/λ
−0.3 −0.3
−0.6 −0.3 0 0.3 0.6 −0.6 −0.6 −0.3 0 0.3 0.6 −0.6
x/λ x/λ
(a) (b)
K=O=4, τ=1.0, d=0.54 [λ] R[τ(r)] [Arbitrary Unit]
K=O=4, τ=1.0, d=0.54 [λ] − SNR=10 [dB]

1 1
0.6 0.8 0.6 0.8
0.6 0.6
d d
0.3 0.4 0.3 0.4
d 0.2 0.2
z/λ 0 z/λ 0
d 0 0
^
−0.3 −0.3
0.6 0.6
0.3 0.3
−0.6 0 y/λ −0.6 0 y/λ
−0.3 −0.3
−0.6 −0.3 0 0.3 0.6 −0.6 −0.6 −0.3 0 0.3 0.6 −0.6
x/λ x/λ
(c) (d)
1 1 R[τ(r)] [Arbitrary Unit]

0.6 0.8 0.6 0.8
d d 0.6 0.6
0.3 0.4 0.3 0.4
d 0.2 0.2
z/λ 0 z/λ 0
0 0
^
d
−0.3 −0.3
0.6 0.6
0.3 0.3
−0.6 0 y/λ −0.6 0 y/λ
−0.3 −0.3
−0.6 −0.3 0 0.3 0.6 −0.6 −0.6 −0.3 0 0.3 0.6 −0.6
x/λ x/λ
(e) (f)
Figure 11. Numerical Assessment (K = 4, O = 4, τ = 1.0, SNR = 10 [dB])—(a,c,e) Actual contrast
function and (b,d,f) MT-BCS reconstructions when the objects distance from the origin is (a,b)
d = dmin = 0.11 [λ] ()
τmax = 0.73), (c,d) d = 0.54 [λ] ()
τmax = 0.70), and (e,f) d = dmax = 0.97 [λ]
τmax = 0.71).
()
Indeed, it turns out that more accurate images of D are yielded when the scatterers are far (i.e.,
d ↑ ⇒ ξ tot ↓) as confirmed by the plot of ξ tot vs. d (Figure 12a). However, it cannot be neglected that a
simple filtering (Figure 12b—τth = 10−3 ) allows one to clearly resolve the scatterer support even in the
most critical case (Figure 11a).
42
J. Imaging 2019, 5, 19
K=O=4, τ=1.0
Total Error, ξtot [Arbitrary Unit] (×10 )

-3
4
SNR=50 [dB]
3.5 SNR=20 [dB]
SNR=10 [dB]
3
SNR=5 [dB]
2.5
1.5
1
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
d/λ
(a)
K=O=4, τ=1.0, d=0.11 [λ] − SNR=10 [dB]

1
0.6 0.8
0.6
0.3 0.4
0.2
z/λ 0
0 ^
−0.3
0.6
0.3
−0.6 0 y/λ
−0.3
−0.6 −0.3 0 0.3 0.6 −0.6
x/λ
(b)
Figure 12. Numerical Assessment (K = 4, O = 4, τ = 1.0, SNR ∈ [5, 50] [dB])—(a) Behavior of the total
error, ξ tot , as a function of the objects distance from the origin, d and (b) filtered (τth = 10−3 ) MT-BCS
reconstruction when the objects distance from the origin is d = dmin = 0.11 [λ] () τmax = 0.73).
The next numerical test is devoted to validate the MT-BCS in a more complex 3D imaging
scenario concerned with non-uniformly
shaped scatterers. More specifically,
K = 3 objects sized
1x = 1y = 1z = 0.125 [λ], 2x , 2y , 2z = [0.125, 0.25, 0.125] [λ], and 3x , 3y , 3z = [0.25, 0.25, 0.125]
[λ] (O = 7) with contrast τ = 1.0 have been imaged (Figure 13a).
Despite the increased complexity and the reduced intrinsic-sparsity order of the scattering
configuration, faithful reconstructions of the 3D contrast distribution are yielded (Figure 13b–e)
even though there is an over-estimation of the scatterers supports when highly-blurred data are at
hand (e.g., Figure 13b vs. Figure 13e) and, consequently, an increase of the reconstruction error with
ξ tot SNR=5 [dB]
the noise level (e.g., ≈ 2.04). Moreover, the previous considerations regarding the fidelity
ξ tot SNR=50 [dB]
of the proposed strategy with respect to the target contrast are confirmed also when dealing with
non-uniformly shaped scatterers (Figure 14).
Finally, it is interesting to underline the advantage of using the MT extension of the BCS-based
approach for solving the 3D-CSI inversion problem instead of its “naive” single-task implementation
(ST-BCS [23]) that does not impose any physical correlation in solving (14). Towards this end, let us
consider the benchmark O = 6 voxel arrangement in Figure 15a with two (K = 2) “L-shaped”
homogeneous (τ = 2.0) objects.
43
J. Imaging 2019, 5, 19
K=3, O=7, τ=1.0

1.4
1.2
0.6 1
0.8
0.3 0.6
0.4
z/λ 0 0.2
0
−0.3
0.6
0.3
−0.6 0 y/λ
−0.3
−0.6 −0.3 0 0.3 0.6 −0.6
x/λ
(a)
K=3, O=7, τ=1.0 − SNR=50 [dB] K=3, O=7, τ=1.0 − SNR=20 [dB]

1.4 1.4
1.2 1.2
0.6 1 0.6 1
0.8 0.8
0.3 0.6 0.3 0.6
0.4 0.4
z/λ 0 0.2 z/λ 0 0.2
0 0
^
^
−0.3 −0.3
0.6 0.6
0.3 0.3
−0.6 0 y/λ −0.6 0 y/λ
−0.3 −0.3
−0.6 −0.3 0 0.3 0.6 −0.6 −0.6 −0.3 0 0.3 0.6 −0.6
x/λ x/λ
(b) (c)
K=3, O=7, τ=1.0 − SNR=10 [dB] K=3, O=7, τ=1.0 − SNR=5 [dB]

1.4 1.4
1.2 1.2
0.6 1 0.6 1
0.8 0.8
0.3 0.6 0.3 0.6
0.4 0.4
z/λ 0 0.2 z/λ 0 0.2
0 0
^
−0.3 −0.3
0.6 0.6
0.3 0.3
−0.6 0 y/λ −0.6 0 y/λ
−0.3 −0.3
−0.6 −0.3 0 0.3 0.6 −0.6 −0.6 −0.3 0 0.3 0.6 −0.6
x/λ x/λ
(d) (e)
Figure 13. Numerical Assessment (K = 3, O = 7, τ = 1.0)—(a) Actual contrast function and

MT-BCS reconstructions when (b) SNR = 50 [dB] () τmax = 1.08); (c) SNR = 20 [dB] ()
τmax = 1.05);
(d) SNR = 10 [dB] ()
τmax = 1.09); and (e) SNR = 5 [dB] ()
τmax = 1.37).
For illustrative purposes, the reconstructions with the MT-BCS (Figure 15b,c) and the ST-BCS
(Figure 15d,e) when processing two different sets of noisy data [SNR = 20 [dB]—Figure 15b,d;
SNR = 10 [dB]—Figure 15c,e] are shown. As it can be visually inferred, the MT strategy turns out to
be more effective than the ST one in both locating and shaping the non-connected scattering regions
as well as in estimating the actual contrast value. Such an inference is quantitatively confirmed by
SNR=20 [dB] SNR=10 [dB]
ξ tot ST − BCS ξ tot ST − BCS
the error indexes since SNR=20 [dB] ≈ 33.36 (Figure 15b vs. Figure 15d) and SNR=10 [dB] ≈ 7.04
ξ tot MT − BCS ξ tot MT − BCS
44
J. Imaging 2019, 5, 19
(Figure 15c vs. Figure 15e) (Table 3). Similar outcomes can also be drawn when changing the contrast
of the scatterers as indicated by the behaviour of the reconstruction error ξ tot versus τ in Figure 16.
K=3, O=7
1
SNR=50 [dB]
ξtot [Arbitraty Unit] (×10-2)
0.8 SNR=20 [dB]

SNR=10 [dB]
0.6 SNR=5 [dB]
0.4
0.2
0
1 1.5 2 2.5 3 3.5 4
τ [Arbitrary Unit]
Figure 14. Numerical Assessment (K = 3, O = 7, SNR ∈ [5, 50] [dB])—Behavior of the total
reconstruction error (ξ tot ) when processing the scattering data with the MT-BCS.
K=2, O=6, τ=2.0

2
0.6 1.6
1.2
0.3 0.8
0.4
z/λ 0
0
−0.3
0.6
0.3
−0.6 0 y/λ
−0.3
−0.6 −0.3 0 0.3 0.6 −0.6
x/λ
(a)
K=2, O=6, τ=2.0 − SNR=20 [dB] − MT−BCS K=2, O=6, τ=2.0 − SNR=10 [dB] − MT−BCS
2 2
0.6 1.6 0.6 1.6
1.2 1.2
0.3 0.8 0.3 0.8
0.4 0.4
z/λ 0 z/λ 0
0 0
^
^
MT − BCS
−0.3 −0.3
0.6 0.6
0.3 0.3
−0.6 0 y/λ −0.6 0 y/λ
−0.3 −0.3
−0.6 −0.3 0 0.3 0.6 −0.6 −0.6 −0.3 0 0.3 0.6 −0.6
x/λ x/λ
(b) (c)
Figure 15. Cont.
45
J. Imaging 2019, 5, 19
K=2, O=6, τ=2.0 − SNR=20 [dB] − ST−BCS K=2, O=6, τ=2.0 − SNR=10 [dB] − ST−BCS

2 2
0.6 1.6 0.6 1.6
1.2 1.2
0.3 0.8 0.3 0.8
0.4 0.4
z/λ 0 z/λ 0
0 0
^
−0.3 −0.3
ST − BCS
0.6 0.6
0.3 0.3
−0.6 0 y/λ −0.6 0 y/λ
−0.3 −0.3
−0.6 −0.3 0 0.3 0.6 −0.6 −0.6 −0.3 0 0.3 0.6 −0.6
x/λ x/λ
(d) (e)
Figure 15. Comparative Assessment (K = 2, O = 6, τ = 2.0)—(a) Actual contrast function and retrieved
solutions by the (b,c) MT-BCS and (d,e) ST-BCS when processing noisy data at (b,d) SNR = 20 [dB]
MT − BCS = 1.96, τ
τmax
() ST − BCS = 0.33) and (c,e) SNR = 10 [dB] ()
)max MT − BCS = 1.68, τ
τmax ST − BCS = 0.36).
)max
K=2, O=6
2
-2
1.6
1.2
0.8
0.4
0
1 1.5 2 2.5 3 3.5 4
τ [Arbitrary Unit]
MT-BCS - SNR=20 [dB] MT-BCS - SNR=10 [dB]

ST-BCS - SNR=20 [dB] ST-BCS - SNR=10 [dB]
CG - SNR=20 [dB] CG - SNR=10 [dB]
Figure 16. Comparative Assessment (K = 2, O = 6, τ ∈ [1.0, 4.0], SNR ∈ [10, 20] [dB])—Behavior of the
total error, ξ tot , as a function of the object contrast, τ, when processing the scattering data with the
MT-BCS, the ST-BCS, and the CG methods.
To conclude the numerical assessment of the reconstruction capabilities of the MT-BCS, it has
been compared with a competitive non-CS state-of-the-art approach. Towards this end, a deterministic
CG-based inversion tool—still based on a CSI formulation of the scattering problem—has been applied
to the same scenario in Figure 15a. The retrieved images of the dielectric profile of the investigation
domain are shown in Figure 17a (SNR = 20 [dB]) and Figure 17b (SNR = 10 [dB]).
46
J. Imaging 2019, 5, 19
K=2, O=6, τ=2.0 − SNR=20 [dB] − CG K=2, O=6, τ=2.0 − SNR=10 [dB] − CG

2 2
0.6 1.6 0.6 1.6
1.2 1.2
0.3 0.8 0.3 0.8
0.4 0.4
z/λ 0 z/λ 0
0 0
^
−0.3 −0.3
0.6 0.6
0.3 0.3
−0.6 0 y/λ −0.6 0 y/λ
−0.3 −0.3
−0.6 −0.3 0 0.3 0.6 −0.6 −0.6 −0.3 0 0.3 0.6 −0.6
x/λ x/λ
(a) (b)
Figure 17. Comparative Assessment (K = 2, O = 6, τ = 2.0)—CG reconstructions when processing noisy
data characterized by (a) SNR = 20 [dB] () CG = 0.33) and (b) SNR = 10 [dB] ()
τmax CG = 0.36).
τmax
Without imposing sparseness priors, only the presence of K = 2 scatterers lying in D can be
deduced, but their contrasts are strongly under-estimated and their supports/shapes are unreliably
predicted. Comparatively, the MT-BCS enables a reduction of the reconstruction error of about
SNR=20 [dB] SNR=10 [dB]
ξ tot CG ξ tot CG
SNR=20 [dB] ≈ 43.75 (Figure 17a vs. Figure 15b—Table 3) and SNR=10 [dB] ≈ 8.92 (Figure 17b vs.
ξ tot MT − BCS ξ tot MT − BCS
Figure 15c—Table 3) times. Such conclusions are not limited to the contrast τ = 1.0, but they also arise
for stronger scatterers as detailed by the plot of ξ tot versus τ in Figure 16.
As for the computational efficiency, the MT-constrained exploitation of a sparseness promoting
inversion technique allows a non-negligible reduction of the computational time, Δt, when processing
3D scattering data (For the sake of fairness, all the computational times refer to non-optimized Matlab
codes executed on a single-core laptop running at 2.20 GHz). Indeed, the MT-BCS is not only faster
Δt MT − BCS
than the ST-BCS, thanks to the joint processing of the data (i.e., ΔtST − BCS
≈ 0.52—Table 3), but it also
Δt MT − BCS
overcomes the CG speed (i.e., ΔtCG
≈ 0.032—Table 3).
Table 3. Comparative Assessment (K = 2, O = 6, τ = 2.0, SNR ∈ [10, 20] [dB])—Inversion

performance indexes.
SN R = 20 [dB] SN R = 10 [dB] Δt
Method
ξ tot ξ int ξ ext ξ tot ξ int ξ ext [s]
MT − BCS 2.56 × 10−4 4.07 × 10−2 0.00 1.39 × 10−3 1.53 × 10−1 3.94 × 10−4 15.12
ST − BCS 8.54 × 10−3 6.28 × 10−1 2.95 × 10−3 9.79 × 10−3 8.16 × 10−1 3.06 × 10−3 29.08
CG 1.12 × 10−2 5.78 × 10−1 7.79 × 10−3 1.24 × 10−2 5.91 × 10−1 8.92 × 10−3 4.66 × 102
5. Conclusions
An innovative approach to efficiently solve the full-vectorial 3D-IS problem has been presented.
The retrieval of the volumetric contrast distribution of sparse non-weak scatterers has been tackled as
a probabilistic CSI-based problem, which has been efficiently solved through a customized MT-BCS
approach. The numerical analysis has pointed out the following key features of the proposed technique:
• Reliable 3D reconstructions of the EM properties of the imaged domain are yielded processing
scattering data also blurred with a non-negligible amount of additive noise;
• The inversion accuracy of the proposed CS-based approach depends on the degree of sparseness
of the actual scenario with respect to the expansion basis at hand. However, it can be
47
J. Imaging 2019, 5, 19
fruitfully and profitably applied when other/different (non-voxel) representations of the contrast
source/contrast function are chosen [46];
• The MT implementation of the BCS-based inversion remarkably overcomes its single-task (ST-BCS)
counterpart thanks to the profitable exploitation of the existing correlations between the V views
and the scattered field components;
• The MT-BCS positively compares with other state-of-the-art approaches, also deterministic and
non-CS, in terms of both reconstruction accuracy and computational efficiency.
Moreover, the methodological advancements of this work with respect to the state-of-the-art on the
topic [47,48] include (i) the generalization of the MT-BCS strategy to handle 3D-IS problems, differently
from previous customizations of such an inversion paradigm which deal only with two-dimensional
formulations [48], (ii) the derivation of a BCS-based imaging approach able to retrieve 3D target contrast
information, unlike state-of-the-art Bayesian CS contributions only dealing with the reconstruction
of equivalent sources [47], and (iii) the analysis and validation of suitable operative guidelines for the
optimal setting of the key calibration parameters of the introduced methodology. Future works will be
aimed at extending the capabilities of the proposed approach to effectively deal with non voxel-sparse
targets as well as with other applicative scenarios of great interest (e.g., subsurface imaging) including
the processing of multi-frequency data [55].
Author Contributions: All authors contributed equally to this work.

Funding: This work has been partially supported by the Italian Ministry of Foreign Affairs and International
Cooperation, Directorate General for Cultural and Economic Promotion and Innovation within the SNATCH
Project (2017–2019) and by the Italian Ministry of Education, University, and Research within the Program “Smart
cities and communities and Social Innovation” (CUP: E44G14000060008) for the Project “WATERTECH—Smart
Community per lo Sviluppo e l’Applicazione di Tecnologie di Monitoraggio Innovative per le Reti di Distribuzione
Idrica negli usi idropotabili ed agricoli” (Grant no. SCN_00489).
References
1. Chen, X. Computational Methods for Electromagnetic Inverse Scattering; Wiley-IEEE: Singapore, 2018.
2. Zoughi, R. Microwave Nondestructive Testing and Evaluation; Kluwer: Amsterdam, The Netherlands, 2000.
3. Ghasr, M.T.; Horst, M.J.; Dvorsky, M.R.; Zoughi, R. Wideband microwave camera for real-time 3-D imaging.
4. Fallahpour, M.; Zoughi, R. Fast 3-D qualitative method for through-wall imaging and structural health
monitoring. IEEE Geosci. Remote Sens. Lett. 2015, 12, 2463–2467. [CrossRef]
5. Benjamin, R.; Craddock, I.J.; Hilton, G.S.; Litobarski, S.; McCutcheon, E.; Nilavalan, R.; Crisp, G.N.
Microwave detection of buried mines using non-contact, synthetic near-field focusing. IEE P-Radar Son. Nav.
2001, 148, 233–240. [CrossRef]
6. Sheen, D.M.; McMakin, D.L.; Hall, T.E. Three-dimensional millimeter-wave imaging for concealed weapon
detection. IEEE Trans. Microwave Theory Technol. 2001, 49, 1581–1592. [CrossRef]
7. Di Donato, L.; Crocco, L. Model-based quantitative cross-borehole GPR imaging via virtual experiments.
IEEE Trans. Geosci. Remote Sens. 2015, 53, 4178–4185. [CrossRef]
8. Catapano, I.; Crocco, L.; Persico, R.; Pieraccini, M.; Soldovieri, F. Linear and nonlinear microwave
tomography approaches for subsurface prospecting: Validation on real data. IEEE Antennas Wirel. Propag. Lett.
2006, 5, 49–53. [CrossRef]
9. Bucci, O.M.; Crocco, L.; Isernia, T.; Pascazio, V. Subsurface inverse scattering problems: Quantifying,
qualifying, and achieving the available information. IEEE Trans. Geosci. Remote Sens. 2001, 39, 2527–2538.
[CrossRef]
10. Bevacqua, M.; Crocco, L.; Di Donato, L.; Isernia, T.; Palmeri, R. Exploiting sparsity and field conditioning in
subsurface microwave imaging of nonweak buried targets. Radio Sci. 2016, 51, 301–310. [CrossRef]
11. Amineh, R.K.; Khalatpour, A.; Nikolova, N.K. Three-dimensional microwave holographic imaging using co-
and cross-polarized data. IEEE Trans. Antennas Propag. 2012, 60, 3526–3531. [CrossRef]
48
J. Imaging 2019, 5, 19
12. Semenov, S.Y.; Bulyshev, A.E.; Abubakar, A.; Posukh, V.G.; Sizov, Y.E.; Souvorov, A.E.; van den Ber, P.M.;
Williams, T.C. Microwave-tomographic imaging of the high dielectric-contrast objects using different
image-reconstruction approaches. IEEE Trans. Microw. Theory Technol. 2005, 53, 2284–2294. [CrossRef]
13. Semenov, S.Y.; Bulyshev, A.E.; Souvorov, A.E.; Nazarov, A.G.; Sizov, Y.E.; Svenson, R.H.; Posukh, V.G.;
Pavlovsky, A.; Repin, P.N.; Tatsis, G.P. Three-dimensional microwave tomography: Experimental imaging of
phantoms and biological objects. IEEE Trans. Microwave Theory Technol. 2000, 48, 1071–1074. [CrossRef]
14. Zhang, Z.Q.; Liu, Q.H. Three-dimensional nonlinear image reconstruction for microwave biomedical imaging.
IEEE Trans. Biomed. Eng. 2004, 51, 544–548. [CrossRef] [PubMed]
15. Bulyshev, A.E.; Semenov, S.Y.; Souvorov, A.E.; Svenson, R.H.; Nazarov, A.G.; Sizov, Y.E.; Tatsis, G.P.
Computational modeling of three-dimensional microwave tomography of breast cancer. IEEE Trans.
Biomed. Eng. 2001, 48, 1053–1056. [CrossRef]
16. Colgan, T.J.; Hagness, S.C.; Van Veen, B.D. A 3-D level set method for microwave breast imaging. IEEE Trans.
17. Winters, D.W.; Shea, J.D.; Kosmas, P.; Van Veen, B.D.; Hagness, S.C. Three-dimensional microwave breast
imaging: dispersive dielectric properties estimation using patient-specific basis functions. IEEE Trans.
Med. Imaging 2009, 28, 969–981. [CrossRef] [PubMed]
18. Grzegorczyk, T.M.; Meaney, P.M.; Kaufman, P.A.; di Florio-Alexander, R.M.; Paulsen, K.D. Fast 3-D
tomographic microwave imaging for breast cancer detection. IEEE Trans. Med. Imaging 2012, 31, 1584–1592.
[CrossRef] [PubMed]
19. Johnson, J.E.; Takenaka, T.; Ping, K.A.H.; Honda, S.; Tanaka, T. Advances in the 3-D forward-backward
time-stepping (FBTS) inverse scattering technique for breast cancer detection. IEEE Trans. Biomed. Eng. 2009,
56, 2232–2243. [CrossRef] [PubMed]
20. Bevacqua, M.T.; Scapaticci, R. A compressive sensing approach for 3D breast cancer microwave imaging
with magnetic nanoparticles as contrast agent. IEEE Trans. Med. Imaging 2016, 35, 665–673. [CrossRef]
21. Bucci, O.M.; Bellizzi, G.; Catapano, I.; Crocco, L.; Scapaticci, R. MNP enhanced microwave breast cancer
imaging: Measurement constraints and achievable performances. IEEE Antennas Wirel. Propag. Lett. 2012,
11, 1630–1633. [CrossRef]
22. Angiulli, G.; Carlo, D.D.; Isernia, T. Matching fluid influence on field scattered from breast tumour: Analysis
using 3D realistic numerical phantoms. Electron. Lett. 2012, 48, 13–14. [CrossRef]
23. Oliveri, G.; Rocca, P.; Massa, A. A Bayesian compressive sampling-based inversion for imaging sparse
scatterers. IEEE Trans. Geosci. Remote Sens. 2011, 49, 3993–4006. [CrossRef]
24. Palmeri, R.; Bevacqua, M.T.; Crocco, L.; Isernia, T.; Di Donato, L. Microwave imaging via distorted iterated
virtual experiments. IEEE Trans. Antennas Propag. 2017, 65, 829–838. [CrossRef]
25. Di Donato, L.; Palmeri, R.; Sorbello, G.; Isernia, T.; Crocco, L. A new linear distorted-wave inversion method for
microwave imaging via virtual experiments. IEEE Trans. Microwave Theory Technol. 2016, 64, 2478–2488. [CrossRef]
26. Di Donato, L.; Bevacqua, M.T.; Crocco, L.; Isernia, T. Inverse scattering via virtual experiments and contrast
source regularization. IEEE Trans. Antennas Propag. 2015, 63, 1669–1677. [CrossRef]
27. Bevacqua, M.T.; Crocco, L.; Di Donato, L.; Isernia, T. An algebraic solution method for nonlinear inverse
scattering. IEEE Trans. Antennas Propag. 2015, 63, 601–610. [CrossRef]
28. Crocco, L.; Di Donato, L.; Catapano, I.; Isernia, T. An improved simple method for imaging the shape of
complex targets. IEEE Trans. Antennas Propag. 2013, 61, 843–851. [CrossRef]
29. Poli, L.; Oliveri, G.; Rocca, P.; Massa, A. Bayesian compressive sensing approaches for the reconstruction of
two-dimensional sparse scatterers under TE illumination. IEEE Trans. Geosci. Remote Sens. 2013, 51, 2920–2936.
[CrossRef]
30. Li, M.; Abubakar, A.; Habashy, T.M. A three-dimensional model-based inversion algorithm using radial
basis functions for microwave data. IEEE Trans. Antennas Propag. 2012, 60, 3361–3372. [CrossRef]
31. Meaney, P.M.; Paulsen, K.D.; Geimer, S.D.; Haider, S.A.; Fanning, M.W. Quantification of 3-D field effects
during 2-D microwave imaging. IEEE Trans. Biomed. Eng. 2002, 49, 708–720. [CrossRef]
32. Semenov, S.Y.; Svenson, R.H.; Bulyshev, A.E.; Souvorov, A.E.; Nazarov, A.G.; Sizov, Y.E.; Pavlovsky, A.V.;
Borisov, V.Y.; Voinov, B.A.; Simonova, G.I.; et al. Three-dimensional microwave tomography: Experimental
prototype of the system and vector Born reconstruction method. IEEE Trans. Biomed. Eng. 1999, 46, 937–946.
[CrossRef]
49
J. Imaging 2019, 5, 19
33. Bucci, O.M.; Isernia, T. Electromagnetic inverse scattering: retrievable information and measurement
strategies. Radio Sci. 1997, 32, 2123–2137. [CrossRef]
34. Fear, E.C.; Li, X.; Hagness, S.C.; Stuchly, M.A. Confocal microwave imaging for breast cancer detection:
Localization of tumors in three dimensions. IEEE Trans. Biomed. Imaging 2002, 49, 812–822. [CrossRef]
35. Ali, M.A.; Moghaddam, M. 3D nonlinear super-resolution microwave inversion technique using time-domain
data. IEEE Trans. Antennas Propag. 2010, 58, 2327–2336. [CrossRef]
36. Abubakar, A.; Habashy, T.M.; Pan, G.; Li, M. Application of the multiplicative regularized Gauss-Newton
algorithm for three-dimensional microwave imaging. IEEE Trans. Antennas Propag. 2012, 60, 2431–2441. [CrossRef]
37. Donelli, M.; Franceschini, D.; Rocca, P.; Massa, A. Three-dimensional microwave imaging problems solved
through an efficient multi-scaling particle swarm optimization. IEEE Trans. Geosci. Remote Sens. 2009, 47,
1467–1481. [CrossRef]
38. De Zaeytijd, J.; Franchois, A.; Eyraud, C.; Geffrin, J. Full-wave three-dimensional microwave imaging
with a regularized Gauss-Newton Method—Theory and experiment. IEEE Trans. Antennas Propag. 2007,
55, 3279–3292. [CrossRef]
39. Harada, H.; Wall, D.J.N.; Takenake, T.; Tanaka, M. Conjugate gradient method applied to inverse scattering
problem. IEEE Trans. Antennas Propag. 1995, 43, 784–792. [CrossRef]
40. Salucci, M.; Oliveri, G.; Anselmi, N.; Viani, F.; Fedeli, A.; Pastorino, M.; Randazzo, A. Three-dimensional
electromagnetic imaging of dielectric targets by means of the multiscaling inexact-Newton method. J. Opt.
Soc. Am. A 2017, 34, 1119–1131. [CrossRef]
41. Estatico, C.; Pastorino, M.; Randazzo, A.; Tavanti, E. Three-Dimensional Microwave Imaging in L P Banach
Spaces: Numerical and Experimental Results. IEEE Trans. Comput. Imaging 2018, 4, 609–623. [CrossRef]
42. Simonov, N.; Kim, B.; Lee, K.; Jeon, S.; Son, S. Advanced fast 3-D electromagnetic solver for microwave
tomography imaging. IEEE Trans. Med. Imaging 2017, 36, 2160–2170. [CrossRef]
43. Wang, X.Y.; Li, M.; Abubakar, A. Acceleration of 2-D multiplicative regularized contrast source inversion
algorithm using paralleled computing architecture. IEEE Antennas Wirel. Propag. Lett. 2017, 16, 441–444. [CrossRef]
44. Oliveri, G.; Salucci, M.; Anselmi, N.; Massa, A. Compressive sensing as applied to inverse problems for
imaging: Theory, applications, current trends, and open challenges. IEEE Antennas Propag. Mag. 2017,
59, 34–46. [CrossRef]
45. Massa, A.; Rocca, P.; Oliveri, G. Compressive sensing in electromagnetics—A review. IEEE Antennas
Propag. Mag. 2015, 57, 224–238. [CrossRef]
46. Anselmi, N.; Oliveri, G.; Hannan, M.A.; Salucci, M.; Massa, A. Color compressive sensing imaging of
arbitrary-shaped scatterers. IEEE Trans. Microware Theory Technol. 2017, 65, 1986–1999. [CrossRef]
47. Oliveri, G.; Ding, P.-P.; Poli, L. 3D crack detection in anisotropic layered media through a
sparseness-regularized solver. IEEE Antennas Wirel. Propag. Lett. 2015, 14, 1031–1034. [CrossRef]
48. Poli, L.; Oliveri, G.; Viani, F.; Massa, A. MT-BCS-based microwave imaging approach through
minimum-norm current expansion. IEEE Trans. Antennas Propag. 2013, 61, 4722–4732. [CrossRef]
49. Bevacqua, M.T.; Crocco, L.; Di Donato, L.; Isernia, T. Non-linear inverse scattering via sparsity regularized
contrast source inversion. IEEE Trans. Computat. Imag. 2017, 3, 296–304. [CrossRef]
50. Qiu, W.; Zhou, J.; Zhao, H.; Fu, Q. Three-dimensional sparse turntable microwave imaging based on
compressive sensing. IEEE Geosci. Remote Sens. Lett. 2015, 12, 826–830. [CrossRef]
51. Haynes, M.; Stang, J.; Moghaddam, M. Real-time microwave imaging of differential temperature for thermal
therapy monitoring. IEEE Trans. Biomed. Eng. 2014, 61, 1787–1797. [CrossRef]
52. Li, M.; Abubakar, A.; Van Den Berg, P.M. Application of the multiplicative regularized contrast source
inversion method on 3D experimental Fresnel data. Inverse Probl. 2009, 25, 1–23. [CrossRef]
53. Ji, S.; Dunson, D.; Carin, L. Multitask compressive sensing. IEEE Trans. Signal Process. 2009, 57, 92–106.
[CrossRef]
54. Tipping, M.E. Sparse Bayesian learning and the relevant vector machine. J. Mach. Learn. Res. 2001, 1, 211–244.
55. Bucci, O.M.; Crocco, L.; Isernia, T.; Pascazio, V. Inverse scattering problems with multifrequency data:
Reconstruction capabilities and solution strategies. IEEE Trans. Geosci. Remote Sens. 2000, 38, 1749–1756. [CrossRef]
50
Journal of
Imaging
Article
Image-Based RCS Estimation from Near-Field Data
Tushar Rajvanshi † , Maria Antonia Maisto † , Angela Dell’Aversano † and Raffaele Solimene ∗,†
Department of Engineering, University of Campania, 81031 Aversa, Italy; [email protected] (T.R.);
[email protected] (M.A.M.); [email protected] (A.D.)
* Correspondence: [email protected]; Tel.: +39-081-50-10-335
Received: 7 May 2019; Accepted: 14 June 2019; Published: 17 June 2019
Abstract: This paper deals with the problem of estimating the RCS from near-field data by
image-based approaches. In particular, a rigorous focusing procedure based on a weighted
adjoint scheme, which is also applicable to an arbitrary measurement curve, is developed.
The developed formalism allows us to address the important question concerning the need to
employ a multi-frequency configuration to estimate the RCS. Accordingly, it is shown that if RCS
is required at a given frequency, then the target image obtained solely at such a frequency can be
exploited provided that the spatial truncation arising from the size of the investigated area is properly
taken into account.
Keywords: RCS estimation; image-based approach; adjoint inversion methods
1. Introduction
The Radar Cross Section (RCS) of a target is a crucial quantity that describes how the target
responds to an impinging electromagnetic wave (i.e., how it scatters the incident wave) across
different directions. As is well known, RCS is formally defined as the distance between the probing
antenna and the target approaching infinity [1]. In practical cases, to measure the RCS, the physical
separation between the target and the illuminating/measuring antenna is required to comply with
the far-field conditions. However, as frequency increases, and depending on the size of the target,
the required separation can soon become very large. Therefore, for a long time, there has been a great
interest in developing methods for predicting the RCS for scattered field measurements taken at
a distance not in far-field [2–6]. Indeed, the possibility of acquiring data in near-zone avoids the
above-mentioned drawback and in principle offers several advantages since measurements can be
taken within an anechoic chamber. On the other hand, near-zone measurements cannot be used
straight away for RCS estimation, due to the wavefront curvature and because in near-zone the
target is not uniformly illuminated. Compact range equipment solves these problems, but requires
high-quality reflectors [4,6]. Alternatively, image-based approaches first obtain a reconstruction of
the target reflectivity and then the RCS is estimated by Fourier, transforming the obtained reflectivity.
Accordingly, those approaches can be regarded as particular near-field to far-field transformations [7].
One can consider compact range methods and image-based approaches to move the job to do from
hardware, i.e., the reflector, to software, i.e., the scattered field data processing algorithms.
In this paper, we are concerned with image-based approaches, in particular under a monostatic
measurement setup.
The underlying working assumption is that the target scatters linearly, i.e., multiple scattering,
shadowing, and creeping waves are considered negligible. Accordingly, the scattering phenomenon
can be described by a linear integral scattering operator and the reconstruction of the target reflectivity
is usually obtained by solving (inverting) that integral equation via some focusing procedure borrowed
from back-propagation/migration algorithms literature [8].

J. Imaging 2019, 5, 61
Focusing entails some smearing in the reconstruction (depending on the measurement

configuration parameters such as frequency band, etc.) which can negatively impact the subsequent
RCS estimation. To compensate for such an effect, the obtained RCS can be normalized by the one
corresponding to a point-like target [9] or, alternatively, the standard focusing kernel can be corrected
by a suitable factor [10].
The first procedure implicitly assumes a spatially invariant behavior (within the image region)
of the imaging procedure which indeed does not hold. The second approach has instead been
demonstrated only for a circular measurement curve. The first contribution of this paper is the
derivation of the focusing procedure under a general framework as far as the measurement curve
is concerned, which actually generalizes the procedure reported in [10] to a generally shaped
measurement curve.
To obtain high-resolution target reconstructions, the imaging stage is usually achieved by using
multi-frequency data. However, target reflectivity in general depends on frequency. As outlined
in [7], in that case, the obtained reflectivity can be considered more like an average over the employed
frequency band. Furthermore, the obtained reflectivity reconstruction is generally used for RCS
estimation just at the central frequency. It is then natural to ask if the imaging procedure can work by
directly using single-frequency data.
This is indeed possible [11] even though RCS estimation is more sensitive to the spatial truncation
determined by the size of the image region.
2. Basics on Image-Based RCS

In this section, we introduce the problem, the adopted notation, and briefly recall image-based
methods present in literature.
The scattering experiment related to the RCS estimation considered in this paper is described
in Figure 1. In particular, we refer to a 2D scalar configuration, i.e., the target is invariant along
the z-axis and the probing field is a cylindrical wave linearly polarized again along z. Accordingly,
the scattered field is also linearly polarized and the RCS is a single scalar and not a dyad, as in the
general 3D vector case. The target is assumed to reside within the image domain D I , whereas the
scattered field is collected in near-field over a curve Γ surrounding D I and for the frequency band Ω.
The usual monostatic configuration is considered. Accordingly, during the data-acquisition stage,
the same antenna acts as transmitter and receiver and then moves around the target to synthesize
the measurement curve, or equivalently the target is rotated over a turntable. In the latter case the
synthesized measurement curve is just a circle, which is the one that is usually considered in literature.
Figure 1. Geometry of the problem.
52
J. Imaging 2019, 5, 61
It is worth remarking that we are considering such a 2D configuration for the sake of simplicity,
since this allows easier production of the numerical examples to be shown later on. However, a similar
configuration is commonly addressed also in 3D cases while estimating the so-called water line RCS [7].
The starting point for problem formulation is the assumption that the target consists of
an ensemble of independent and non-directional scattering centers [9]. Since multiple interactions,
creeping waves and shadowing effects are neglected, the target reflectivity γ(r) and the scattered field
are linked by the following linear integral operator

ES (k, ro (s)) = A(k, ro (s), r)e− jφ[k,ro (s),r] γ(r)d2 r (1)
DI
where k is the wavenumber and ro (s) ∈ Γ is the measurement curve parametrized with respect to s.
Moreover, the phase term takes into account the propagation path from the antenna to the target and
back, i.e.,
φ(k, ro , r) = 2k|ro (s) − r|
whereas the amplitude term is given by
αC (k ) P2 [θ (ro (s), r)]

A(k, ro (s), r) =
|ro (s) − r|
with α being a constant term that varies depending on whether a 3D or 2D geometry is being considered,
P[θ (ro (s), r)] is the antenna directivity pattern assumed to be 1 for θ (ro (s), r) = 0, (with θ (ro (s), r)
being the angle between the ro (s) − r and the antenna broadside direction) and C (k ) collects the
antenna frequency behavior and the frequency part of the propagator (Green function) amplitude.
In particular, for the case at hand, the relevant Green function is proportional to a Hankel function of
zero order and second kind. P[θ (ro (s), r)] and C (k) are assumed to be known by a preliminary stage of
measurement and calibration using a reference target of known RCS such as a metallic sphere, a plate
or the like.
At this juncture, some further considerations concerning the model (1) are in order. First, it is
noted that in (1) each single point in the image domain has been considered to be being in the far-field
of the transmitting/receiving antenna. If this is not the case, a linear transformation could be still
established by invoking the plane-wave spectrum representation for both the Green function and the
antenna response. This circumstance is not considered in this manuscript. Second, field data, instead
of voltage, are being considered. Voltage is actually what one can measure. However, assuming
field data does not impair the generality of the theoretical/numerical analysis we intend to pursue.
Finally, the target reflectivity is in general frequency dependent; we denote such a dependence by
γ f (k ). However, if the target can be represented by an ensemble of point-like scatterers, the frequency
dependent part of reflectivity is a priori known and hence can be considered embodied within the C (k )
term [10]. However, we will turn back on this assumption in the following sections.
The near-field (1) is in general much different from the corresponding far-field and hence cannot
be used directly for estimating the target RCS. This of course is due to the wavefront curvature of the
impinging electromagnetic wave and to the non-uniform illumination of the target. The aim of the
imaging stage is just to compensate for such wavefront curvature and amplitude behavior. This is
achieved by processing the scattered field data through a focusing operator that actually “translates”
the scattered field data into an image of the target under test. Formally, this is written as

γ̃(r) = f c [k, ro (s), r] ES [k, ro (s)]dkds (2)
Ω×Γ
where f c [k, ro (s), r] is the focusing kernel which is commonly chosen equal to e jφ[k,ro (s),r] /A[k, ro (s), r].
Clearly, γ̃ is a filtered (blurred) version of the actual reflectivity. That filtering depends on the
53
J. Imaging 2019, 5, 61
configuration parameters, i.e., Ω × Γ, which reflect the properties of the imaging point-spread
function, i.e.,

A[k, ro (s), r ] j[φ[k,ro (s),r]−φ[k,ro (s),r ]]
ps f (r, r ) = e dkds (3)
Ω×Γ A[k, ro (s), r]
so that (2) can be equivalently rewritten as

γ̃(r) = ps f (r, r )γ(r )d2 r (4)
DI
Once the reflectivity has been estimated, the RCS can be computed by using its (2D) definition equation

E (k, ro ) 2
σ(k, θo ) = lim 2πro S
(5)
ro →∞ E inc
with ro ≡ (ro , θo ) being the observation point. Hence, on exploiting (1) and (2), (5) yields
2

σ (k, θo ) = αγ f (k ) γ̃(r)e2jkr̂o ·r d2 r (6)
D I
Equation (6) allows recognition that under the assumed linear scattering model, the RCS depends
of the Fourier transform of the reflectivity function projected over the so-called Ewald disc (resp. sphere
for the 3D case) [12]. Of course, because of the blurring introduced by the imaging procedure, some
error will corrupt the estimation (6). In order to mitigate such an error, a common way to proceed is to
normalize (6) by the RCS of the point-spread function [9], which is computed by Fourier transforming
ps f (r, 0). It is clear that behind this procedure there is the implicit assumption of considering (4) as
a convolution which in general does not hold. Differently, in [10], it is shown that by introducing
a suitable correction term, the focusing procedure can be approximated as a Fourier transformation.
Accordingly, γ̃ proves to be a windowed (in spatial spectrum domain) Fourier transform of the γ;
therefore, normalization is no longer required.
We will come back to these important points in the next section where we introduce an imaging
procedure for a more general (with respect to the usual circle) measurement curve.
3. Adjoint Inversion for Generic Measurement Curve

According to the previous section, the first step in any image-based RCS estimation method is to
solve the integral Equation (1) for the reflectivity γ. Formally, this entails finding the inverse of the
linearized scattering operator
AS : γ → ES (7)
AS being just the integral operator in (1). A very common way to invert (7) is to adopt the adjoint
operator A†S instead of the inverse of AS . A number of popular methods such as time-reversal,
reverse-migration, back-propagation, etc. are adjoint-based imaging schemes [8]. Inversion through the
adjoint allows us to deal with the ill-posedness of the problem in that it retunes a stable reconstruction
procedure. However, it is not a Tichonov regularization scheme as the corresponding reconstruction
fails to converge to the generalized solution even in absence of noise [8]. Their great diffusion is due to
their simple physical understanding and the possibility to be often implemented via FFT. In particular,
adjoint inversion succeeds in compensating the phase in correspondence to the actual scatterers’s
position but the amplitude is not addressed properly. For this reason, adjoint inversion is often paired
with a pre-weighting (filtering) stage, so that the image/reconstruction is obtained as [13]
A†S : WES → γ̃ (8)
54
J. Imaging 2019, 5, 61
where W (k, ro , r) is just the weighting function. In particular, on exploiting (1), the reconstruction
operator (8) can be explicitly written as

γ̃(r) = W [k, ro (s), r]e jφ[k,ro (s),r] ES [k, ro (s)]dkds (9)
Ω×Γ
which yields the point-spread function

ps f (r, r ) = W [k, ro (s), r] A[k, ro (s), r ]e j{φ[k,ro (s),r]−φ[k,ro (s),r ]} dkds (10)
Ω×Γ
Please note that the usual choice for the weighting function, as done in the previous section, is to
W [k, ro (s), r] = 1/A[k, ro (s), r]. On the other hand, it is desirable that the point-spread function be as
close as possible to a delta function. If this is achieved, the adjoint inversion actually approximates
a regularized reconstruction [14]. To cope with this question, the weighting function must be properly
chosen as detailed below.
We start by introducing the variable
1
w[k, ro (s), r, r ] = − ∇x φ[k, ro (s), x]|x=r +λ(r−r ) dλ (11)
0
which allows rewriting of the phase term in (10) as
φ[k, ro (s), r] − φ[k, ro (s), r ] = −w[k, ro (s), r, r ] · (r − r ) (12)
and the point-spread function (10) as

ps f (r, r ) = W (w, r) A(w, r )e− jw·(r−r ) |J| d2 w (13)
Ωw
with J being the Jacobian of the transformation that maps w in (k, s). Equation (13) shows the
point-spread function as pseudodifferential operators which enjoy the so-called pseudolocal
property [15]. This is the very mathematical rationale that justifies performing the adjoint inversion to
retrieve object singularities, even though in order to correctly retrieve the singularity amplitudes the
weighting function must be properly chosen. Also, (13) makes it immediately clear that the leading
order contribution occurs for r − r = 0. Accordingly, the approximation
w[k, ro (s), r, r ] = w[k, ro (s), r, r] = −∇x φ[k, ro (s), x]|x=r (14)
is made in (13). This leads to the following point-spread function approximation

ps f (r, r ) e− jw·(r−r ) d2 w (15)
Ωw
once the weighting function is chosen as
W (w, r) = 1/ [ A(w, r) |J|] (16)
It is seen that (15) is a filtered version of a Dirac delta. Moreover, Equation (15) makes it clear the
spectral content of the target that can be retrieved. Indeed, when the observation curve Γ goes around
the target it results that
Ωw = {w : 2k min ≤ w ≤ 2k max } (17)
where k min and k max are the wavenumbers corresponding to the lowest and highest adopted
frequencies. This result could have been expected since it exactly coincides to what can be retrieved
by a far-field configuration. However, it must be kept in mind that (17), and of course (15), holds
55
J. Imaging 2019, 5, 61
only approximately true because of the approximation (14). However, (14) is expected to work well in
reproducing the main beam of the point-spread function, which in turn plays the major role in the
filtering introduced by the imaging procedure.
It is interesting to highlight that (15) holds true regardless of the shape of the observation
curve Γ. Indeed, the shape of Γ only enters in the choice of the weighting function through the
Jacobian term. In particular, when Γ is a circle surrounding the image region (as it is commonly
assumed) simple calculations show that (15) exactly returns the imaging kernel introduced in [10],
with the correcting term there introduced just being given by 1/|J|. In this regard, the formulation
introduced in this paper generalizes previous literature results, which indeed have mainly considered
circular measurement curves. Previous discussion can be summarized by the following statement:
the resolution achievable during the image stage does not depend on the shape of the measurement
curve. Moreover, by exploiting techniques similar to those in [16], an analytical expression for the
point-spread function can be obtained. In particular, it can be shown that the point-spread function
in (15) is given as
2π [ψ(2k max |r − r |) − ψ(2k min |r − r |)]
ps f (r, r ) (18)
| r − r |2
with x
ψ( x ) = yJ0 (y)dy = xJ1 ( x ) (19)
0
and J0 (·) and J1 (·) being Bessel functions of zero and first order. In Figure 2, the comparison between
the actual point-spread function and the one returned by (18) is shown. As can be appreciated, the two
curves overlap very well, this means that the leading term approximation exploited to derive (18)
works fine.
1 1
0.9 0.9
0.8 0.8
0.7 0.7
Normalized |psf|
Normalized |psf|
0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
-1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1
y[m] x[m]
Figure 2. Comparison between the normalized point-spread function amplitude (blue line) and the
one returned by (18) (red dotted line). The left panel refers to the cut along the y-axis; the right one to
the cut along the x-axis. The adopted frequency band is Ω = [0.5, 1.5] GHz while the observation curve
is an ellipse with axes of 2 m and 2.5 m, respectively.
56
J. Imaging 2019, 5, 61
A further point that is worth stressing is that for a full-view scan (i.e., for a measurement curve
running all around the image region) the spectral domain Ωw does not depend on the image point r.
This means that the image kernel (15) can be actually considered to be convolution; however, this
does not hold true for (10), since to get (15) a spatially varying kernel was required. Accordingly,
the normalization procedure in [9] can be expected to work only when the image points are close to
the reference one (i.e., the one for which the point-spread function is computed and transformed).
Finally, it is remarked that to speed up the imaging procedure FFT-based routines are often
employed [17]. According to the previous formulation, this can be achieved for a generic measurement
curve. Of course, the usual interpolation and resampling step, which is required to map data over
a uniform grid, in general depends on the particular curve under concern. In the following numerical
examples we do not follow such a procedure, rather we employ (9) with the weighting function chosen
as in (16).
4. Single-Frequency Imaging
As clearly stated above, the previous RCS estimation procedure is founded on the assumption that
the targets are frequency independent or their frequency behaviors are equal and known. The latter
holds true for point-like targets but for most practical cases, even when targets can be still described
by an ensemble of scattering centers, the different contributions depend on the operating frequency.
Accordingly, the reflectivity returned by (9) is actually an averaged version performed over the
frequency band. This circumstance in general impacts negatively on the RCS computation. That is
why, the obtained image is usually used to compute the RCS only at the central (of the adopted band)
frequency [7].
Let us relax the above-mentioned assumption. Hence, here we consider the cases in which the
target frequency behavior is unknown, or it consists of scattering centers whose reflectivity coefficients
differently depends on frequency. In these cases, the frequency behavior cannot be singled out and
embodied within the scattering operator since actually γ = γ(k, r). Accordingly, the scattered field is
yielded by
ES [k, ro (s)] = A[k, ro (s), r]e− jφ[k,ro (s),r] γ(k, r)d2 r (20)
DI
and the estimated reflectivity (through imaging) and the actual one are linked as
γ̃(k,r)
0 12 3

γ̃(r) = dk f c [k, ro (s), r] A[k, ro (s), r ]e− jφ[k,ro (s),r ] γ(k, r )dsd2 r (21)
Ω DI Γ
It is seen that γ̃(r) is now given as the coherent summation of the single-frequency images γ̃(k, r)
and as such it does not represent the (regularized) solution of (20). This could have been expected
since both scattered field data and unknown reflectivity depend on the frequency. As discussed above,
this leads to a degradation of accuracy while estimating the RCS.
The arising question is whether, in order to estimate the RCS at a given frequency, a frequency
band must necessarily be employed. The answer to this question seems positive if high-resolution
radar imaging is aimed. At the other hand, single-frequency images are not affected by previous
drawback but the performance in the reconstruction in general results much lower. Hence, the question
can be rephrased as whether image degradation (due to single-frequency data) impairs RCS estimation.
To this end, we particularize (15) to the single-frequency case, for example by considering k = k av
(i.e., the average one)

ps f s f (r, r ) e− jw·(r−r ) d2 w (22)
Cw
where the subscript s f stands for single-frequency and Cw = {w : w = 2k}. (22) can be easily
computed and the result is
ps f s f (r, r ) 4πkJ0 (2k|r − r |) (23)
57
J. Imaging 2019, 5, 61
In Figure 3 the comparison between the actual point-spread function and the one returned by (23)
is shown. Again, the estimated psf allows us to obtain a good approximation of the actual one.
To understand how resolution degrades when single-frequency data are used, Equation (23) can
be compared to (18). It is clear that now the spatial spectrum that can be retrieved about the unknown
reflectivity is much lower since Cw consists of a single circle instead of the circular annulus Ωw .
Nonetheless, as shown in [16], it is the relative bandwidth (i.e., (k max − k min )/k av ), rather than the
frequency band, that plays the major role. This is somewhat different from the common belief that
the frequency band is necessary and is a consequence of the full-view (i.e., measurements are taken
all around the target) configuration. However, we do not want to dwell any further on that point
since this is not the focus of the paper. The point is whether image degradation impairs the possibility
of obtaining RCS. To this end, we once again remark that (22) introduces a filtering that allows
retention of only the target reflectivity over the circle Cw , which is the Ewald circle corresponding to
the wavenumber k and which is, according to (6), what is necessary to compute the RCS.
Finally, by looking at Figure 3, it can be seen that the side-lobes are higher than the multi-frequency
case. This allows expectation that while comparing multi-frequency and single-frequency RCS
estimations, the spatial truncation introduced by the size of the investigated area will play a crucial role.
1 1
0.9 0.9
0.8 0.8
0.7 0.7
Normalized |psf|
Normalized |psf|
0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
-1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1
y[m] x[m]
Figure 3. Comparison between the normalized point-spread function amplitude (blue line) and the
one returned by (23) (red dotted line). The left panel refers to the cut along the y-axis whereas the right
one to the cut along the x-axis. The frequency is 1 GHz while the observation curve is an ellipse with
axes 2 m and 2.5 m.
To check previous arguments, we conclude this section by running a simple case.

In particular, we consider two point-like objects whose reflectivity is given by γ( x, y) =
0.1( j k k )2 [δ( x, y − 0.5) + δ( x, y + 0.5)]. The electric field is collected over an ellipse whose axes are
min
3 m and 3.5 m and the working frequency is f = 1.75 GHz. Since the far-field distance is 11.7 m,
the ellipse lies in the near-field region of two point scatterers. In Figure 4 the RCS at that frequency
is shown. The green lines refer to the actual RCS, while the blue and red lines to the ones computed
58
J. Imaging 2019, 5, 61
from the multi-frequency (Ω = [0.5, 3] GHz) and single-frequency images, respectively. In particular,
in the panel a) D I = [(−1, 1) × (−1, 1)] is considered whereas for panel b) D I = [(−2, 2) × (−2, 2)].
From the Figure 4a) it can be appreciated that the multi-frequency configuration allows us to obtain
a better RCS estimation. However, when D I increases (see Figure 4b) the red line becomes very similar
to the blue one. This confirms that single-frequency RCS is feasible and competitive with respect to the
multi-frequency one provided the size of the investigated area is enlarged.
-10
[dB]
-20
-30
0 50 100 150 200 250 300 350
a) [deg]
0
-10
[dB]
-20
-30
0 50 100 150 200 250 300 350
b) [deg]
Figure 4. Normalized RCS (in db scale) at f = 1.75 GHz. The green lines refer to the actual RCS, while
the blue and red lines to the ones computed from the multi-frequency and single-frequency images,
respectively. (a) D I = [(−1, 1) × (−1, 1)] and (b) D I = [(−2, 2) × (−2, 2)].
5. Numerical Examples
In this section, some further examples are addressed to check RCS estimation.
In all the following examples, the RCS is evaluated at f avg = 4.5 GHz, the field is collected
over an ellipse whose axes are 30λ avg and 35λ avg with λ avg the wavelength at f avg and D I =
[(−20λ avg , 20λ avg ) × (−20λ avg , 20λ avg )]. Moreover, when the multi-frequency configuration is
exploited Ω = [3, 6] GHz. The target, always √ a perfect electric conducting scatterer, is contained
within a circular image region with radius 5 2λ avg . Accordingly, the far-field distance is 400λ avg
and the ellipse lies in the near-field region of the object. Three different objects, shown in Figure 5,
are considered. √
The first object is a bar of length 10 2λ avg . The corresponding multi-frequency and
single-frequency images are reported in the panels (a) and (b) of Figure 6, while the RCS estimations
are compared in (c). As can be seen, despite the evident degradation of the reconstructed image,
single-frequency configuration allows us to obtain an RCS estimation which is very similar to the
multi-frequency one; indeed the three curves overlap very well.
59
J. Imaging 2019, 5, 61
√
The circular object (of radius 5 2λavg ) is addressed in Figure 7. As can be seen, the multi-frequency
reconstruction allows us to clearly discern the shape of the target which is completely lost in the
single-frequency case. Nonetheless. the RCS estimation is comparable with the multi-frequency one.
The last considered example, the missile shaped target is addressed in Figure 8. The same
discussion as for the previous examples still holds. In this case, it must be remarked that even
though the main RCS features (spikes) are captured, both the multi-frequency and the single-frequency
approaches show some deviation from the actual RCS. This can be ascribed to the model error (for
example multiple scattering, which is more relevant for this complex shaped target) that inherently
affects image-based procedure.
10 10
5 5
y/ avg
y/ avg
0 0
-5 -5
-10 -10
-10 -5 0 5 10 -10 -5 0 5 10
a) x/ avg b) x/ avg
5
y/ avg
-5
-5 0 5
c) x/ avg
Figure 5. Contours of the objects under test.
20 20
10 10
y/ avg
y/ avg
0 0
-10 -10
-20 -20
-20 -10 0 10 20 -20 -10 0 10 20
a) x/ avg b) x/ avg
0
-10
-20
[dB]
-30
-40
-50
0 50 100 150 200 250 300 350
c) [deg]
Figure 6. Object in Figure 5a. (a) Multi-frequency image. (b) Single-frequency image. (c) Normalized
RCS at f = 4.5 GHz. The green line refers to the actual RCS, while the blue and red lines to the ones
computed from multi-frequency and single-frequency images, respectively.
60
J. Imaging 2019, 5, 61
20 20
y/ avg 10 10
y/ avg
0 0
-10 -10
-20 -20
-20 -10 0 10 20 -20 -10 0 10 20
x/ avg x/ avg
-10
-20
[dB]
-30
-40
-50
0 50 100 150 200 250 300 350
[deg]
Figure 7. Object in Figure 5b. (a) Multi-frequency image. (b) Single-frequency image. (c) Normalized
RCS at f = 4.5 GHz. The green line refers to the actual RCS, while the blue and red lines to the ones
computed from the multi-frequency and single-frequency images, respectively.
20 20
10 10
y/ avg
y/ avg
0 0
-10 -10
-20 -20
-20 -10 0 10 20 -20 -10 0 10 20
a) x/ avg b) x/ avg
0
-10
-20
[dB]
-30
-40
-50
0 50 100 150 200 250 300 350
c)
[deg]
Figure 8. Object in Figure 5c. (a) Multi-frequency image. (b) Single-frequency image. (c) Normalized
RCS at f = 4.5 GHz. The green line refer to the actual RCS, while the blue and red lines to the ones
computed from the multi-frequency and single-frequency images, respectively.
61
J. Imaging 2019, 5, 61
6. Conclusions
In this paper, the problem of estimating the RCS from near-field data by image-based approaches
has been addressed.
The first contribution we conveyed in this manuscript was the rigorous derivation of a focusing
procedure based on a weighted adjoint scheme which generalizes for an arbitrary measurement curve
the results presented in [10] that have been developed for a circular measurement curve.
Secondly, the important question concerning the necessity to use a multi-frequency configuration
to estimate the RCS has been studied. Accordingly, a deep analysis of the introduced analytical
tools highlighted that if RCS is required at a given frequency, then the target image obtained solely
at such a frequency can be in principle exploited. This opens the possibility of employing cheaper
measurement systems and to take into account target frequency dependence. However, the spatial
truncation introduced by the size of the investigated area must be properly taken into account.
Several numerical examples corroborated the possibility of computing the RCS from
single-frequency image.
For the sake of simplicity, the study has been developed for a 2D scalar configuration.
The generalization to 3D real scenario is a possible future development.
Author Contributions: Conceptualization, R.S. and M.A.M.; methodology, R.S. and M.A.M.; software, T.R and
A.D.; validation, T.R. and A.D.; formal analysis, R.S.; investigation, M.A.M.; resources, R.S., M.A.M., T.R. and
A.D.; data curation, R.S., M.A.M., T.R. and A.D.; writing–original draft preparation, R.S., M.A.M., T.R. and A.D.;
writing–review and editing, R.S., M.A.M., T.R. and A.D.; visualization, R.S., M.A.M., T.R. and A.D.; supervision,
R.S.; project administration, R.S.; funding acquisition, R.S.
Funding: This research received no external funding.
Acknowledgments: The authors kindly thank Giuseppina Nuzzo for proofreading the manuscript.
References
1. Balanis, A.C. Antenna Theory: Analysis and Design, 3rd ed.; John Wiley and Sons: Hoboken, NJ, USA, 2005.
2. Falconer, D.G. Extrapolation of Near-Field RCS Measurements to the Far Zone. IEEE Trans. Antennas Propag.
1988, 36, 822–829. [CrossRef]
3. Cown, B.J.; Ryan, C.E. Near-field scattering measurements for determining complex target RCS. IEEE Trans.
4. Theron, I.P.; Walton, E.K.; Gunawan, S. Compact range radar cross-section measurements using a noise radar.
5. LaHaie, I.J. Overview of an image-based technique for predicting far-field radar cross section form near-field
measurements. IEEE Antennas Propag. Mag. 2003, 45, 159–169. [CrossRef]
6. Sevgi, L.; Rafiq, Z.; Majid, I. Radar Cross Section (RCS) Measurements. IEEE Antennas Propag. Mag. 2013, 55,
277–291. [CrossRef]
7. Odendaal, J.W.; Joubert, J. Radar cross section measurements using near-field radar imaging. IEEE Trans.
Instrum. Meas. 1996, 45, 948–954. [CrossRef]
8. Solimene, R.; Catapano, I.; Gennarelli, G.; Cuccaro, A.; Dell’Aversano, A.; Soldovieri, F. SAR imaging algorithms
and some unconventional applications: A unified mathematical overview. IEEE Signal Process. Mag. 2014, 31,
90–98. [CrossRef]
9. Broquetas, A.; Palau, J.; Jofre, L.; Cardama, A. Spherical wave near-field imaging and radar cross-section
measurement. IEEE Trans. Antennas Propag. 1998, 46, 730–735. [CrossRef]
10. Osipov, A.; Kobayashi, H.; Suzuki, H. An Improved Image-Based Circular Near-Field-to-Far-Field
Transformation. IEEE Trans. Antennas Propag. 2013, 61, 989–993. [CrossRef]
11. Schnattinger, G.; Mauermayer, R.A.M.; Eibert, T.F. Monostatic Radar Cross Section Near-Field Far-Field
Transformations by Multilevel Plane-Wave Decomposition. IEEE Trans. Antennas Propag. 2014, 62, 4259–4268.
[CrossRef]
62
J. Imaging 2019, 5, 61
12. Devaney, A.J. Mathematical Fundations of Imaging Tomography and Wavefield Inversion; Cambridge University
Press: New York, NY, USA, 2012.
13. Solimene, R.; Cuccaro, A. Back-propagation imaging by exploiting multipath from point scatterers.
14. Cuccaro, A.; Solimene, R. Inverse Source Problem for a Host Medium Having Pointlike Inhomogeneities.
IEEE Trans. Geosci. Remote Sens. 2018, 56, 5148–5159. [CrossRef]
15. Cheney, M.; Bonneau, R.J. Imaging that exploits multipath scattering from point scatterers. Inverse Probl.
2004, 30, 1691–1711. [CrossRef]
16. Solimene, R.; Cuccaro, A.; Ruvio, G.; Tapia, D.F.; O’Halloran, M. Beamforming and holography image
formation methods: An analytic study. Opt. Express 2016,8, 9077–9093. [CrossRef] [PubMed]
17. Vaupel, T.; Eibert, F. Comparison and Application of Near-Field ISAR Imaging Techniques for Far-Field
Radar Cross Section Determination. IEEE Trans. Antennas Propag. 2006, 54, 144–151. [CrossRef]
63
Journal of
Imaging
Article
A Minimum Rank Approach for Reduction of
Environmental Noise in Near-Field Array
Antenna Diagnosis
Marco Donald Migliore *,† , Fulvio Schettino † , Daniele Pinchera † , Mario Lucido † and
Gaetano Panariello †
Department of Electrical and Information Engineering, University of Cassino and Southern Lazio,
Via Di Bisaio 43, 03043 Cassino, Italy; [email protected] (F.S.); [email protected] (D.P.);
[email protected] (M.L.); [email protected] (G.P.)
* Correspondence: [email protected]; Tel.: +39-0776-299-3550
† Current address: DIEI, Via Di Biasio 43, 03043 Cassino, Italy.
Received: 5 February 2019; Accepted: 28 April 2019; Published: 2 May 2019
Abstract: A method to filter out the contribution of interference sources in array diagnosis is proposed.
The interference-affected near field measured on a surface is treated as a (complex-data) image.
This allows to use some modern image processing algorithms. In particular, two strategies widely
used in image processing are applied. The first one is the reduction of the amount of information
by acquiring only the innovation part of an image, as currently happens in video processing.
More specifically, a differential measurement technique is used to formulate the estimation of the
array excitations as a sparse recovery problem. The second technique has been recently proposed in
video denoising, where the image is split into a low-rank and high-rank part. In particular, in this
paper the interference field is filtered out using sparsity as discriminant adopting a mixed minimum
1 norm and trace norm minimization algorithm. The methodology can be applied to both near and
far field measurement ranges. It could be an alternative to the systematic use of anechoic chambers
for antenna array testing.
Keywords: antenna array; near-field measurements; 5G communication; array diagnosis;

rank minimization; compressed sensing; antenna testing
1. Introduction
Sophisticated radiating systems as active array antennas, and massive MIMO arrays will play a
relevant role in the forthcoming 5G communication systems [1]. Due to the high levels of electronic
devices integration required in 5G antennas, no physical connectors are generally available. This yields
a radically new connectorless measurement paradigm in which over-the-air (OTA) measurements
will have a relevant role. Furthermore, mass production of these new antennas requires new, fast and
reliable antenna testing methods. In this framework, near-field measurements represent the most
interesting solution due to the accuracy and small dimension of the test set compared to far-field and
compact range antenna measurement systems.
In near-field measurements [2] the field radiated by the Antenna Under Test (AUT) is measured
on a scanning surface placed at short distance (5λ − 7λ, λ being the free space wavelength) from
the AUT in an anechoic chamber in order to avoid reflections and stray signals [3]. The propagation
process from near-field to far-field conditions is simulated by a proper software. Even if the scanning
area could be any sufficiently smooth surface, planar surfaces are most commonly adopted for array
antennas.
The main advantages of near-field set-ups are well known: they allow to perform accurate
measurements in a controlled environment. However, a further advantage, that is often underestimated,

J. Imaging 2019, 5, 51
is the possibility of using sophisticated data processing algorithms to reduce the cost of the measurement
process without affecting the accuracy. This possibility will be exploited in this paper with reference to
array diagnosis.
On the other hand, near-field measurements suffer from two main drawbacks. The first one is
the time required to collect the data on the observation surface using standard near-field set-ups [4].
With reference to this point, sparse recovery techniques have been recently proposed in the framework
of antenna diagnosis from planar near-field measurements [5–8] in order to reduce the number of
measured data and consequently the measurement time. The method has been successfully tested
from data acquired in anechoic chamber [9,10].
A further problem is the cost of large anechoic chambers. Regarding this point, a large effort
has been devoted to the reduction of the so-called truncation error, caused by a limited scanning
area, in order to use smaller and less expensive anechoic chambers [11–14]. However, a further and
more drastic solution is to perform measurements in a non anechoic environment, avoiding the use of
expensive anechoic chambers.
The aim of this contribution is to investigate a technique that avoids the use of expensive anechoic
chambers by filtering out the interference of undesired electromagnetic sources.
It is worth noting that other techniques for filtering interference signals in antenna measurements
have been proposed. A partial list is reported in the references section [15–20]. The strategies
followed in literature are based on a complete characterization of the environment in order to
subtract the environment response [15], on the equivalent source reconstruction using inverse linear
approaches [16], on the use of suitable base representations [19,20]. Large effort has also been devoted
to interference filtering from amplitude-only measurements [17,18]. Generally speaking, these methods
are ‘general purpose’, in the sense that they can be used in general near-field measurement systems,
and do not take explicitly into account the small number of failures in array testing. In the proposed
method this characteristic is explicitly exploited to reduce the set of possible array excitations, allowing
to identify the failures and to filter the interference contributions in the same step.
The basic idea is to use differential measurements in order to obtain a sparse representation of the
AUT excitations [21]. Such sparseness property of the radiating source is used as a-priori information
in order to distinguish the AUT contribution from the contribution of scattering objects in measured
data, that are characterized by a low rank field distribution.
It is worth noting that the method proposed in this paper strictly resembles the methods used in
image processing. In practice, the interference-affected near field measured on a surface is treated as a
(complex-data) image. This allows to use some modern image processing algorithms. In particular, two
strategies widely used in image processing are applied in this paper. The first one is the reduction of
the amount of information by acquiring only the innovation part of an image, as currently happens in
video processing, using a differential measurements. The second technique has been recently proposed
in video denoising, where the image is split into a low-rank and high-rank part [22]. In particular,
in this paper the interference field is filtered out using sparsity as discriminant adopting a mixed
minimum 1 norm and trace norm minimization algorithm.
2. Rank and Sparsity of the Feld Radiated by an Electric Dipole

Before introducing the filtering technique, it is useful to briefly discuss the general idea at the
base of the filtering procedure.
Let us consider a harmonic electromagnetic source consisting in an elementary electric dipole
directed along the y direction (Figure 1). The dipole can model an element of an array, or also a
scattering point caused by objects in the environment where the measurement system is placed.
The field of this dipole is observed on a square surface having dimension L × L (L = 20λ) placed on
the z = d plane with a uniform planar grid at 0.2 λ sampling step. The field on the observation points
is sampled and the data are collected on an equispaced grid and the measured values are collected in
the matrix X̄.
65
J. Imaging 2019, 5, 51
L
y
x L
z
source
Obsevation
plane
d
Figure 1. Geometry of the problem.
In Figure 2, the rank of the matrix X̄ is evaluated versus the distance d between the source and
the observation plane (blue curve, left scale). We can note that the rank rapidly decreases. As a
consequence, at sufficiently large distance X̄ tends to be a low rank matrix.
22 12000
20
10000
18
16 8000
normalized l 1 norm
14
rank
6000
12
10 4000
8
2000
6
4 0
0 10 20 30 40 50 60 70 80 90 100
d/lambda
Figure 2. Blue curve (left scale): rank of the field on the observation plane; red curve (right scale): 1
norm normalized to the maximum of the field amplitude on the observation plane; the observation
plane is 20λ × 20λ; d is the distance between the source point and the observation plane.
This ‘smoothing’ process in the propagation process is a general property of the field radiated
by an electromagnetic source [23]. Loosely speaking, the propagation acts as a spatial low pass filter,
smoothing the fast spatial variations of the field. Consequently, the field on the observation surface,
being smoother, tends to require less basis functions for its representation.
The spreading of the field on the observation surface caused by the filtering property of the
propagation phenomenon can also be quantified in terms of ‘sparseness’ of the field. For this purpose
we should evaluate the so called 0 norm, i.e., the number of elements of the matrix different from zero.
66
J. Imaging 2019, 5, 51
The 0 norm suffers from a number of drawbacks that prevent its use in practical problems. Instead of
the 0 norm, we will estimate the degree of sparsity using the 1-norm, or 1 norm, of the matrix X̄, i.e.,
the sum of all the amplitudes of the entries of the matrix. The more the field is concentrated on the
observation surface, the smaller the 1-norm is. This point will be discussed in more detail in the next
section. In this section the goal is to give an intuitive explanation of the usefulness of the 1 norm.
The 1 norm of the matrix, normalized to the maximum amplitude of the entries of the matrix,
is plotted in Figure 2 versus the distance d between the source and the measurement plane (red curve,
right scale). We can see that the normalized 1 norm of the matrix increases rapidly with distance,
as we should expect.
This simple example shows that rank and degree of sparsity can be used to distinguish contributions
from sources in different positions provided that the plane is positioned in the right position. In particular,
if the plane is placed very close to the ‘desired’ source, and sufficiently far from the interference source,
it is possible to filter the undesired sources by subtracting the low rank contribution.
This observation is at the basis of the method proposed in this contribution to filter undesired
field reflections.
3. The Array Failure Detection Algorithm with Reflection Filtering

Let us consider an Antenna Under Test (AUT) consisting of a planar N × N array affected by a
number of fault elements (Figure 3). Let Σ be the plane where the AUT aperture lies, and X AUT ∈ C N × N
the matrix collecting the currents of the radiating elements of the AUT.
In differential measurements [5], we consider also an array without failures, called ‘golden array’,
whose currents matrix is XGOLD . XGOLD (as well as the field radiated by XGOLD ) can be obtained
by full-wave numerical simulations, or by measurements in a controlled environment (i.e., in an
anechoic chamber).

Figure 3. Measurement set-up; the data are collected on the surface Ω placed at a distance d from
the AUT and are affected by a scattered field and Gaussian noise; some elements of the AUT are
malfunctioning (red squares).
The field of the AUT and of the Golden array are measured on a plane Ω placed at distance d
from the AUT in a square lattice of M × M points. The data are collected in the matrices Y AUT and
YGOLD , and the following quantities
X = X AUT − XGOLD (1)

Y = Y AUT − YGOLD (2)
are evaluated. Since the number of fault elements is much smaller than the number of elements of the
array, the X matrix is sparse [24].
67
J. Imaging 2019, 5, 51
Now, let us suppose that there is an interference source and let us call Xs the matrix collecting
the equivalent currents generated by the field of the interference source on the plane Σ, i.e., the plane
where the AUT is placed.
Consequently, the equivalent currents on Σ are given by the superposition of the equivalent
currents associated to the AUT and to the interference source, i.e., X AUT + Xs . Note that on this plane
Xs tends to be a low rank matrix as discussed in the previous section.
The field measured on the observation surface is the superposition of the field radiated by these
two contributions plus noise:
Ym = A(X AUT + Xs ) + Yn (3)
wherein A is the radiation operator, i.e., the operator mapping the equivalent current matrix on Σ
into the matrix collecting the field on Ω [21] and Yn is the matrix collecting the measurement noise
contribution at the receiver.
The differential measured matrix is consequently:
Ŷ = Ym − YGOLD = A(X + Xs ) + Yn (4)
Rigorously, in order to distinguish the sparse contribution of the AUT and the low rank
contribution of the interference source on Σ, the following problem must be solved:
min rank(Xs ) + αX0

subject to A(X + Xs ) − Ŷ2 ≤ (5)
wherein rank(X) is the rank of the matrix X, Xs 0 is the 0 norm of the matrix Xs , α is a regularization
parameter and depends on the level of the noise affecting the data.
Since both rank minimization and 0 minimization are non convex functions the solution of (5)
requires a computational expensive exhaustive search. Furthermore, 0 norm is instable in presence
of noise.
In order to solve the problem it is advantageous to substitute the original problem with a suitably
relaxed version.
In particular, the 0 norm can be substituted by the 1 entrywise matrix norm [24],
X1 = ∑ | xk,h | (6)

k,h
wherein xk,h is the (k, h) entry of the matrix X, while the rank function can be well approximated by
the trace norm (also called Schatten 1-norm or nuclear norm) [25,26]:
r
Xs ∗ = ∑ σk (7)
k =1
where σk is the k-th singular value of Xs and r is its rank [25,27].

It is interesting to note that nuclear norm and 1 norm have some similarities since in some way
the nuclear norm is to the rank functional what the convex 1 -norm is to the 0 -norm in the sparse
recovery area, Figure 4b. In fact, the nuclear norm can be seen as a relaxed version of the rank norm,
while the 1 norm can be considered a relaxed version of the 0 norm. While 0 norm counts the
number of elements different from zero, the 1 norm sums up their amplitude. In the same way,
while the rank function counts the number of non-zero singular values, the nuclear norm sums their
amplitude. In order to clarify this point, let us recall that in sparse recovery the goal is to identify
the sparsest vector (i.e., the vector having the largest number of null components) compatible with
the available data [24]. This requires to minimize the so-called 0 norm, wherein 0 is the number of
68
J. Imaging 2019, 5, 51
non null elements of the unknown vector. Such a minimization is a challenging non convex problem.
For the sake of simplicity, let us consider a 3 entries vector, x = { x, y, z}. The vector is supposed to be
1-sparse, i.e., only one of the three entries of the vector is different from zero. Let us consider the convex
hull of the 1-sparse vectors. Such a convex hull turns out to be the unit ball of the 1 norm, wherein
the 1 norm is x1 = | x | + |y| + |z|. A graphical picture of the unit 1 ball is drawn in Figure 4a.
The solution of the 1 minimization (red point in Figure 4a) is the tangent point between the affine
space associated to the available data (drawn as a red line in Figure 4a) and the scaled convex hull.
The minimization of the trace norm works in the same way, but operating on the singular values of
the matrix.
Figure 4. (a) geometrical picture of the 1 minimization; (b) geometrical picture of the trace
norm minimization.
Consequently, in practice the solution of the problem requires the following minimization
procedure involving two different definitions of matrix 1-norm:
min αXs ∗ + X1

subject to A(X + Xs ) − Ŷ2 ≤ (8)
i.e., a weighted minimization of the Schatten 1-norm (i.e., the trace norm) of Xs and of the entrywise
1-norm of X. The regularization parameter α can be estimated using the L-shape curve adopted also in
Tikhonov regularization [28].
The above minimization is a convex problem and can be solved by means of the powerful and
efficient algorithms available in many numerical libraries.
The algorithm can handle also multiple scattering interference sources. In this case, the field of
each interference source gives a low rank matrix on the observation plane. Consequently, the problem
is to identify a set of low-rank matrices and a sparse matrix. i.e.,
L
min ∑ αl Xsl ∗ + X1
l =1

L

subject to A X + ∑ Xsl − Ŷ ≤ (9)
l =1

2
where in L is the number of interference sources and Xsl is the low rank matrix associated to the l-th
interference source.
4. Numerical Examples
In this section some numerical results are shown. The AUT is a 7 × 7 planar array with λ/2
inter-element distance, centered on the x, y plane of a Cartesian coordinate system (see Figure 3).
69
J. Imaging 2019, 5, 51
The data are collected on a 21 × 21 points λ/2 uniform grid placed on the plane Ω at distance d = 7λ
from the AUT aperture plane. An undesired source is placed at { x = 0, y = 2.2λ, z = −8λ}. The data
are affected by −45 dB level Gaussian noise. We suppose that the AUT is affected by three fault
elements. The excitation of the non fault elements is one, while the three fault elements have zero
excitation. The amplitude of the excitations of the 49 radiating elements of the AUT are plotted in
Figure 5a (left figure) in false colors (1 = red, 0 = yellow).
The proposed filtering technique is applied to the measured data. The amplitude of the array
excitations is shown in Figure 5c (right figure). The three defects are clearly visible.
As a comparison, the same data have been elaborated using the method [5] consisting of 1
minimization without filtering procedure. The result is plotted in Figure 5b (central figure), showing a
less effective identification of the failures.
In particular, the presence of the undesired source makes two broken elements barely identifiable,
while the proposed technique is able to clearly identify all three elements.
The estimation algorithm is also stable compared to the noise level. For example, in Figure 6 the
estimation of the failures of the AUT is shown in case of −35 dB noise level. The figure shows that the
proposed method still gives acceptable results (Figure 6c, right figure), while the standard method fails
to identify at least one failure (Figure 6b, central figure).
In order to show the performance of the algorithm in case of multiple interference sources, two
sources placed at ( x = 0, y = 3.2λ, z = −8λ) and ( x = 4, y = 0λ, z = −10λ) are considered. The data
are corrupted by −45 dB level Gaussian noise The solution using the filtering technique is shown in
Figure 7c, while the solution not implementing the filtering method is shown in Figure 7b, confirming
again an improvement in the estimation of the differential excitation.
Finally, an example of identification of the failures in a larger array (81 radiating elements) is
reported in Figure 8. The plot shows that standard technique completely fails to identify the fault
elements, while the proposed technique is able to identify the area where two fault elements are placed.
The position of the third fault element is not detected, but the figure shows a variation of the excitations
on the left upper corner of the array.
In Table 1 the results of the simulations are briefly compared in a quantitative way. As figure of
merit, the Mean Square Error between the reference differential amplitude excitations of the AUT and
the amplitude excitation of the retrieved differential excitations are reported using the interference
source filtering algorithm (4th column and using standard algorithm (5th column of the Table).
The CPU time required by the filtering algorithm is reported in the 6th column. Loosely speaking, also
the quantitative parameter chosen shows an improvement in the differential excitation reconstruction
using the proposed algorithm. The improvement becomes more relevant increasing the number of
interference sources. For example, in the case of two sources, the MSE decreases from 1.5 dB to −3.7
dB. Even if this improvement is numerically lower than the 1 source case, it is practically much more
relevant, and allows to pass from no failure detection to an effective failure detection, as shown in
the previous section. Regarding the computation time, the examples were obtained on an Mc Air 11’
with i7 processor using CVX using only one core. The computation time is less than two minutes, and
increases almost linearly with the number of interference sources. However, these values are only
indicative, and can give an erroneous idea of the computational time required in real applications.
70
J. Imaging 2019, 5, 51
Table 1. First column: number of the example; second column: number of the elements of the AUT;
third column: number of interference sources; fourth column: Mean Square Error of the amplitude
of the differential excitations using the proposed technique; fifth column: Mean Square Error of the
amplitude of the differential excitations without using the proposed technique; sixth column: CPU
time (seconds) required by the filtering program.
Example Array Elements Number of Interf. MSE Filt MSE no Filt. CPU Time
1st 7×7 1 −7.5 dB −3.7 dB 95 s
2rd 7×7 1 −9.5 dB −3.7 dB 77 s
3rd 7×7 2 −3.7 dB 1.5 dB 167 s
4rth 9×9 1 −4.8 dB 3.9 dB 78 s
Figure 5. 1st example: normalized excitation amplitude of the radiating elements (linear scale in false
colors: yellow = null amplitude, red = unit amplitude); (a) exact array excitations; (b) excitations
obtained without filtering; (c) excitations obtained using the proposed filtering method; 7 × 7 planar
array with λ/2 inter-element distance, 21 × 21 measurement points, d = 7λ, measured data affected
by interference field radiated by a source placed at ( x = 0, y = 2.2λ, z = −8λ) and by −45 dB level
Gaussian noise.
Figure 6. 2nd example;: normalized excitation amplitude of the radiating elements (linear scale in
false colors: yellow = null amplitude, red = unit amplitude); (a) exact array excitations; (b) excitations
Gaussian noise.
71
J. Imaging 2019, 5, 51
a b c
-1 -1 -1
0 0 0
1 1 1
-1 0 1 -1 0 1 -1 0 1
Figure 7. 3rd example: normalized excitation amplitude of the radiating elements (linear scale in
by interference field radiated by a source placed at ( x = 0, y = 3.2λ, z = −8λ) and a source placed at
( x = 4, y = 0λ, z = −10λ). The data are corrupted by −45 dB level Gaussian noise.
Figure 8. 4th example: normalized excitation amplitude of the radiating elements (linear scale in
Gaussian noise.
5. Conclusions
In this paper a novel filtering algorithm of signals in planar near-field measurements is described.
The method allows the filtering strategy of interference sources in array diagnosis. The technique is
simple and numerically efficient, since it allows the use of convex minimization procedures.
The basic idea is to take advantage of the characteristics of the electromagnetic propagation in
terms of ‘spreading’ of the field distribution. Briefly, the equivalent current distribution on the array
is strongly concentrated on the radiating elements, while the contribution of interference sources
placed far from the plane of the array is smoother. Accordingly, it is possible to distinguish these two
contributions looking for a ‘sparse’ distribution and a ‘low rank’ distribution. Numerical examples
carried out in some simple cases confirm the effectiveness of this approach.
As discussed in the paper, the method proposed in this paper strictly resembles the methods used
in image processing. In practice, the interference-affected near field measured on a surface is treated as
a (complex-data) image. This allows to use some modern image processing algorithms. In particular,
two strategies widely used in image processing are applied in this paper. The first one is the reduction
72
J. Imaging 2019, 5, 51
the amount of information by acquiring only the innovation part of an image, as currently happens in
video processing, using differential measurements. The second technique has been recently proposed
in video denoising and splits the image into a high-sparse and high-rank part.
The results have been obtained using a small laptop computer and CVX program. As stressed
in the Introduction, the aim of this paper is to introduce the technique, and for this purpose the
computer and program adopted are acceptable. However, CVX is a slow program developed mainly
for research-stage applications. More powerful and efficient algorithms are available under payment.
The use of these algorithms on powerful parallel computers drops the computational time drastically.
Even if computational time is not an issue in near-field measurements, since it is usually a fraction
compared to the time required for data acquisition, the possibility of fast reconstruction opens the
thrilling possibility of online failure identification, i.e., identification of failures on-site while antenna
works [29], filtering the environmental noise of the site where the antenna is placed. This interesting
possibility encourages to continue the investigation on the rank properties of the field in the framework
of antenna measurements.
Author Contributions: Conceptualization M.D.M.; software, D.P. and M.L.; formal analysis G.P. and F.S.
Funding: This work was supported by the MIUR Program Dipartimenti di Eccellenza 2018–2022.
References
1. Agiwal, M.; Roy, A.; Saxena, N. Next Generation 5G Wireless Networks: A Comprehensive Survey.
IEEE Commun. Surv. Tutor. 2016, 18, 1617–1655. [CrossRef]
2. Yaghjian, A. An overview of near-field antenna measurements. IEEE Trans. Antennas Propag. 1986, 34, 30–45.
[CrossRef]
3. Newell, A.C. Error analysis Techniques for Planar Near-Field Measurements. IEEE Trans. Antennas Propag.
1988, 36, 754–768. [CrossRef]
4. Migliore, M.D. Near Field Antenna Measurement Sampling Strategies: From Linear to Nonlinear
Interpolation. Electronics 2018, 7, 257. [CrossRef]
5. Migliore, M.D. A Compressed Sensing Approach for Array Diagnosis From a Small Set of Near-Field
Measurements. IEEE Trans. Antennas Propag. 2011, 59, 2127–2133. [CrossRef]
6. Oliveri, G.; Carlin, M.; Massa A. Complex-Weight Sparse Linear Array Synthesis by Bayesian Compressive
Sampling. IEEE Trans. Antennas Propag. 2012, 60, 2309–2326. [CrossRef]
7. Oliveri, G.; Rocca, P.; Massa, A. Reliable diagnosis of large linear arrays—A bayesian compressive sensing
approach. IEEE Trans. Antennas Propag. 2012, 60, 4627–4636. [CrossRef]
8. Salucci, M.; Gelmini, A.; Oliveri, G.; Massa, A. Planar arrays diagnosis by means of an advanced Bayesian
compressive processing. IEEE Trans. Antennas Propag. 2018, 66, 5892–5906. [CrossRef]
9. Fuchs, B.; Le Coq, L.; Migliore, M.D. Fast Antenna Array Diagnosis from a Small Number of Far-Field
Measurements. IEEE Trans. Antennas Propag. 2016, 64, 2227–2235. [CrossRef]
10. Costanzo, S.; Borgia, A.; di Massa, G.; Pinchera, D.; Migliore, M.D. Radar Array Diagnosis from
Undersampled Data Using a Compressed Sensing/Sparse Recovery Technique. J. Electr. Comput. Eng.
2013, 2013, 627410. [CrossRef]
11. Bolomey, J.; Bucci, O.M.; Casavola, L.; D’Elia, G.; Migliore, M.D.; Ziyyat, A. Reduction of truncation
error in near-field measurements of antennas of base-station mobile communication systems. IEEE Trans.
12. Salucci, M.; Migliore, M.D.; Oliveri, G.; Massa, A. Antenna measurements-by-design for antenna qualification.
13. Martini, E.; Breinbjerg, O.; Maci, S. Reduction of Truncation Errors in Planar Near-Field Aperture Antenna
Measurements Using the Gerchberg-Papoulis Algorithm. IEEE Trans. Antennas Propag. 2008, 56, 3485–3493.
[CrossRef]
73
J. Imaging 2019, 5, 51
14. Migliore, M.D.; Salucci, M.; Rocca, P.; Massa, A. Truncation-Error Reduction in Antenna Near-Field
Measurements Using an Overcomplete Basis Representation. IEEE Antennas Wirel. Propag. Lett. 2019,
18, 283–287. [CrossRef]
15. Burnside, W.D.; Gupta, I.J. A method to reduce stray signal errors in antenna pattern measurements.
16. Alvarez, Y.; Las-Heras, F.; Pino, M.R. The sources reconstruction method for amplitude-only field
measurements. IEEE Trans. Antennas Propag. 2010, 58, 2776–2781. [CrossRef]
17. Brown, T.; Jeffrey, I.; Mojabi, P. Multiplicatively regularized source reconstruction method for phaseless
planar near-field antenna measurements. IEEE Trans. Antennas Propag. 2017, 65, 2020–2031. [CrossRef]
18. Quijano, J.L.A.; Vecchi, G. Field and source equivalence in source reconstruction on 3D surfaces.
Prog. Electromagn. Res. 2010, 103, 67–100. [CrossRef]
19. Bucci, O.M.; D’Elia, G.; Migliore, M.D. A general and effective clutter filtering strategy in near-field antenna
measurements. Proc. IEE Microw. Antennas Propag. 2004, 151, 227–235. [CrossRef]
20. Quijano, J.L.A.; Vecchi, G.; Li, L.; Sabbadini, M.; Scialacqua, L.; Bencivenga, B.; Mioc, F.; Foged, L.J. 3D spatial
filtering applications in spherical near field antenna measurements. In Proceedings of the AMTA 2010
Symposium, Atlanta, GA, USA, 10–15 October 2010.
21. Migliore, M.D. On the Sampling of the Electromagnetic Field Radiated by Sparse Sources. IEEE Trans.
22. Ji, H.; Liu, C.; Shen, Z.; Xu, Y. Robust video denoising using low rank matrix completion. In Proceedings of
the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, San Francisco, CA,
USA, 13–18 June 2010; pp. 1791–1798.
23. Migliore, M.D. Minimum Trace Norm Regularization (MTNR) in Electromagnetic Inverse Problems.
24. Migliore, M.D. A simple introduction to compressed sensing/sparse recovery with applications in antenna
measurements. IEEE Antennas Propag. Mag. 2014, 56, 14–26. [CrossRef]
25. Candes, E.J.; Tao, T. The Power of Convex Relaxation: Near-Optimal Matrix Completion. IEEE Trans. Inf.
Theory 2010, 56, 2053–2080. [CrossRef]
26. Fuchs, B.; Le Coq, L.; Migliore, M.D. On the Interpolation of Electromagnetic Near Field without Prior
Knowledge of the Radiating Source. IEEE Trans. Inf. Theory 2017, 65, 3568–3574. [CrossRef]
27. Keshavan, R.H.; Montanari, A.; Oh, S. Matrix Completion From a Few Entries. IEEE Trans. Inf. Theory 2010,
56, 2980–2998. [CrossRef]
28. Brancaccio, A.; Migliore, M.D. A Simple and Effective Inverse Source Reconstruction with Minimum a Priori
Information on the Source. IEEE Geosci. Remote Sens. Lett. 2017, 14, 454–458. [CrossRef]
29. Migliore, M.D.; Pinchera, D.; Lucido, M.; Schettino, F.; Panariello, G. A Sparse Recovery Approach for
Pattern Correction of Active Arrays in Presence of Element Failures. IEEE Antennas Wirel. Propag. 2015, 14,
1027–1030. [CrossRef]
74
Journal of
Imaging
Article
Qualitative Methods for the Inverse Obstacle
Problem: A Comparison on Experimental Data
Martina T. Bevacqua 1,2, *,† and Roberta Palmeri 1,†
1 DIIES, Department of Information Engineering, Infrastructures and Sustainable Energy, Università
Mediterranea di Reggio Calabria, via Graziella, Loc. Feo di Vito, 89124 Reggio Calabria, Italy;
[email protected]
2 CNR-IREA, National Research Council of Italy, Institute of Electromagnetic Sensing of the Environment,
via Diocleziano 328, 80124 Napoli, Italy
* Correspondence: [email protected]; Tel.: +39-09-651-693-262
† The paper is a result of a joint contribution of both the authors, from conceptualization to final writing
and editing.
Received: 7 March 2019; Accepted: 8 April 2019; Published: 10 April 2019
Abstract: Qualitative methods are widely used for the solution of inverse obstacle problems. They
allow one to retrieve the morphological properties of the unknown targets from the scattered field by
avoiding dealing with the problem in its full non-linearity and considering a simplified mathematical
model with a lower computational burden. Very many qualitative approaches have been proposed
in the literature. In this paper, a comparison is performed in terms of performance amongst three
different qualitative methods, i.e., the linear sampling method, the orthogonality sampling method,
and a recently introduced method based on joint sparsity and equivalence principles. In particular,
the analysis is focused on the inversion of experimental data and considers a wide range of (distinct)
working frequencies and different kinds of scattering experiments.
Keywords: inverse obstacles problem; inverse source problem; joint sparsity; linear sampling method;
microwave imaging; orthogonality sampling method
1. Introduction
The inverse scattering problem aims to retrieve the electromagnetic properties of unknown targets
by processing the field they scatter [1,2]. This problem can be of interest in biomedical imaging in
order to discriminate between healthy and malignant biological tissues, identify anomalies inside the
human body, and build an electromagnetic model of a specific anatomical district [3–5]. Moreover, it is
relevant in other applications, such as underground prospecting as well as non-destructive testing of
materials, in order to detect and characterize buried targets and internal defects [6,7].
Due to the non-linearity and ill-posedness of this class of inverse problems [1,8], the quest to
develop reliable, accurate, and effective solution methods is still ongoing. As witnessed by many
papers published in literature, several researchers and scholars are very active on this topic, covering
its challenging computational, algorithmic, modeling, and experimental aspects.
Due to the mathematical difficulties of the problem, many studies have been focused on the
introduction of solution methods for the corresponding inverse obstacle problem, whose scope is to
recover just the morphological information about the targets, namely their supports, by giving up their
electromagnetic properties [9–16]. Such methods are known as qualitative methods and are different
from quantitative methods, which aim at retrieving both electrical and morphological properties of
the region of interest [2,9]. Indeed, they tackle the problem in its full non-linearity, without involving
any approximation and allowing for a widening of the class of retrievable objects; unfortunately, such
methods are characterized by a higher computational burden and longer processing time [2]. On the

J. Imaging 2019, 5, 47
other hand, qualitative methods are characterized by simplified mathematical models and a lower
computational burden [9].
Among qualitative methods, it is worth mentioning sampling methods that include the linear
sampling method (LSM) [12] (and the related factorization method [13]), as well as the orthogonality
sampling method (OSM) [14]. The main idea of these methods is that of sampling the investigation
domain in an arbitrary grid of points and computing at each point an indicator function whose
energy will assume different values depending on whether the sampled point belongs or not to the
scatter. Although both LSM and OSM belong to the class of sampling methods, they exhibit different
features. For instance, the LSM indicator function is computed by solving an auxiliary linear problem
whose solution involves the adoption of a regularization technique. In a different fashion, the OSM
does not require the solution of a linear problem since the indicator function is just related to the
evaluation of a scalar product. Moreover, the OSM seems to be more flexible with respect to the
measurement configurations.
By taking advantage of recent results in the area of compressive sensing and sparsity promoting
techniques [17], a new qualitative method has been very recently introduced in the literature that,
provided some conditions hold true, is able to retrieve the boundary of unknown targets rather than
their support [15,16]. Such a method exploits the equivalence theorem in the case of dielectric obstacles
and takes advantage of the particular distribution assumed by the induced currents in case of perfect
electric conductors [18]. Differently from sampling methods, it does not sample the investigation
domain in a grid of points, and it does not require the selection of a fixed threshold to discriminate
between points inside and outside the targets. However, if compared with LSM and OSM, it shows a
heavier computational burden.
Encouraged by the remarkable properties of the methods discussed above, in the following,
some comparisons and performance analysis are provided. In particular, the tests are performed
against experimental data [19] and consider a wide range of working frequencies and different kinds
of (monochromatic) data.
The paper is organized as follows. In Section 2, the inverse scattering problem is formulated.
Moreover, a brief summary of the three considered qualitative methods is given, including a discussion
on their expected advantages and disadvantages. In Section 3, some examples considering experimental
Fresnel data are reported. Finally, conclusions follow. Throughout the paper we consider the canonical
2D scalar problem (transverse magnetic (TM) polarized fields) and we assume and drop the time

harmonic factor exp jωt .
2. Materials and Methods
2.1. Mathematical Formulation of the Problem

Let D denote the region being tested where targets with support Σ are located.
Let χ(r) = s (r)/b − 1 be the contrast function that relates the electromagnetic properties of
the scatterers to those of the host medium, wherein s and b are the complex permittivities of the
scatterers and the background medium, respectively. Let Es ∞ (r̂m , r̂t ) be the scattered far-field pattern
measured on a closed circle curve Γ in the far zone of the scatterers in direction r̂m when a unit
amplitude plane wave impinges from direction r̂t . Then, the equations describing the scattering
problem can be expressed as [1]:

Es ∞ (r̂m , r̂t ) = Gb ∞ (r̂m , r)W (r , r̂t )dr = Ae [W ] (1)
D

E(r, r̂t ) = Ei (r, r̂t ) + Gb (r , r)W (r , r̂t )dr = Ei + Ai [W ] (2)
D
76
J. Imaging 2019, 5, 47
where Ei and E are the incident and total field, respectively, W = χE is the contrast source, and r = (x, y)
scans the investigation domain D. Moreover, Gb ∞ is the Green’s function in the far-field zone pertaining
to the background medium. Finally, Ae and Ai are short notation for the integral radiation operators.
In the inverse obstacle problem, one aims at retrieving the support Σ of the target, by giving up
its electromagnetic properties, from the (noise corrupted) measured data Es ∞ when the targets are
illuminated by a known incident field Ei [9]. Such a problem is non-linear as well as ill-posed because
of the properties of the involved operator Ae [1,8].
2.2. Linear Sampling Method

The LSM is one of the most popular qualitative methods to retrieve objects’ support from the
measurements of the corresponding scattered field data [12]. More in detail, it consists in sampling
the region under test into an arbitrary grid of points and solving for each point the so-called far field
integral equation given by:
Es ∞ (r̂m , r̂t ) ξ(rs , r̂t ) dr̂t = Gb ∞ (r̂m , rs ) (3)
Γ
wherein ξ represents the unknown function and rs the sampling point. Despite the linearity of
Equation (3), the evaluation of the solution ξ is not straightforward due to the compactness of
the operator at the left hand side, which implies ill-posedness in its inversion [12,20]. Hence, a
regularization technique is required. An effective choice is that of solving it by adopting the Tikhonov
regularization [20]. Then, the estimation of the unknown support is pursued by evaluating the l2 norm
(i.e., the “energy”) of the unknown function ξ with respect to rt , given by:
2

ILSM (rs ) = ξ(rs , r̂t )22 = ξ(rs , r̂t ) dr̂t (4)
Γ
The above defined indicator depends on the sampling point rs and assumes a finite value when the
sampling point belongs to the unknown object, while it blows up elsewhere [12]. As such, by selecting
a fixed threshold, the indicator described in Equation (4) allows to discriminate between points inside
and outside the scatterers and finally retrieve Σ [20]. It is worth noting that the computational burden
is low, as the solution of Equation (4) just involves the evaluation of the singular values decomposition
(SVD) of the measured data [20].
2.3. Orthogonality Sampling Method

The OSM has recently been introduced in the literature [14]. It consists in computing the reduced
scattered field from the far-field pattern Es ∞ (rm , rt ). Such a reduced field can be computed by evaluating
the scalar product between the far field and the Green’s function in far-field zone, that is [14]:

1 ∞
Es red (rs , r̂t ) = Es ∞ (r̂m , r̂t ) e jkb rs ·r̂m dr̂m = Es , Gb ∞ Γ (5)
Γ γ
where ·, · denotes the scalar product and γ is a constant for a fixed frequency [14]. In a nutshell,
Equation (5) tests the orthogonality relation between the far-field pattern and the Green’s function in
the far-field zone (apart from a constant). Then, the OSM indicator IOSM (rs ) is defined as:
2
red
IOSM (rs ) = Es red rs , r̂t 22 = Es (rs , r̂t ) dr̂t (6)
Γ
and achieves large values when the far-field pattern approaches the one for the background Green’s
function for the considered sampling point [14]. Again, by selecting a fixed threshold, it is possible to
discriminate sampling points belonging to the target support [14].
77
J. Imaging 2019, 5, 47
2.4. Shape Reconstruction Via Joint Sparsity Based Inverse Source and Equivalence Principles
In order to deal with inverse obstacle problems, one can also take advantage of the joint solution
of a number of related inverse source problems that consist of recovering the induced currents from the
knowledge of the measured far field pattern Es ∞ [15,16]. The strategy is obviously suggested by the
fact that the support of W is exactly equal to Σ whatever incident field Ei is. Note that inverse source
problems are linear (see Equation (1)) but are still (severely) ill-posed such that some additional care
is needed.
In this respect, note that when a generic electromagnetic field is propagating in the space in the
presence of dielectric obstacles, it induces some currents inside the obstacles, which are expected to be
different from zero inside Σ and in turn become sources of the scattered field data. However, by virtue
of the equivalence theorem [18] (which is a key point of the approach), the measured scattered field
can be eventually conceived as the one radiated by some equivalent surface currents, i.e., electric Ws
and magnetic Wsmx and Wsmy surface currents, which can play the role of the original ones.
The main advantage of considering the equivalent currents rather than the original ones is that
they are distributed over the boundary of Σ. As a consequence, they can be identified by only a few
non-zero pixels in the investigation domain and, hence, can be considered sparse with respect to the
pixel basis representation. Then, a possible idea for their retrieval is that of looking for the sparsest
current distributions, which contemporarily satisfy the measured data. To this end, sparsity promoting
approaches can be exploited [17].
Interestingly, the property of sparsity displayed by the equivalent currents holds true regardless
of the direction of illumination r̂t or the kind of measurement configuration. In order to enforce this
joint sparsity among the different scattering experiments, an auxiliary variable B can be defined as
the common upper bound on the amplitude of the equivalent currents for the different scattering
illumination conditions. Accordingly, the problem is recast as [15,16]:
minB(r)1 (7a)

s.t. Es ∞ − Ae [Ws ] − Amx [Wsmx ] − Amy Wsmy ≤ δ (7b)
2

Ws (r, r̂t ) ≤ B(r)

Wsmx (r, r̂t ) ≤ B(r), ∀ r̂t ∈ Γ (7c)
Wsmy (r, r̂t ) ≤ B(r)
where ·1 is the l1 norm, and Amx and Amy are the integral radiation operators that relate the magnetic
surface currents to the scattered electric fields.
The variable B does not depend on the direction of illumination r̂t , but only on the position r in
the investigation domain. Moreover, it is expected to assume non-zero values only on the boundary of
Σ such that it will act as a “boundary” indicator. For this reason, the method is called boundary retrieval
through inverse source and sparsity (B-IS). As a consequence, differently from LSM and OSM, there is
no need for a thresholding to retrieve the support from the indicator map, as the approach directly
retrieves the boundary of the targets.
Note the approach can be also applied to the case of perfect electrical conductors. In fact, in such
a case, the induced currents W exist only on the boundary of Σ and there is no need for introducing
equivalent surface currents. Then, the optimization problem (7) is further simplified [15,16].
As a final comment, note that the B-IS method belongs to the class of the joint sparsity-promoting
procedures. In particular, the joint sparsity used herein is not enforced by means of the mixed norm
l1 − l2 as in References [21,22], but through the definition of the auxiliary variable B.
2.5. Applicability and Limititations of the Three Methods

In LSM, the far field Equation (3) can be interpreted as an attempt to synthetize induced currents
such that their radiating part is focused and/or circularly symmetric with respect to the given sampling
78
J. Imaging 2019, 5, 47
point [20,23]. As such, it correctly works as long as the object is convex and the non-radiating
component of the induced currents is negligible [20]. Usually, when the product of the maximum
amplitude of the contrast χ times the maximum size of the target is large, which corresponds to an
increasing amount of non-radiating sources, it does not guarantee reliable reconstructions.
The OSM is based on a test of orthogonality, which allows one to compute the reduced scattered
field. As far as its physical interpretation is concerned, it has been recently argued in Reference [24]
that the reduced scattered field is related to the radiative part of the currents. As such, similarly to
LSM, the OSM indicator does not take into account the non-radiating part of the currents, so that,
again, limitations are expected with increasing electrical dimensions of the scatterer.
While both the LSM and OSM act on volumetric currents, which are located on all points belonging
to the support of the target, B-IS aims at jointly solving several inverse source problems, wherein
one is looking for currents that, under some reasonable hypotheses, are located on the boundary of
Σ. Notably, the optimization problem (7) does not aim at retrieving exactly the equivalent currents
but rather to find a common upper bound to their amplitudes. Also note the possible (superficial)
non-radiating currents do not seem to play any role in the process, while the possible non-radiating
volumetric currents are filtered out by the sparsity regularization. For all the above, the B-IS method
can be understood as more robust with respect to the presence of non-radiating currents than LSM and
OSM. Moreover, the upper bound in Equation (7) can be understood as a trade-off between sparsity
promoting and data fitting. As a consequence, B-IS will be able to retrieve at least the convex-hull of
the targets. Remarkably, probably because of the use of the l1 norm rather than the l0 pseudo-norm,
one can also be able to retrieve the shape of some non-convex target [15]. As a consequence of all the
above, B-IS is expected to fail when the boundary of the target is not sparse (for instance in case of a
ragged boundary), or when, according to the sparsity theory [17], the number of measurements is not
sufficiently large with respect to the degree of sparsity of the target contour.
Concerning the requirements about the kind of measured data, the LSM needs to correctly discretize
the integral operator in Equation (3) by collecting a sufficient number of data in multiview-multistatic
measurement configuration. As a consequence, even if the product of the maximum amplitude of the
contrast χ times the maximum size of the target is not large, the undersampling of the measured data can
compromise the reconstructions. On the other side, the OSM requires an accurate discretization of the
integral in Equation (5) over the measurements curve, and hence, a sufficient number of measurements.
However, differently from LSM, OSM can compensate the low number of data points by considering
other kind of experiments different from multiview-multistatic ones. In fact, OSM and B-IS exhibit
wide flexibility with respect to processed data and can offer the possibility of combining information
arising from frequency diversity without the need of a posteriori merging the single results. To this
end, the OSM indicator can be easily modified by integrating over a given frequency band [14,24], and
the B-IS indicator function implicitly defined by Equation (7c) can be redefined as the upper bound for
the different sources at the different frequencies [15].
Another important difference among the methods regards the computational burden. The OSM
indicator does not involve the solution of an ill-posed problem but only the evaluation of a scalar
product. This implies more robustness with respect to measurement errors on the scattered data and a
very low computation burden. On the contrary, the LSM indicator function is computed by solving an
auxiliary linear problem, whose solution involves the adoption of a regularization technique. However,
the main computational effort of the method is the evaluation of the SVD of measured data, whose
dimension is only dictated by the number of scattering experiments. This task has to be carried out
only once, since the kernel of the linear system is the same for all sampling points. With respect to the
LSM and the OSM, B-IS exhibits a heavier computational burden, which drastically increases with
the number of unknowns, since an l1 norm optimization problem has to be solved. On the other side,
the problem underlying the B-IS belongs to the class of convex programming problems, whose single
optimal solution can be reached using any common local optimization procedure.
79
J. Imaging 2019, 5, 47
In terms of reconstruction accuracy, both the LSM and OSM require the selection of a fixed
threshold to discriminate between points inside and outside the targets. In a different fashion, the
B-IS does not provide an indicator map but directly looks for the boundary of the target. As such, it is
expected to be more accurate.
These aspects, which are briefly summarized in Table 1, will be further investigated in the
following numerical section.
Table 1. Summary of strong and weak points of the three solution methods.
LSM OSM B-IS

Reconstruction accuracy medium medium high
Computational burden low very low high
Flexibility with respect to the kind of data low high high
3. Results
In this section, an accurate comparative study was performed against the experimental data set
provided by the Institute Fresnel of Marseille [19], typically adopted to benchmark inverse scattering
procedures. In particular, four targets from the 2001 Fresnel database were considered, i.e.,:
• the DielTM target, which consists of a dielectric homogeneous cylinder of radius 1.5 cm and
relative permittivity 3 ± 0.3
• the RectTM_Dece target, which is a rectangular metallic target of 25.4 mm × 12.7 mm not centered
with respect to the azimuthal positioner axis;
• the U-TM shaped target, which is a metallic U-shaped target with dimension 80 × 50 mm2
• the TwinDielTM target, which consists of two identical dielectric homogeneous cylinders of radius
1.5 cm and relative permittivity 3 ± 0.3.
The complete description of the targets and the measurement set-up can be found in Reference [19],
while a schematic sketch is shown in Figure 1. The data were collected under a partially aspect-limited
configuration, where primary sources completely surrounded the targets, but for each illumination,
the measurements were taken only on an angular sector of 240◦ achieved by excluding the 120◦ sector
centered on the incidence direction.
Figure 1. Targets from 2001 Fresnel dataset: (a) DielTM target; (b) RectTM_Dece target; (c) U-TM-shaped
target, and (d) TwinDielTM target.
For the dielectric targets, the working frequencies could change among the interval 1–8 GHz with
a step of 1 GHz. Also, 12 GHz and 16 GHz data were available. On the other side, for the metallic
targets the working frequencies can change among the interval 2–16 GHz with a step of 2 GHz. For the
sake of brevity, only some significant reconstructions corresponding to some frequencies will be shown
in the following. Moreover, for each target, two different datasets have been considered. The first one
consisted of a 36 × 36 multiview-multistatic data matrix (the number of measurements and views M is
equal to 36), while the second one corresponded to a data matrix with halved dimensions, that is M
equal to 18. The investigated area of 0.1775 m × 0.1775 m was discretized in 40 × 40 cells.
In the following examples the single frequency indicators were normalized with respect to their
maximum values. In particular, the LSM indicator were rescaled as described in Reference [25].
80
J. Imaging 2019, 5, 47
Note that the number of independent data were lower with respect to the numbers of unknowns.
This is a common problem of all inverse scattering and inverse obstacles problems. In fact, one is not
able to collect data at will, as only a finite number of independent experiments can be performed [26].
However, the qualitative methods we were comparing did not look for a quantitative reconstruction in
every pixel of the scenario, but they looked for some qualitative behavior of the indicator. Moreover,
all three methods somehow exploited additional regularizations.
3.1. Convex Dielectric Target

The results corresponding to the single dielectric circular cylinder are reported in Figure 2. As can
be seen, at the lowest frequencies, the three methods could localize the cylinder (whose diameter was
0.1λb at 1 GHz and 0.2λb at 2 GHz (λb was the wavelength in free space)). Its size was overestimated,
especially by LSM and OSM, while instead the B-IS method was not able to correctly identify its shape.
In the frequency range of 3–8 GHz, the performance of the three methods improved and the support
was correctly retrieved. This circumstance also held true at higher frequencies, although some artifacts
and ambiguities started to emerge. For instance, at 12 GHz, the OSM indicator was characterized
by a hole at the center of the cylinder, while at 16 GHz, some artifacts appeared in the background.
On the other hand, the B-IS indicator wrongly suggested some pixels inside the target, but such a
circumstance did not compromise the correct estimation of the target support.
When a reduced amount of data was processed, both OSM and LSM did not work fine at the
frequencies of 12 and 16 GHz, as shown in Figure 3. This was due to the fact that the amount of data,
which was needed in order to correctly discretize the far field Equation (3), as well as to correctly
compute the reduced scattered field Equation (5), was larger than M, whereas the size of the target
became large in terms of wavelength. The B-IS was instead robust with respect to the reduction of data.
In fact, the same performance (with respect to the case M = 36) was guaranteed when M = 18.
In order to quantitatively assess the quality of the reconstructions, the following dimensional
error errΔ was evaluated:
rrec − rre f
errΔ = (8)
rre f
where rrec and rre f being the radius of the reconstructed and the reference cylinders, respectively. This
metric allows one to appreciate the performance of the method in reconstructing the size of the target.
In evaluating these errors, two different methods have been adopted to binarize the LSM and OSM
indicator map. As the indicator maps are continuous functions, we first binarized them by applying
the Canny edge detector [27,28] to its dB plot and set it to one of the pixels internal to the identified
contour (see Table 2). The second method instead consisted of replacing all pixels in the input image
with value greater than a given level L ∈ ] 0, 1] with the value 1 and replaced all other pixels with
the value 0 (see Table 3). Note that the B-IS method did not require the adoption of a binarization
technique. By comparing Tables 2 and 3, one can conclude that:
• the B-IS was more accurate in estimating the radius of the cylinder and it was robust with respect
to the reduction of M;
• the choice of the thresholding techniques could impact the reconstruction of the support of
the target. Indeed, the dimensional errors in Table 2 were different from the one in Table 3.
In particular, the Canny edge detector-based approach led to an overestimation of the radius of
the cylinder, while the second approach (L = 0.8) implied more accurate estimations, as witnessed
by the lower dimensional errors. Notably, the accuracy of the reconstructions when LSM and
OSM were adopted strongly depended on how the indicator map was binarized and on the
adopted threshold.
Finally, in order to emphasize the flexibility of OSM and B-IS with respect to the measurement
configuration, the inversion of a 72 × 9 data matrix was performed. In such a case the number of
experiments was reduced to 9, while the measurements were increased to 72. The corresponding
81
J. Imaging 2019, 5, 47
indicators are reported in Figure 4. B-IS was able to quite accurately retrieve the target support,
while LSM only localized the target. Finally, differently from LSM, OSM was able to provide a
rough reconstruction of the support, as it was sensitive with respect to the number of measurements
rather than to the number of experiments.
Figure 2. The DielTM target. From left to right: 2 GHz, 6 GHz, 12 GHz, and 16 GHz. From top to
bottom: LSM, OSM, and B-IS indicators. The number of measurements and views was M = 36.
Figure 3. The DielTM target. From left to right: 2 GHz, 6 GHz, 12 GHz, and 16 GHz. From top to
82
J. Imaging 2019, 5, 47
Table 2. Dimensional errors for the DielTM target. LSM and OSM indicators were binarized by
applying the Canny edge detector. The symbol * indicates the cases wherein the indicator maps were
not reliable, and the dimensional error was not computed.
LSM OSM B-IS

Freq (GHz)
M = 36 M = 18 M = 36 M = 18 M = 36 M = 18
2 1.4 0.47 0.97 0.97 0.18 0.18
6 0.55 0.55 0.69 0.74 0.18 0.11
12 1 * 1 * 0.18 0.18
16 1 * 0.18 * 0.18 0.11
Table 3. Dimensional errors for the DielTM target. LSM and OSM indicators were binarized by
selecting L = 0.8. The symbol * indicates the cases wherein the indicator maps were not reliable, and the
dimensional error was not computed.
LSM OSM B-IS

Freq (GHz)
M = 36 M = 18 M = 36 M = 18 M = 36 M = 18
2 0.37 0.32 0.11 0.11 0.18 0.18
6 0.22 0.09 0.27 0.28 0.18 0.11
12 0.47 * 0.47 * 0.18 0.18
16 0.03 * 0.47 * 0.18 0.11
Figure 4. The DielTM target at 12 GHz. From left to right: LSM, OSM, and B-IS indicators. The dimension
of the data matrix was 72 × 9.
3.2. Convex Metallic Target

Figure 5 depicts the reconstructions corresponding to the metallic rectangular targets pertaining
to 2, 6, 10, and 16 GHz and M = 18. As can be observed, at the lowest frequency, the methods were just
able to localize the scatterer, especially the B-IS method, which identified only some pixels in the center
of the target. In fact, at 2 GHz, the target was just 0.2λb × 0.1λb large. By increasing the frequency,
the capability of the methods in retrieving the shape of the metallic target improved. For instance, at
6 GHz, both LSM and OSM correctly retrieved the unknown support, while B-IS overestimated its
vertical dimension. However, in the frequency range of 8–16 GHz, both LSM and OSM came to exhibit
some artifacts and did not retrieve an indicator map that was monotonically decreasing outside the
actual support of the scatterer. This was correlated to the fact that the electromagnetic size of the target
became larger and the amount of data needed to correctly solve Equation (3), as well as compute the
reduced field Equation (5), was much larger than M. Indeed, by considering M = 36, reliable results,
which are not shown for the sake of brevity, could be achieved. Notably, when M = 18, the B-IS was
still able to reconstruct the actual support of the rectangular object also in the range of 8–16 GHz, as
shown in Figure 5k,l.
83
J. Imaging 2019, 5, 47
Figure 5. The RectTM_Dece target. From left to right: 2 GHz, 6 GHz, 10 GHz, and 16 GHz. From top to
3.3. Non-Convex Metallic Target

The same behavior observed in the previous examples also holds true in case of the U-TM-shaped
target (see Figure 6). In fact, at 2 GHz, the target was detected by all three methods and approximately
localized by LSM and OSM, but only its convex-hull was retrieved by LSM and OSM. On the other
hand, at higher frequencies, the support was only approximately retrieved by all three methods.
In particular, some artifacts were present due to the mutual interactions between the “arms” of the
target, especially in the LSM and OSM indicators, which were sensitive to the non-radiating part of the
currents. Note that such a target belongs to the class of non-convex targets, so that it represents a more
challenging case with respect to the previous ones.
Figure 6. The U-TM-shaped target. From left to right: 2 GHz, 6 GHz, and 10 GHz. From top to bottom:
LSM, OSM, and B-IS indicators. The number of measurements and views was M = 18.
84
J. Imaging 2019, 5, 47
No significant reconstructions were obtained by means of the three methods for M = 18 and
a working frequency belonging to 12–16 GHz. In case of M = 36, the reconstructions are shown in
Figure 7. As it can be seen, B-IS correctly retrieved the support of the target, apart from some spurious
pixels, while the LSM did not provide reliable results due to the relevant presence of non-radiating
currents and a high degree of non-linearity. Finally, the OSM indicator provided a rough support
estimation only at 12 GHz.
Figure 7. The U-TM-shaped target. From top to bottom: 12 GHz and 16 GHz. From left to right: LSM,
OSM, and B-IS indicators. The number of measurements and views was 36.
3.4. Multiple Dielectric Targets

The results pertaining to the two identical cylinders are shown in Figure 8 for a number of
experiments and measurements corresponding to M = 18. At 1 GHz, the methods could only detect
the presence of an obstacle, but they could not identify the presence of two disjoint objects. On the
contrary, at 2GHz, it was possible to distinguish two different targets; in particular, from the OSM map,
one could also understand their shape. Note that the distance between the targets was 60 mm, which
corresponded to 0.2λb at 1 GHz and 0.4λb at 2 GHz. When the frequency increased, both LSM and
OSM worked fine. However, when the frequency was 6 GHz, some undesired artifacts occurred.
On the contrary, the B-IS method is able to provide quite accurate reconstructions of the support
also in the frequency range 6–8 GHz. When the frequency and the size of the overall target’s system
were further increased, completely meaningless reconstructions were obtained by LSM and OSM, while
B-IS indicator at least localized the two targets (see Figure 9g). In order to improve the reconstruction
accuracy of the three methods, it was necessary to process a higher amount of data. This was witnessed
by the reconstructions reported in Figure 9 corresponding to M = 36 at both 8 GHz and 16 GHz.
However, at 16 GHz, the increase of the number of data was not enough for LSM and OSM to provide
indicator maps free of artifacts.
85
J. Imaging 2019, 5, 47
Figure 8. The TwinDielTM target. From left to right: 1 GHz, 2 GHz, 6 GHz, and 8 GHz. From top to
Figure 9. The TwinDielTM target. From left to right: 16 GHz (M = 18), 8 GHz (M = 36), and 16 GHz
(M = 36). From top to bottom: LSM, OSM, and boundary indicators. The number of measurements
and views was M = 18 for the first column and M = 36 for the last two columns.
86
J. Imaging 2019, 5, 47
4. Conclusions
The paper contributes to understand and discuss some strong and weak points of three qualitative
methods for the inverse obstacle problem. In particular, a comparison among the popular LSM,
the OSM, and the more recent B-IS was reported by considering the processing of the 2001 Fresnel
dataset, which included both dielectric and metallic targets, as well as non-convex and multiple
targets. Different frequencies and amount of data were considered for each method. In all cases,
because of the characteristics of the Fresnel measurements setup, data were aspect-limited, i.e., not all
incidence observation angle couples were available (for instance the monostatic data was missing).
Also, scattered field data turned out to be undersampled at the higher frequencies with respect to the
criteria (depending on the dimension of the investigation domain) discussed in Reference [26].
Results suggested that at lower frequencies, reliable reconstructions were guaranteed if the size of
the target was higher than 0.2λb . Indeed, for lower dimensions, the three methods could only localize
the target, while its size was overestimated. Such a result was fully consistent with expected limitations
on resolution implied by the diffraction theorem [29] and the results in Reference [26]. When the
frequency was much larger, LSM and OSM began to exhibit some artifacts and, if the amount of data
was not sufficient or due to the presence of non-negligible non-radiating sources, they could provide
unreliable reconstructions. Notably, the B-IS seemed to be more robust with respect to LSM and OSM
when undersampling data and/or increasing the frequency. In fact, it could still provide accurate
indicator maps in a number of cases where LSM and OSM failed.
The examples have also proved that both OSM and LSM were sensitive to the choice of the
binarization technique to discriminate between points inside and outside the scatterer. On the other
hand, both OSM and LSM had a negligible computational burden, while B-IS had to solve a CP problem
whose computational burden increased with the cells used to discretize D, as well as with M. In the
considered examples, the computational time amounted to a few seconds for OSM and LSM, and to a
few minutes for B-IS.
Future work will be devoted to extending such a performance comparison to other qualitative
methods, for instance, the modified version of LSM [30,31].
Author Contributions: The authors contribute equally.

Funding: This work was partly supported by the Italian Ministry of Education, Universities, and Research through
the Project of National Interest (PRIN) ‘Field and Temperature Shaping for MWI Hyperthermia - FAT SAMMY’
under Grant 2015KJE87K.
Acknowledgments: The authors want to acknowledge Prof. Tommaso Isernia for the fruitful discussions, which
inspired the writing of this paper.
References
1. Colton, D.; Kress, R. Inverse Acoustic and Electromagnetic Scattering Theory, 2nd ed.; Springer: Berlin, Germany,
1998; ISBN 9781461449423.
2. Pastorino, M. Microwave Imaging; John Wiley: New York, NY, USA, 2010.
3. Conceicção, R.C.; Mohr, J.J.; O’Halloran, M. (Eds.) An Introduction to Microwave Imaging for Breast Cancer
Detection, 1st ed.; Biological and Medical Physics, Biomedical Engineering; Springer International Publishing:
Basel, Switzerland, 2016.
4. Crocco, L.; Conceicção, R.C.; James, M.L.; Karanasiou, I. (Eds.) Emerging Electromagnetic Technologies for Brain
Diseases Diagnostics, Monitoring and Therapy; Springer: Cham, Switzerland, 2018.
5. Bevacqua, M.T.; Bellizzi, G.G.; Crocco, L.; Isernia, T. A Method for Quantitative Imaging of Electrical
Properties of Human Tissues from Only Amplitude Electromagnetic Data. Inverse Probl. 2019, 35, 025006.
[CrossRef]
6. Turk, A.S.; Hocaoglu, A.K.; Vertiy, A.A. (Eds.) Subsurface Sensing; Wiley: Hoboken, NJ, USA, 2011.
7. Persico, R. Introduction to Ground Penetrating Radar: Inverse Scattering and Data Processing; Wiley: Hoboken,
NJ, USA, 2014.
87
J. Imaging 2019, 5, 47
8. Bertero, M.; Boccacci, P. Introduction to Inverse Problems in Imaging; Institute of Physics: Bristol, UK, 1998.
9. Cakoni, F.; Colton, D. Qualitative Methods in Inverse Scattering Theory; Springer: Berlin, Germany, 2006.
10. Ammari, H.; Iakovleva, E.; Lesselier, D.; Perrusson, G. MUSIC-type electromagnetic imaging of a collection
of small three-dimensional inclusions. Siam J. Sci. Comput. 2007, 29, 674–709. [CrossRef]
11. Tortel, H.; Micolau, G.; Saillard, M. Decomposition of the time reversal operator for electromagnetic scattering.
J. Electromagn. Waves Appl. 1999, 13, 687–719. [CrossRef]
12. Colton, D.; Haddar, H.; Piana, M. The linear sampling method in inverse electromagnetic scattering theory.
Inverse Probl. 2003, 19, 105–137. [CrossRef]
13. Kirsch, A.; Grinberg, N.I. The Factorization Method for Inverse Problems; Oxford University Press: Oxford, UK, 2008.
14. Potthast, R. A study on orthogonality sampling. Inverse Probl. 2010, 26, 074015. [CrossRef]
15. Bevacqua, M.; Isernia, T. Shape reconstruction via equivalence principles, constrained inverse source
problems and sparsity promotion. Prog. Electromagn. Res. 2017, 158, 37–48. [CrossRef]
16. Bevacqua, M.T.; Isernia, T. Boundary Indicator for Aspect Limited Sensing of Hidden Dielectric Objects.
IEEE Geosci. Remote Sens. Lett. 2018, 15, 838–842. [CrossRef]
17. Donoho, D. Compressed sensing. IEEE Trans. Inf. Theory 2006, 52, 1289–1306. [CrossRef]
18. Franceschetti, G. Electromagnetics: Theory, Techniques, and Engineering Paradigms; Springer Science & Business
Media: Berlin, Germany, 2013.
19. Belkebir, K.; Saillard, M. Special section: Testing inversion algorithms against experimental data. Inverse Probl.
2001, 17, 1565–2028. [CrossRef]
20. Catapano, I.; Crocco, L.; Isernia, T. On simple methods for shape reconstruction of unknown scatterers.
21. Kowalski, M. Sparse regression using mixed norms. Appl. Comput. Harmon. Anal. 2009, 27, 303–324.
[CrossRef]
22. Fornasier, M.; Rauhut, H. Recovery algorithms for vector-valued data with joint sparsity constraints. SIAM J.
Numer. Anal. 2008, 46, 577–613. [CrossRef]
23. Di Donato, L.; Bevacqua, M.; Isernia, T.; Catapano, I.; Crocco, L. Improved quantitative microwave
tomography by exploiting the physical meaning of the Linear Sampling Method. In Proceedings of the 5th
European Conference on Antennas and Propagation, Rome, Italy, 11–15 April 2011; pp. 3828–3831.
24. Bevacqua, M.T.; Palmeri, R.; Isernia, T.; Crocco, L. Physical Interpretation of the Orthogonality Sampling
Method. In Proceedings of the 2nd URSI Atlantic Radio Science Meeting (AT-RASC), Gran Canaria, Spain,
28 May–1 June 2018.
25. Di Donato, L.; Bevacqua, M.T.; Crocco, L.; Isernia, T. Inverse Scattering Via Virtual Experiments and Contrast
Source Regularization. IEEE Trans. Antennas Propag. 2015, 63, 1669–1677. [CrossRef]
26. Bucci, O.M.; Isernia, T. Electromagnetic inverse scattering: Retrievable information and measurement
strategies. Radio Sci. 1997, 32, 2123–2138. [CrossRef]
27. Canny, J. A computational approach to edge detection. IEEE Trans Pattern Anal. Mach. Intell. 1986, 8, 679–698.
[CrossRef] [PubMed]
28. Catapano, I.; Crocco, L.; D’Urso, M.; Isernia, T. On the Effect of Support Estimation and of a New Model in
2-D Inverse Scattering Problems. IEEE Trans. Antennas Propag. 2007, 55, 1895–1899. [CrossRef]
29. Devaney, A.J. Mathematical Foundations of Imaging, Tomography and Wavefield Inversion; Cambridge University
Press: Cambridge, UK, 2012.
30. Crocco, L.; Di Donato, L.; Catapano, I.; Isernia, T. An Improved Simple Method for Imaging the Shape of
Complex Targets. IEEE Trans. Antennas Propag. 2013, 61, 843–851. [CrossRef]
31. Agarwal, K.; Chen, X.; Zhong, Y. A multipole-expansion based linear sampling method for solving inverse
scattering problems. Opt. Express 2010, 18, 6366–6381. [CrossRef] [PubMed]
© 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
88
Journal of
Imaging
Article
A Finite-Difference Approach for Plasma Microwave
Imaging Profilometry
Loreto Di Donato 1,2, *, David Mascali 2 , Andrea F. Morabito 3 and Gino Sorbello 1,2
1 Department of Electrical, Electronics and Computer Engineering (DIEEI), University of Catania,
viale A. Doria 6, 95126 Catania, Italy
2 Laboratori Nazionali del Sud, National Institute of Nuclear Physics (INFN), Via Santa Sofia 62,
95125 Catania, Italy
3 Department of Information Engineering, Infrastructures and Sustainable Energy (DIIES), University
“Mediterranea” of Reggio Calabria, via Graziella, Loc. Feo di Vito, 89124 Reggio di Calabria, Italy
Received: 9 July 2019; Accepted: 9 August 2019; Published: 12 August 2019
Abstract: Plasma diagnostics is a topic of great interest in the physics and engineering community
because the monitoring of plasma parameters plays a fundamental role in the development and
optimization of plasma reactors. Towards this aim, microwave diagnostics, such as reflectometric,
interferometric, and polarimetric techniques, can represent effective means. Besides the above,
microwave imaging profilometry (MIP) may allow the obtaining of tomographic, i.e., volumetric,
information of plasma that could overcome some intrinsic limitations of the standard non-invasive
diagnostic approaches. However, pursuing MIP is not an easy task due to plasma’s electromagnetic
features, which strongly depend on the working frequency, angle of incidence, polarization, etc., as
well as on the need for making diagnostics in both large (meter-sized) and small (centimeter-sized)
reactors. Furthermore, these latter represent extremely harsh environments, wherein different systems
and equipment need to coexist to guarantee their functionality. Specifically, MIP entails solution of an
inverse scattering problem, which is non-linear and ill-posed, and, in addition, in the one-dimensional
case, is also severely limited in terms of achievable reconstruction accuracy and resolution. In this
contribution, we address microwave inverse profiling of plasma assuming a high-frequency probing
regime when magnetically confined plasma can be approximated as both an isotropic and weak
penetrable medium. To this aim, we adopt a finite-difference frequency-domain (FDFD) formulation
which allows dealing with non-homogeneous backgrounds introduced by unavoidable presence of
plasma reactors.
Keywords: microwave plasma diagnostics; electromagnetic inverse scattering; microwave imaging

profilometry; finite-difference methods
1. Introduction
Plasma diagnostics using active microwave techniques is one of the most promising tools to
optimize the production of plasma in both large and compact reactors because of its non-invasive nature
and relatively modest access requirements [1]. Among the others, microwave imaging reflectometry
(MIR), interferometric and polarimetric techniques are currently under investigation to extract plasma
proprieties useful to model and optimize the heating process. Such information is crucial to study
and improve plasma generation based on Electron Cyclotron Resonance [2] as well as innovative
heating schemes such as the ones employing Electrostatic Bernstein Waves in overdense plasma [3].
Unfortunately, the diagnostic approaches mentioned above show some limitations. As an example,
MIR, a multifrequency radar-like technique, mainly used in large reactors, such as tokamak and
stellarator, is commonly used to infer electronic density distribution and fluctuation, but suffers from

J. Imaging 2019, 5, 70
the impossibility of performing diagnostics of non-increasing plasma profiles since inner reflecting
cutoff layers are “hidden” from the outer ones [4]. On the other hand, line-integrated measurements,
such as those actually provided by interferometric and polarimetric techniques, although employable
on compact reactors such as the electron cyclotron ion source (ECRIS) [5], cannot give local information
about electron density, but only a line-of-sight integrated value.
For these reasons, it is worth devoting interest towards microwave imaging tomographic
approaches, such as those actually used in medical and subsurface microwave diagnostics [6], which
are in principle able to “locally” monitor plasma uniformity and instabilities.
In a recent paper, we investigated the possibility of obtaining plasma imaging profilometry
(MIP) by means of electromagnetic inverse scattering techniques requiring only measurements of the
reflection coefficient in a quite large frequency band. In particular, adopting a high-frequency probing
regime (i.e., probing frequencies much larger than the cyclotron frequency), i.e., when the plasma can
be assumed to be both isotropic and weak scatterer, linear imaging recovery techniques are able to
achieve quantitative characterization of plasma profiles [7]. The results have shown that if the recovery
process is properly formulated as a sparse optimization problem, compressive sensing (CS)-based
strategies are able to achieve nearly optimal results. However, free-space homogeneous background
has been considered, which is an assumption suitable only if multiple reflections introduced by the
metallic environment (reactor walls) are neglected or, in any case, filtered out.
With the aim to get a further step in the analysis of such a problem, in this paper we introduce
a frequency-domain finite-difference (FDFD) formulation for the solution of the inverse scattering
problem in order to take easily into account the effect of non-homogeneous backgrounds related to the
metallic environment of the plasma chamber.
On the other hand, besides the modeling aspects, solution of the inverse scattering problem
in presence of a metallic surface, or, in general, non-homogeneous background, can bring some
advantages due to the possibility to achieve enhanced resolution in the reconstruction process [8,9].
Roughly speaking, this enhancement can be explained due to the multiple reflected waves which can
bring additional information conveyed back by the recovery process. Interestingly, this may be of
fundamental importance in the solution of the one-dimensional inverse scattering problem (also known
as inverse profiling) wherein the number and kind of retrievable information is severely limited [10].
The paper is structured as follows. In Section 2, the adopted mathematical formulation
for the solution of the forward and inverse scattering problem is described. In Section 3 the
microwave imaging profilometry for inhomogeneous plasma slabs recovery is formulated. Finally, in
Section 4, the developed approach is validated against simulated data dealing with homogeneous and
non-homogeneous backgrounds as well as reflection-only and reflection-transmission measurement
configuration. Conclusions follow.
2. Mathematical Formulation
Let us assume a bounded, simply connected, domain D located in an non-homogeneous
background medium, wherein some electromagnetic primary sources j0 are located in a region
characterized by relative permittivity and permeability distribution μr (r ) and r (r ), r spanning
the space D. After normalizing the electric field, magnetic field and the primary sources according to
4 1
ẽ = μ00 e, h̃ = h and j̃ = c0 j0 , with c0 = √ the speed of light, Maxwell equations read:
0 0 μ 0
c0 ∇ × ẽ = −jωμr (r )h̃

(1)
c0 ∇ × h̃ = jωr (r )ẽ + j̃
0
where ω is the angular frequency and the time factor exp{jωt} is assumed and dropped for the
time-harmonic fields. Normalized electric and magnetic fields, ẽ and h̃, are of the same order of
magnitude and this is an advantage in formulating the perfectly matched layer (PML) [11]. If, for the
90
J. Imaging 2019, 5, 70
sake of a simpler notation, we drop the tilde ˜ symbol and the c0 is included in the ∇× definition, we
can write:
5 × e = −jωμr (r )h
∇
(2)
5 × h = jωr (r )e + j0
∇
In the above formulation, to satisfy different boundary conditions, a lossy PML [12] can be
introduced to truncate the computational domain, as commonly done in numerical solver. By choosing
proper PML parameters, various boundary conditions can seamlessly introducing ranging from a
perfect absorbing boundary condition to a PEC (or PMC) boundary condition, the latter case when no
PML is used [11].
The background field (i.e., the field without the unknown profile) defined as a plane wave
propagating in a non-homogeneous, eventually lossy, background with permittivity and permeability,
respectively, given by:
σ (r )
b (r ) = b (r ) − jb (r ) = b (r ) − j
0 ω
μb (r ) = μb (r ) − jμb (r )
satisfies the following equations:
5 × ebck = −jωμb (r )hbck

∇
(3)
5 × hbck = jωb (r )ebck + j0
∇
To formulate the scattering problem, we can add and subtract the quantities jωb (r )e and
jωμb (r )h at the right hand side member of Equation (2) to obtain:
5 × e = −jωμb (r )h − jω [μr (r ) − μb (r )] h

∇
(4)
5 × h = jωb e + jω [r (r ) − b (r )] e + j0
∇
and, finally, subtracting (3) to (4), the equations governing the scattered fields, defined as esct = e − ebck
and hsct = h − hbck , are given by:

5 × esct = −jωμb (r )hsct − jωμb (r ) μr (r )−μb (r )
∇ μb (r )
h
(5)
5 × hsct = jωb (r )esct + jωb (r ) r (r )−b (r )
∇ b ( r )
e
wherein the quantities under the square brackets can be easily recognized as the contrast functions:
μr (ω, r ) − μb (ω, r )
χμ (ω, r ) = ,
μb (ω, r )
r (ω, r ) − b (ω, r )
χ (ω, r ) = ,
b (ω, r )
which relate the dielectric and magnetic properties of the anomalies, i.e., scattering objects, to those of
the embedding background medium at each frequency.
Adopting a dual grid [13,14] and a proper discretization, in operator notation, the above equations
can be rewritten as:

C jω Mb esct 0 −jωμb χμ e
= , (6)
−jω Eb C hsct jωb χ 0 h
91
J. Imaging 2019, 5, 70
wherein () is used to indicate the field integrated over the appropriate line element of the primal or
6 Moreover, the matrix operator C known in the literature as incidence
dual mesh (e ≈ eΔL, h ≈ hΔL).
matrix (see for example section §7.3.3 and Equation (7.1) of Tonti’s book [13] or [15]) denotes the
proper derivative operations and is a sparse matrix. Moreover, Mb and Eb are matrices accounting for
ΔS 6
the magnetic and dielectric non-homogeneous background properties, respectively, wherein Eb ≈ b ΔL
6 and ΔS
and Mb ≈ μb ΔS , with ΔL and ΔS refer to lines and surfaces of the primal mesh, and ΔL 6 refer
6
ΔL
to the dual mesh [13]. In the remaining of the paper for the sake of a simpler notation we drop the
symbol.
The formal solution of the linear system (6) can be rewritten as:

esct 0 −jωμb χμ e
= Hb−1 . (7)
hsct jωb χ 0 h
The above equation is the key equation for the inverse scattering problem adopting a
finite-difference formulation. As a matter of fact, the measured scattered fields, i.e., the scattered field
recorded on the data domain S (for the case at hand, the same positions z = zS at different frequencies),
can be represented by the following operator notations:
( *
esct (zS , ωn ) 0 −jωμb χμ e
= MS Hb−1 , (8)
hsct (zS , ωn ) jωb χ 0 h
where MS is an operator that extracts the field values at the observation point z = zS at different
frequencies. Substituting [esct , hsct ] T = [e , h] T − [ebck , hbck ] T into (7) , we obtain the domain
(or object) equation:
( *
ebck e 0 −jμb ωχμ e
= − Hb−1 , (9)
hbck h jωb χ 0 h
that can be sampled with an operator, say it M D , that extracts fields only in the imaging domain D at
the discretization grid points z D [16], i.e., :
( *
ebck (z D , ωn ) e ( z D , ωn ) 0 −jωμb χμ e
− = M D Hb−1 . (10)
hbck (z D , ωn ) h ( z D , ωn ) jωb χ 0 h
Equations (8)–(10) are the basic equations to solve the inverse scattering problem which is
non-linear since both the contrast functions χ and χμ and the field e and h are unknowns of the
problem [17]. Moreover, the problem is ill-posed, too, and effective regularization strategies must be
adopted for stable solutions against the presence of noise on data [18].
3. Microwave Imaging Profilometry
3.1. Linearized Approach for Plasma Slabs

Let us consider a one-dimensional non-homogeneous slab extending in the range [z1 , z2 ]
embedded in a non-homogeneous background, see Figure 1a. Normal incidence plane waves are used
to probe the slab at difference frequencies with phase reference equal to zero at z = z(View) (z ≥ z1 ).
Accordingly, the scattered field can be gathered at a given abscissa z1 ≤ zS = z(Meas) ≤ zmin (reflection
measurement) or zmax ≤ zS = z(Meas) ≤ z2 (transmission measurement).
92
J. Imaging 2019, 5, 70
(a) (b)
Figure 1. Sketch of the single view normal incidence measurement. The imaging domain extents in the
range [zmin , zmax ] and the scattered field data at different frequencies are gathered at a given observation
points external to this range. (a) Homogeneous background: configuration with only an active (tx-rx)
antenna (neglecting the rightmost dashed antenna) and an active (tx-rx) and a passive (rx) antenna
(considering the rightmost dashed antenna). (b) Non-homogeneous background: configuration with
an active antenna (tx-rx) with a PEC surface.
To retrieve nonmagnetic (μr (r ) = μb ) dielectric profile of inhomogeneous isotropic plasma

slabs, we consider a first order linearized approach under the Rytov approximation, instead of
solving the exact optimization problems stated by Equations (8)–(10). Linearized approaches, such
as the Born (BA) and Rytov (RA) approximations, can be used to recover small and extended (with
respect to the wavelength) weak slabs, respectively, i.e., slabs showing low contrast (||χ || << 1).
In particular, the Rytov approximation is useful for scatterers far extended in terms of the probing
wavelengths (anyway, RA reduces to BA when the scatterers become smaller and smaller with respect
to the probing wavelengths [19]). In this respect, to fulfill model hypothesis, it is worth noticing that
assuming a relatively high-frequency probing band (with respect to the cyclotron frequency) entails
several positive fallouts. First, at probing frequencies much higher than the cyclotron frequency, the
plasma can be assumed as an isotropic medium since the permittivity tensor becomes diagonal with
the terms all equals each other [2,20]. Second, at probing frequencies such that ω 2 >> ν2 the plasma
contrast becomes:
ω 2p ω 2p ν ω 2p ω 2p ν
χ = − 2 −j ≈ − 2 −j 3 , (11)
ω +ν 2 ω (ω + ν )
2 2 ω ω
wherein: 7
n e e2
ωp = , (12)
m e 0
ne , e and me being the electron density, electron charge, and electron mass, respectively. Under the
above approximations, considering the frequency difference (FD) formulation described in Section 2,
the unknown total electric field can be approximated by the incident field according to the RA, to
obtain the following linear equation:

fsct = MS R5 −1 χ ebck , (13)
b
wherein fsct = esct /ebck and R5 −1 , replacing H

5 −1 unless unessential constant, takes into account the
b b
normalization by the incident field values at the measurement point z(Meas) introduced by the Rytov
approximation [19] and the frequency dependence of the (multifrequency) scattering operator. As a
result, in order to solve for the plasma constitutive parameters ω 2p and ν, the frequency dependent
93
J. Imaging 2019, 5, 70
complex valued problem (13), can be recast in terms of frequency independent real valued problem as
already exploited in [7], i.e.,:
⎡ ⎤ ⎡ ⎤⎡ ⎤
Re(fsct ) 5 −1 ebck )
− ω12 MS Re(R 1 5 −1 ebck )
MS Im(R ω 2p
b ω3 b
⎢ ⎥ ⎢ ⎥⎢ ⎥
⎣ ⎦=⎣ ⎦⎣ ⎦. (14)
Im(fsct ) 5 −1 ebck )
− ω13 MS Im(R 5 −1 ebck )
− ω12 MS Re(R ω 2p ν
b b
Problem (14) can be solved as a whole or by splitting it into two subsequent subproblems
wherein the first step copes with the retrieval of the plasma frequency (i.e., the real part of the plasma
contrast) and the second one dealing with the retrieval of the collision rate which is related to the
imaginary part of the plasma contrast. This approach has been proposed in [7] wherein, considering
the very low effect of the plasma losses on scattered field, the overall problem can be “decoupled”.
This notwithstanding, in the following, we concentrate our attention only on the retrieval of the plasma
angular frequency. As a matter of fact, experimental studies state that the collision rate is almost
one order of magnitude smaller that the cyclotron frequency, which in turn, is much lower that the
probing frequency. This entails to cope with a lossless plasma and the whole entire problem (14) will
be reduced to the following one:
⎡ ⎤ ⎡ ⎤
Re(fsct ) −ωmax
2 ω 5 −1 ebck )
MS Re(R
ω2 b
⎢ ⎥ ⎢ ⎥ ω 2p
⎣ ⎦=⎣ ⎦ , (15)
ωmax
2
Im(fsct ) −ωmax
2 ω 5 −1 ebck )
MS Im(R
ω3 b
with ωmax
2 the maximum probing angular frequency used as normalization constant introduced for
numerical stability of the inversion process [7].
3.2. Sparsity-Promoting Recovery Approaches

Equation (15) is a linear but still ill-posed one which must be solved in a regularized fashion.
However, as stated in [10] and investigated in the case of plasma [7], minimum energy solutions
approaches, such as Tikonov regularization [21,22] are completely unable to recover neither the
shape nor the constitutive parameters of arbitrarily shaped plasma slabs. For this reason, CS-based
approaches are worth to be considered, since they are in principle suitable to overcome the intrinsic
limitations of the minimum energy solution related to the kind and number of actual parameters which
can be conveyed back from multifrequency single view inverse scattering problem.
Let us consider a generic linear problem of the following kind:
Φx = f , (16)
x is the N × 1 unknown vector and Φ is the M × N matrix, the so-called sensing matrix, and f the
M × 1 data vector of the multifrequency scattered field ordered according to (15). Let us suppose that
a convenient representation matrix Ψ exists such that the unknown can be expanded as x = Ψs with
only a small number of the coefficients s different from zero. According to CS theory, it is possible
to solve the inverse problem (16) even if M << N but is sufficiently larger than the number S of
coefficients different from zero (with S < M < N). However, it is worth underlining that the number
of the measurements M should be anyway in the order of the degrees of freedom of the scattered
field [7], which in the case of the slab can be calculated as [10]:

√ λM λm
I = 2Lslab b − , (17)
c0 c0
Lslab being the extent of the inhomogeneous slab, and λm and λ M the minimum and maximum
wavelength used to probe the scenario. Accordingly, it is possible to solve the inverse linear problem
94
J. Imaging 2019, 5, 70
ΦΨs = P s = f by means of different optimization constrained problems [23,24] based on the

minimization of the 1 -norm of the unknown when represented in a proper sparse basis. The commonly
adopted one in imaging problems, is the rectangular pulse basis function, commonly known as pixel
and voxel in 2D and 3D, respectively. However, adopting such a basis, sparse recovery approaches are
successful only if the scatterer resembles a point-like scatterer, whereas, in the case at hand, the plasma
embeds almost the whole chamber which, in turn, is several probing wavelengths extended.
On the other hand, if a step-wise constant representation of the unknown profile is adopted,
which is a reasonable assumption for the problem at hand, 1 optimization can be still exploited.
One possibility is offered CS-based approaches involving as objective function the first derivative of
the unknown [7]. For these reasons, we could consider the following recovery problems:
" "
" d(ω̃ 2 ) "
" p "
argmin " " subject to P ω̃ 2p − f 2 ≤ δ, ω̃ 2p ≥ 0, (18)
" dz "
1
ω2
wherein ω̃ 2p = ω2 p . In (18), · 1 and · 2 denote the 1 - and L2 -norms, respectively, and the
max
parameter δ, which depends on the measurement error (noise on data) and model error (field
approximation), is a positive defined parameter. In (18) the minimization of the 1 -norm promotes
the search of solutions with a sparse derivative among all the solutions consistent with the measured
data (within the given error threshold δ). In addition, taking into account the real positive value of the
unknowns, a further constraint (ω̃ 2p ≥ 0) can be enforced in (18) in order to reduce the search space
by convex optimization and improve the reconstruction results. The approach (18) is known as basis
pursuit denoising (BPDN) or least absolute shrinkage and selection operator (LASSO) problem [24].
In [7], the sparsity-promoting approaches (18) have been proved to be capable in retrieving
synthetic step-wise constant axial plasma profile which can be represented with few coefficients in
terms of their first derivative. However, as the 1D single view multifrequency inverse scattering
problem is severe limited, by its very nature, in terms of number of degrees of freedom (I), the number
of non-null coefficient of any suitable representation basis should stay very few in order to achieve
reliable reconstructions. This means that when smoother ad smoother profiles are considered, the
adopted sparsity-promoting approach (18) is expected to be unsatisfactory.
For this reason, a “relaxed” version of (18) can be considered, i.e.,:
" "
" d(ω̃ 2 ) "
" p "
argmin P ω̃ 2p − f 2 subject to " " ≤ γ, ω̃ 2p ≥ 0, (19)
" dz "
1
wherein the role of the objective function with that of the constraints are exchanged. Approach (19)
is expected to work better than (18) with profiles which are not exactly step-wise constant. Also, in
this case, the assessment of the parameter γ is not a simple task. In this respect, prior information
can be exploited to assess the optimal choice of the regularization parameter that mainly depends
on the sparsity of the unknown at hand. However, the optimal choice of this threshold cannot be a
priori established, pointing out one could a priori establish the range of the threshold and compare the
results obtained for different values of γ.
4. Numerical Assessment Towards Benchmark Examples

The two benchmarks we refer to as “flat-top” and “hollow-core”, respectively, represent expected
electron density axial distribution in ECRIS-like devices [5]. The synthetic data have been generated
considering the convolution between a rectangular window and a Hann window large 10 cm and
0.7 cm, respectively with a discretization grid of 256 cells. As a result, smoother profiles than those
tackled in [22] can be considered. The total and the background fields have been simulated in the
frequency range [17.5–34.5] GHz considering 87 evenly spaced frequency values in the solution of the
forward problem stated by (3), thus the scattered field data has been obtained as difference between
95
J. Imaging 2019, 5, 70
them. The solution of the forward problem can be performed by means of a standard conjugate
gradient scheme as that proposed in [16]. The choice of the adopted frequency band is related to
several requirements. The first one is the need to consider magnetically confined plasma as isotropic
and weak scattering medium, and such assumption is fulfilled when the probing frequency is much
higher than the cyclotron frequency, hereafter considered at 2.45 GHz. The second one, is related to the
need for working with a large bandwidth in order to achieve the number of degrees of freedom I as
large as possible, see Equation (17). Last, but not least, at this frequency band, also in centimeter-sized
compact reactors, the propagation can be reasonably approximated as one-dimensional one along the
chamber’s axis. According to the above reasonings, we have set the frequency band of the fundamental
mode in a rectangular waveguide WR34 that can be used to feed a horn antennas opening in the front
and back walls of the plasma reactor [3].
We have considered three kind of measurement/background configurations which resemble
possible experimental setup:
1. A single transmitting and receiving antenna measuring the reflection coefficient in a free-space
homogeneous background (reflection-only measurement);
2. Single transmitting and receiving antenna measuring the reflection coefficient in presence of a
PEC surface;
3. A transmitting and receiving antenna measuring the reflection coefficient and a receiving antenna
measuring the transmission coefficient in a homogeneous free-space background (reflection and
transmission measurement).
The three cases above can be seen a possible operational scenario where the reflection from
metallic environments introduced by the plasma reactor can be neglected or not. This is the case,
for example, when large and small reactors are in order, respectively.
To simulate the above measurement/background configurations, an absorbing, or free-space
condition, can be very easily introduced, in the case at hand, by considering losses in PML layers
(z1 − d ≤ z ≤ z1 and z2 ≤ z ≤ z2 + d) where the medium is gradually modified introducing losses
according appropriate permittivity and permeability r [1 − jσ(z)], μr [1 − jσ (z)] to have ideally no
reflections from the PML [11,12,14].
For the sake of simplicity, we
have
chosen a simple polynomial grading for the PML layers
m m
conductivity profiles: σ (z) = σmax z1d−z for z1 − d ≤ z ≤ z1 and σ(z) = σmax z−dz2 for z2 ≤ z ≤
z2 + d; but more complex profiles can be used [25]. On the other hand, when PEC layer is considered,
the PML is completely removed and perfect reflecting surface boundary conditions are enforced at
z = z2 .
It is worth pointing out that for the solution of the forward problem, the plasma profiles are
assumed to be lossy with the collision rate profile one order of magnitude smaller than the plasma
angular frequency. As a result, this introduces a model error in the scattered field data which avoids
the so-called “inverse crime”. Furthermore, the scattered field data have been corrupted by random
gaussian white noise of 5%. Finally, the number of the cells in the computational grid are set to
N = 256.
Problem (19) has been solved by means of the environment CVXPY 1.0 for the solution of convex
optimization problems [26]. In doing this, the choice of the threshold value γ has been set equal to
the exact 1 -norm of the actual unknown. Although this exact value cannot be chosen in any actual
experiment, it allows an understanding of how the value of γ can be set on the base of expected a
priori information on the degree of sparsity of the profile. Furthermore, for our investigation, such a
choice allows the making of a fair comparison among the achieved results in the three different cases.
For the two benchmarks profiles at hand it is γ = 6.22 (flat-top) and γ = 7.14 (hollow-core)
Finally, the required time to solve the inverse scattering problems takes very negligible time
(below few seconds on a standard PC) due to the one-dimensional nature of the problem and then the
size of the relevant matrix operator.
96
J. Imaging 2019, 5, 70
4.1. Reflection-Only Measurement with Single Antenna

We have considered free-space background with PML layers with thickness of 2 cm, σ = 6, m = 4,
for a sketch the reader can refer to Figure 1a without considering the rightmost antenna. On the other
hand, we have considered a non-homogeneous background with a PEC boundary condition on the
right hand side of the computational domain (z2 = 25 cm) to simulate the plasma chamber with only
one antenna, see Figure 1b.
The imaging domain, [zmin , zmax ], is long 20 cm with the source point z(View) and the measurement
point z(Meas) placed at 2.5 cm far from the imaging domain both in the case of single antenna and
two antennas.
The imaging results are reported in Figures 2 and 3 for both the homogeneous and
non-homogeneous background, respectively. As it can be see, in the case of PEC layer, the
xtrue − xrec 22
reconstructions brings better result with a reconstruction error (defined as err = ) of 9.39%
xtrue 22
for the flat-top and 13.03% for the hollow core, with the plasma profile not heavily underestimated as
in the case of the homogeneous background, wherein the reconstruction errors achieve 47.85% for the
flat-top and 27.02% for the hollow core. This means that the presence of a reflecting surface allows
to the achievement of better results in terms of accuracy of the reconstructed profiles. On the other
hand, it is worth underlining that while the reconstruction of the flat-top profile is almost satisfactory,
Figure 3a, in the case of the hollow-core profile, neither the presence of the PEC allows imaging of
the main feature, although differences about the two different recovered profiles are evident, see
Figure 3a,b. This may be due to a lower degree of sparsity of the hollow-core profile than that of
the flat-top one and to the requirement by (19) to find solutions (profiles) with bounded 1 norm
compatible with the objective function.
(a) (b)
Figure 2. Reconstruction of the flat-top (a) and hollow core (b) plasma profiles using one tx-rx antenna
in homogeneous vacuum background by means of approach (19). Blue continuous line: electronic
density profile, red dashed line: retrieved electronic density profile.
97
J. Imaging 2019, 5, 70
(a) (b)
Figure 3. Reconstruction of the flat-top (a) and hollow core (b) plasma profiles using single tx-rx antenna in
non-homogeneous background with a PEC reflecting surface by means of approach (19). Blue continuous
line: electronic density profile, red dashed line: retrieved electronic density profile.
4.2. Reflection-Transmission Measurement with Two Antennas

In this case, two antennas are considered, see Figure 1a, with one (active) antenna radiating the
probing field and measuring the reflection coefficient, and a second (passive) antenna measuring
only the transmitting coefficient. Opposite to the case with a single antenna, in this case the
non-homogeneous background accounting for two PEC surfaces has not been considered. First,
from a modeling point of view, the use of PEC surfaces entails an electromagnetic cavity which
enforces solutions only at the resonance frequencies given by f n = 2Lnc 0
cavity
, i.e., when the cavity length
(Lcavity ) is an integer multiple of the probing half-wavelength, for a given finite set of integer n∗
indexes depending on the adopted frequency band. Secondly, when the plasma partially embeds the
cavity, the resonance frequencies shift in frequency, so that the expected large error on the background
field approximation prevent using a linear model such as the Rytov approximation. Finally, from an
applicative point of view, since in general, the front and back walls of the plasma reactors host two
aperture antennas [3], just a (possible small) fraction of the transmitted microwave power will be
reflected, while the most part will be gathered by the antennas. For these reasons, we have considered
only homogeneous background.
The imaging results are reported in Figure 4 and the reconstruction errors achieve err = 1.01%
and err = 3.56% for the flat-top and hollow core, respectively. Accordingly, it can be shown as the
presence of a second passive antenna achieve better results in terms of accuracy of the retrieved profiles
when compared to the single antenna in free-space homogeneous background. On the other hand,
while the reconstruction of the flat-top profile is nearly optimal, also in this case, the reconstruction
of the hollow-core profile is not satisfactory to apprise the main feature of the profile, although some
small differences between the two retrieved profile are evident (as in the case of single antenna and
PEC surface) and the same possible explanations hold true.
98
J. Imaging 2019, 5, 70
(a) (b)
Figure 4. Reconstruction of the flat-top (a) and hollow core (b) plasma profiles using two antennas
in homogeneous free-space background by means of approach 19. Blue continuous line: electronic
density profile; red dashed line: retrieved electronic density profile.
5. Discussion of the Results through Singular Value Decomposition Analysis

To understand the different results achieved under the two different considered measurement
configurations and related backgrounds, we have considered the SVD analysis of the relevant scattering
operator P for the different analyzed cases.
First of all, it is worth underlining that even if we are not considering minimum energy solution
approaches (L2 -norm), the SVD analysis can be retained still valid to understand the main features
among the different measurement configurations and backgrounds, since we are considering sparse
promoting approaches rather than exact compressive sensing-based approaches. Indeed, (19) entails
a standard minimum energy solution objective function equipped with a regularization (penalty)
1 -norm enforcing step-wise sparsity. For this reason, we numerically perform the SVD of the relevant
scattering operator with the aim to give a qualitative interpretation of the different results shown in
the previous Section.
Let us denote with {un , σn , vn } the singular value decomposition of the matrix operator P
mapping the data-to-unknown relationship in (19), wherein un represents the left singular vectors
(spanning the object space), vn the right singular vectors (spanning the data space) and σn the singular
values, in such way P un = σn vn and P + vn = σn un , wherein + stand for the adjoint of the operator [18].
For all the measurement configurations adopted in the numerical analysis, we have analyzed the
following metrics based on the SVD analysis, i.e.,:
1. behavior of the singular values —the logarithmic plot of the singular values as ordered in
non-increasing fashion. Indeed, as the scattering operator is a compact one [18], its singular
values exhibits an exponential decay after a given threshold index I (analytically expressed
by (17)), which indicates the maximum number of the degrees of freedom, and hence of the
parameters which can be conveyed back by the recovery procedure;
2. spectral coverage (SC) defined as:
NT
SC = ∑ |Ũn |2 (20)
n =0
wherein Ũ is the Fourier Transform of the left singular vectors and NT the truncation index used
as regularization parameter in TSVD approach [10]. It is a measure of the class of profiles which
can be actually retrieved in the object domain by the inversion process [10].
99
J. Imaging 2019, 5, 70
3. point spread function (PSF) defined as [27]:
NT
PSF (z − z0 ) = ∑ u n ( z ) u n ( z0 ) ∗ (21)
n =0
wherein ()∗ stands for conjugation, and z0 the abscissa with respect the PSF is considered. For the
case at hand, we set z0 = 12.5 cm, which is the center of the imaging domain. The point spread
function is a direct measure of the ultimate attainable spatial resolution [27].
The SVD analysis is reported in Figure 5. It can be seen that the singular values are different
for the considered configurations. Indeed, for the PEC case the lower order singular values show a
greater magnitude while vanishing to zero (like in the case of the free-space background configuration).
On the other hand, in the case of two antennas in homogeneous background, the magnitude of the
singular values keeps almost always greater than the single antenna (without and with PEC).
(a) (b)
(c)
Figure 5. Singular value decomposition analysis. (a) Singular values, (b) spectral coverage and (c) point
spread function in the case of single antenna free-space background (solid blue line), single antenna
inhomogeneous PEC surface (red dashed line) and two antennas free-space background (dot black
line). The cutoff value is NT = 40 according to the change of slope in the singular values behavior,
while the spatial frequency in SC are normalized to the extent of the imaging domain.
The SVD analysis give us the fundamental answer to understand the differences in the
reconstruction process for the three adopted measurement configurations. Indeed, two main comments
are in order. First, the presence of a perfect reflecting plane (PEC) introduces an increase in the singular
values magnitude as compared to the homogeneous case. This is actually shown by the blue and red
dashed line in Figure 5a. Interestingly, such an increase entails the advantage to deal with a more
100
J. Imaging 2019, 5, 70
stable inverse scattering problem and with an increase of the information content [9]. Second, when we
move to the case of two antennas, it can be easily understood as this configuration is almost equivalent
to the case of single antenna operating in presence of a PEC surface. Indeed, this can be explained
considering that the PEC surface acts as a “secondary” antenna which radiates (reflects) the field in
such a way a virtual transmission coefficient can be measured by the actual antenna, as superimposed
to the reflection coefficients that it would be measured without the PEC surface. As a result, when
two antennas are considered, the reconstruction result achieves better results as in presence of a PEC
surface. This analysis shows that the use of a second passive antenna is almost equivalent to have a
reflecting surface behind the plasma slab. Anyhow, in all those cases wherein free-space approximation
can be assumed (f.i., large meter-sized machine), two antennas can be used to achieve an enhancement
in the reconstruction process. This has an immediate fallout in the design of the measurement setup
depending on the constraints enforced by the experimental facilities.
On the other hand, the SVD analysis shows also that SC and PSF attained in three measurement
configurations are identical, unless some very negligible differences due to numeric evaluation of
the singular value system. These are confirmed by the same reconstruction performance achieved by
the single antenna with PEC surface and two antennas in homogeneous background, for which the
harmonic content is almost exactly the same. As a result, the improvement in the reconstruction in
presence of a PEC surface can be attributed only to the difference in the singular values spectrum.
6. Conclusions
In this paper, we have addressed microwave imaging profilometry through the solution of an
inverse scattering problem with a finite-difference formulation which takes into account, under the
adopted assumptions, the effects of the metallic case represented by the plasma reactor. It has been
shown that the presence of a metallic surface profoundly benefits the inversion recovery process
when faced with sparsity-promoting approaches. In addition, the developed analysis shows the
equivalence among two possible measurement configurations which are a single transmitting-receiving
antenna in presence of a PEC surface and two antennas (one active and the second passive) able to
measure also the transmission coefficients. Although the imaging capabilities are enhanced in presence
of totally reflecting surface and with two antennas, microwave imaging profilometry still shows
severe limitations in terms of accuracy in the recovery of arbitrary-shaped electron density profiles,
i.e., not exactly step-wise constant profiles. Further activities are currently being devoted to exploit
more efficient unknown representations and possible a priori information in order to achieve better
reconstructions accuracy and to deal with a class of profiles wider than that considered in this paper.
Author Contributions: Conceptualization, L.D.D. and G.S.; methodology, L.D.D., D.M., A.F.M. and G.S.;
writing—original draft preparation, L.D.D, A.F.M., D.M. and G.S. ; writing—review and editing, L.D.D. and G.S..
Funding: This research has been partially supported by the project CHANCE and PANDORA.
Acknowledgments: The authors would to thank L. Celona and G. Torrisi at the Laboratori Nazionali del Sud
(Istituto Nazionale di Fisica Nucleare, Catania, Italy) for their fruitful suggestions and discussions.
References
1. Park, S.; Jang, J.; Choe, W. Sparse data recovery of tomographic diagnostics for ultra-large-area plasmas.
Plasma Sour. Sci. Technol. 2019, 28, 035012. [CrossRef]
2. Geller, R. Electron Cyclotron Resonance Ion Sources and ECR Plasmas; Routledge: London, UK, 2018.
3. Torrisi, G.; Sorbello, G.; Leonardi, O.; Mascali, D.; Celona, L.; Gammino, S. A new launching scheme for ECR
plasma based on two-waveguides-array. Microw. Opt. Technol. Lett. 2016, 58, 2629–2634. [CrossRef]
4. Mazzucato, E.; Munsat, T.; Park, H.; Deng, B.; Domier, C.; Luhmann, N., Jr.; Donné, A.; van de Pol, M.
Fluctuation measurements in tokamaks with microwave imaging reflectometry. Phys. Plasmas 2002,
9, 1955–1961. [CrossRef]
101
J. Imaging 2019, 5, 70
5. Mascali, D.; Torrisi, G.; Leonardi, O.; Sorbello, G.; Castro, G.; Celona, L.; Miracoli, R.; Agnello, R.; Gammino, S.
The first measurement of plasma density in an ECRIS-like device by means of a frequency-sweep microwave
interferometer. Rev. Sci. Instrum. 2016, 87, 095109. [CrossRef] [PubMed]
6. Pastorino, M. Microwave Imaging; Wiley Online Library: Hoboken, NJ, USA, 2010.
7. Di Donato, L.; Morabito, A.F.; Torrisi, G.; Isernia, T.; Sorbello, G. Electromagnetic Inverse Profiling for Plasma
Diagnostics via Sparse Recovery Approaches. IEEE Trans. Plasma Sci. 2019, 47, 1781–1787. [CrossRef]
8. Gilmore, C.; LoVetri, J. Enhancement of microwave tomography through the use of electrically conducting
enclosures. Inverse Probl. 2008, 24, 035008. [CrossRef]
9. Solimene, R.; Maisto, M.A.; Pierri, R. Inverse scattering in the presence of a reflecting plane. J. Opt. 2015,
18, 025603. [CrossRef]
10. Pierri, R.; Persico, R.; Bernini, R. Information content of the Born field scattered by an embedded slab:
multifrequency, multiview, and multifrequency—Multiview cases. JOSA A 1999, 16, 2392–2399. [CrossRef]
11. Sullivan, D.M. Electromagnetic Simulation Using the FDTD Method; John Wiley & Sons: Hoboken, NJ, USA , 2013.
12. Sacks, Z.S.; Kingsland, D.M.; Lee, R.; Lee, J.F. A perfectly matched anisotropic absorber for use as an
absorbing boundary condition. IEEE Trans. Antennas Propag. 1995, 43, 1460–1463. [CrossRef]
13. Tonti, E. The Mathematical Structure of Classical and Relativistic Physics; Springer: Berlin, Germany, 2013.
14. Laudani, R.; Calvagna, A.; Sorbello, G.; Janner, D.; Tramontana, E. Analysis of the Discretization Error at
Material Interfaces in Staggered Grids. IEEE Trans. Antennas Propag. 2010, 58, 1653–1661. [CrossRef]
15. Clemens, M.; Weiland, T. Discrete electromagnetics: Maxwell’s equations tailored to numerical simulations.
Int. Compumag Soc. Newsl. 2001, 8, 13–20.
16. Abubakar, A.; Hu, W.; Van Den Berg, P.; Habashy, T. A finite-difference contrast source inversion method.
17. Colton, D.; Kress, R. Inverse Acoustic and Electromagnetic Scattering Theory; Springer-Verlag: Berlin, Germany, 1992.
18. Bertero, M.; Boccacci, P. Introduction to Inverse Problems in Imaging; Institute of Physics: Bristol, UK, 1998.
19. Slaney, M.; Kak, A.C.; Larsen, L.E. Limitations of imaging with first-order diffraction tomography. IEEE Trans.
Microw. Theory Tech. 1984, 32, 860 – 874. [CrossRef]
20. Stix, T. Waves in Plasmas; American Institute of Physics: College Park, MD, USA, 1992.
21. Di Donato, L.; Palmeri, R.; Sorbello, G.; Isernia, T.; Crocco, L. Assessing the capabilities of a new
linear inversion method for quantitative microwave imaging. Int. J. Antennas Propag. 2015, 2015.
doi:10.1155/2015/403760. [CrossRef]
22. Di Donato, L.; Palmeri, R.; Sorbello, G.; Isernia, T.; Crocco, L. A new linear distorted-wave inversion method
for microwave imaging via virtual experiments. IEEE Trans. Microw. Theory Tech. 2016, 64, 2478–2488.
[CrossRef]
23. Donoho, D.L. Compressed sensing. IEEE Trans. Inf. Theory 2006, 52, 1289–1306. [CrossRef]
24. Baraniuk, R.G. Compressive sensing [lecture notes]. IEEE Signal Process. Mag. 2007, 24, 118–121. [CrossRef]
25. Rumpf, R.C. Simple implementation of arbitrarily shaped total-field/scattered-field regions in finite-difference
frequency-domain. Prog. Electromagn. Res. 2012, 36, 221–248. [CrossRef]
26. CVXPY 1.0. 2019. Available online: https://www.cvxpy.org/ (accessed on 9 July 2019).
27. Pierri, R.; Brancaccio, A.; Leone, G.; Soldovieri, F. Electromagnetic prospection via homogeneous and
inhomogeneous plane waves: the case of an embedded slab. AEU-Int. J. Electron. Commun. 2002, 56, 11–18.
[CrossRef]
102
Journal of
Imaging
Article
Developments in Electrical-Property Tomography
Based on the Contrast-Source Inversion Method
Reijer Leijsen 1,† , Patrick Fuchs 2,† , Wyger Brink 1 , Andrew Webb 1 and Rob Remis 2, *
1 Department of Radiology, C.J. Gorter Center for High Field MRI, Leiden University Medical Center,
2333ZA Leiden, The Netherlands; [email protected] (R.L.); [email protected] (W.B.);
[email protected] (A.W.)
2 Circuits and Systems Group, Delft University of Technology, Faculty of Electrical Engineering,
Mathematics and Computes Science, 2628CD Delft, The Netherlands; [email protected]
* Correspondence: [email protected]; Tel.: +31-15-278-1442
Received: 4 December 2018; Accepted: 24 January 2019; Published: 1 February 2019
Abstract: The main objective of electrical-property tomography (EPT) is to retrieve dielectric tissue
parameters from B̂1+ data as measured by a magnetic-resonance (MR) scanner. This is a so-called
hybrid inverse problem in which data are defined inside the reconstruction domain of interest.
In this paper, we discuss recent and new developments in EPT based on the contrast-source
inversion (CSI) method. After a short review of the basics of this method, two- and three-dimensional
implementations of CSI–EPT are presented along with a very efficient variant of 2D CSI–EPT called
first-order induced current EPT (foIC-EPT). Practical implementation issues that arise when applying
the method to measured data are addressed as well, and the limitations of a two-dimensional
approach are extensively discussed. Tissue-parameter reconstructions of an anatomically correct
male head model illustrate the performance of two- and three-dimensional CSI–EPT. We show
that 2D implementation only produces reliable reconstructions under very special circumstances,
while accurate reconstructions can be obtained with 3D CSI–EPT.
Keywords: electromagnetic inverse scattering problems; magnetic resonance imaging; electrical-property

tomography; nonlinear optimization; contrast-source inversion
1. Introduction
The conductivity and permittivity values of different tissue types are of great importance in a
variety of medical applications. In magnetic-resonance (MR) safety [1] and hyperthermia-treatment
planning [2], for example, conductivity tissue profiles are required to determine the specific absorption
rate (SAR). Conductivity may also serve as a biomarker in oncology or in acute-stroke imaging [3].
Permittivity is important since it affects the spatial distribution of the transmitted electromagnetic field
responsible for spin excitation.
Typically, tissue conductivity and permittivity values are measured ex vivo for a particular range
of frequencies [4]. Other methods require elaborate hardware, such as electrical impedance tomography
(EIT) [5] or microwave-imaging methods [6]. The objective of electrical-property tomography (EPT)
is to retrieve these dielectric tissue values in vivo using an MR scanner and standard measurement
protocols [3,7]. Specifically, with an MR scanner, the so-called B̂1+ -field, defined as B̂1+ = ( B̂x + j B̂y )/2,
can be measured at a particular frequency of an operation called the Larmor frequency. This frequency
is proportional to the magnitude of static background field B0 via relation f = γ B0 , where γ =
42.577 MHz T−1 is the proton gyromagnetic ratio divided by 2π, leading to MR operating frequencies
of 128 and 298 MHz for a 3 and 7 T scanner, respectively.

J. Imaging 2019, 5, 25
Reconstruction of dielectric tissue parameters is based on the measured B̂1+ -field, and what
sets EPT apart from other more common inversion and imaging problems is that the measured
B̂1+ -field has its support inside the reconstruction domain. The EPT reconstruction problem therefore
belongs to the class of so-called hybrid inverse problems [8], and several EPT techniques have been
proposed to reconstruct the conductivity and permittivity profiles based on these internal B̂1+ data.
Loosely speaking, these techniques can be divided into local differential-based approaches (see, e.g.,
References [9–12]) and global integral-based approaches (see e.g., References [13–18]). Combinations
of local and global methods have been developed as well [19,20].
In this paper, we focus on a global integral-based EPT reconstruction method called contrast
source inversion (CSI)–EPT, where a CSI approach [21–23] is taken to solve an EPT reconstruction
problem. In particular, in CSI–EPT, the reconstruction problem is formulated as an optimization
problem in which an objective function is iteratively minimized. This objective function consists of a
term that measures the mismatch between modeled and measured data (data mismatch), and a term
that measures the discrepancy in satisfying Maxwell’s equations within the reconstruction domain
using a global integral field representation (consistency mismatch). Including the second consistency
term in the objective function is crucial to the performance of CSI, as shown in Reference [24].
Minimization of the objective function is carried out by iteratively updating a contrast function,
which describes the dielectric constitution of the body part of interest, and a so-called contrast source,
which is the product of the contrast function and the electric-field strength. Updating takes place
by fixing one variable and updating the other. More precisely, the contrast function is first fixed,
the contrast source is updated, and subsequently the contrast source is fixed and the contrast function
is updated.
The CSI–EPT method was originally introduced in Reference [14], where it was shown that
CSI–EPT is able to reconstruct strongly inhomogeneous conductivity and permittivity profiles within
the center slice of an object placed in the center of a body coil in a 3 T MR scanner. The method was
initially implemented for E-polarized electromagnetic fields in two-dimensional (2D) configurations in
which the electrical field is parallel to the bore axis (z-axis) and the magnetic field is purely transverse,
because it is significantly less complex than full three-dimensional implementation. The use of a 2D
approach was justified since it was shown that the electromagnetic field in the midplane of a birdcage
coil essentially has an E-polarized field structure [25]. An efficient alternative to CSI–EPT, first-order
induced-current EPT or foIC-EPT, was presented in Reference [20] as well. This method exploits the
structure of the two-dimensional E-polarized field to efficiently reconstruct the tissue profiles in the
midplane of the transmit coil. The foIC-EPT method is significantly faster than CSI–EPT and produces
reconstructions in real time with essentially the same quality as 2D CSI–EPT.
The CSI–EPT method has recently been extended to three-dimensional (3D) configurations in
Reference [26]. With this 3D implementation of CSI–EPT, volumetric conductivity and permittivity
profiles are obtained, and it is no longer necessary to restrict the reconstruction domain to the midplane
of a transmit coil. Moreover, 3D CSI–EPT is based on vectorial 3D Maxwell equations and no
(E-polarized) field structure is assumed to be present as in the case of a 2D approach. Unfortunately,
computation times dramatically increase compared with 2D CSI–EPT and foIC-EPT and, depending
on the configuration, it may take 3D CSI–EPT hours or even days to converge even on dedicated
high-performance computers or servers. Apart from possible preconditioning techniques that may be
applied to accelerate the convergence of 3D CSI–EPT, 2D CSI–EPT or foIC-EPT may be preferable in
practice, since reconstruction times are significantly shorter compared with 3D approaches.
In this paper, we thoroughly investigate this issue and compare reconstructions obtained with
2D CSI–EPT, foIC-EPT, and 3D CSI–EPT. Reconstruction artifacts in the conductivity and permittivity
profiles, the modeled B̂1+ -field, and the internal electric field are carefully studied. Our analysis shows
that only under very special conditions is a 2D approach justified. Even if the electromagnetic field has
an E-polarized field structure in the midplane of the transmit coil, imposing a two-dimensional field
104
J. Imaging 2019, 5, 25
structure is generally too limiting an approximation unless the body part of interest and transmit coil
strictly satisfy the longitudinal invariance condition.
This paper is organized as follows. In Section 2, the 2D and 3D CSI–EPT method is briefly
reviewed, and the governing integral representations are presented. A variant of 2D CSI–EPT, foIC-EPT,
is also presented, and a detailed analysis of the performance of all three reconstruction methods is
presented in Section 3 using a realistic head model from Virtual Family [27]. A discussion with
conclusions can be found in Section 4. Finally, we note that the position vectors in 2D and 3D are
denoted by ρ and x, respectively, and we use an exp(+jωt) time convention.
2. Theory
As mentioned above, the CSI–EPT algorithm operates on two unknowns and is based on two
fundamental equations. Specifically, the unknowns in CSI–EPT are contrast function χ̂ andthe contrast
source ŵ, and the fundamental equations are the data equation and object or state equation.
The contrast function describes the dielectric contrast of the body with respect to free space and is
given by χ̂(x) = ε r (x) − 1 − jσ(x)/ωε 0 , where ε r (x) and σ(x) are, respectively, the unknown relative
permittivity and conductivity profiles of the body, ε 0 is the permittivity of free space, and ω is the
Larmor frequency of operation. The contrast function has bounded domain Dbody that is occupied
by the body as its support, that is, the contrast function vanishes for x ∈ / Dbody . Finally, we note that
the contrast function is dimensionless, and that its real part is determined by the permittivity profile,
while its imaginary part is determined by the body’s conductivity profile.
The contrast source in CSI–EPT is defined as ŵ = χ̂Ê, where Ê(x) is the electric-field strength.
Note that it is common to refer to ŵ as a contrast source even though it is expressed in volts per meter
and is actually scaled electric-field strength. The electric-field strength is obviously also unknown,
since the dielectric constitution of the body is unknown. Even though this field is not of primary
interest in EPT, CSI–EPT does provide electric-field reconstructions that may be used to reconstruct the
local time-averaged power density that is dissipated into heat [1].
To arrive at the two fundamental equations of CSI–EPT, we set up a scattering formalism
in which we make use of the linearity of Maxwell’s equations and exploit the fact that the body
occupies a bounded domain Dbody . In particular, we first determine the electromagnetic field that is
present inside an empty birdcage coil. In practice, this so-called background field is computed using
electromagnetic-simulation software and we denote it by {Êb , B̂b }. We note that the assumption is
made here that external currents are impressed and field-independent. Consequently, antenna loading
is not directly taken into account. The total electromagnetic field in presence of the body is denoted by
{Ê, B̂}, and using the linearity of Maxwell’s equations this field can be written as
{Ê, B̂} = {Êb , B̂b } + {Êsc , B̂sc }, (1)
where {Êsc , B̂sc } is the scattered electromagnetic field due to the presence of the body. For this field,
we have the integral representations

BJ EJ
B̂sc (x) = Ĝ (x, x ) · ŵ(x ) dV and Êsc (x) = Ĝ (x, x ) · ŵ(x ) dV, (2)
x ∈Dbody x ∈Dbody
EJ BJ
where Ĝ and Ĝ are essentially the electric current to the electric field and electric current to
magnetic field Green’s tensors of the background medium. Note that these are the Green’s tensors of a
homogeneous background medium, and the presence of the coil is not taken into account. Explicit
expressions for these tensors are given below.
105
J. Imaging 2019, 5, 25
Having these integral representations at our disposal, we can now present the basic CSI–EPT
equations. We start with the equation that relates the measured B̂1+ -field to the contrast source.
In particular, using the integral representation for the scattered magnetic field of Equation (2), we have
B̂xsc + j B̂ysc
1
B̂1+;sc (x) = = ∑ BJ
Ĝxk (x, x ) + jĜyk
BJ
(x, x ) ŵk (x ) dV, (3)
2 2 x ∈Dbody k= x,y,z
which can be written more compactly as
B̂1+;sc (x) = Gdata {ŵ}(x) for x ∈ Dbody , (4)
where linear data operator Gdata is implicitly defined in Equation (3). Equation (4) is known as the data
equation and relates unknown contrast source ŵ to the scattered B̂1+ -field. Note that this scattered field
is known, since B̂1+;sc (x) = B̂1+ (x) − B̂1+;b (x) and the total B̂1+ -field is known through measurements,
while background field B̂1+;b (x) is known through simulations. The real phase is generally not known
in practice, and tranceive-phase approximation is often used, which can lead to reconstruction artefacts
at higher frequencies [28].
The second basic CSI–EPT equation, called the object or state equation, is obtained from the
integral representation for the scattered electric field as given by the second equation of Equation (2).
Using the definition of the scattered electric field Êsc = Ê − Êb , this integral representation can be
written as
EJ
Ê(x) − Ĝ (x, x ) · ŵ(x ) dV = Êb (x) (5)
x ∈Dbody
and multiplying the above equation by contrast function χ̂, we arrive at

EJ
ŵ(x) − χ̂(x) Ĝ (x, x ) · ŵ(x ) dV = χ̂(x)Êb (x) for x ∈ Dbody , (6)
x ∈Dbody
which can be written more compactly as
ŵ(x) − χ̂(x)Gbody {ŵ} = χ̂(x)Êb (x) for x ∈ Dbody , (7)
where linear operator Gbody is implicitly defined in Equation (6).

To summarize, the two fundamental unknowns in CSI–EPT are contrast function χ̂ and contrast
source ŵ, and the basic CSI–EPT equations are the data Equation (4) and the object Equation (7).
Now, suppose we have available an approximation for the contrast function and contrast source.
We denote these approximants by χ̃ and w̃, respectively, and, in order to measure how well these
approximations satisfy the data and object equations, we introduce the data and object residuals as
r̂d (x) = B̂1+;sc (x) − Gdata {w̃}(x) for x ∈ Dbody , (8)
and
r̂o (x) = χ̃(x)Êb (x) − w̃(x) + χ̃(x)Gbody {w̃}(x) for x ∈ Dbody , (9)
respectively, and measure their magnitudes using L2 -norms

r̂d 2body = |r̂d (x)|2 dV and r̂o 2body = |r̂o (x)|2 dV. (10)
x∈Dbody x∈Dbody
In CSI–EPT, these norms are used to define objective function
r̂d 2body r̂o 2body

F (χ̃, w̃) = + (11)
B̂1+;sc 2body χ̃Êb 2body
106
J. Imaging 2019, 5, 25
and the goal is to find a contrast function and contrast source that minimizes this objective function.
We note that including the two-norm of the object residual in the objective function (second term on
the right-hand side of Equation (11)) is crucial to the success of CSI, since it has been shown that a
contrast-source inversion approach without this term produces unsatisfactory results in general [24].
In CSI–EPT, finding the desired contrast function is now realized by minimizing the objective
function in a “fix-one-minimize-for-the-other” approach. The iterative process continues until a
predefined maximum number of iterations or specified tolerance level of the objective function has
been reached. Specifically, the basic CSI–EPT algorithm is as shown in Listing 1.
Listing 1. Contrast-source inversion–electrical-property tomography (CSI–EPT).
• Given initial guesses χ̃[0] and w̃[0] for the contrast function and contrast source, respectively,
• For k = 1, 2, ...
1. Fix the contrast to χ̃[k−1] and update the contrast source according to the update formula
w̃[k] = w̃[k−1] + α[k] v[k] .
2. Compute the corresponding electric-field strength Ê[k] according to (cf. Equation (5))
Ẽ[k] (x) = Êb (x) + Gbody {w̃[k] }(x).
3. Knowing contrast source w̃[k] and corresponding electric-field strength Ẽ[k] , determine
contrast function χ̃[k] from constitutive relation w̃[k] = χ̃[k] Ẽ[k] by solving
least-squares problem χ̃Ẽ[k] − w̃[k] 2body for minimum norm contrast function χ̃.
4. Stop if objective function is smaller than user-specified tolerance level, or if maximum
number of iterations have been reached.
• End.
Polak–Ribière update directions are usually taken for update direction v[k] in Step 1 of
the algorithm, but Fletcher–Reeves or Hesteness–Stiefel update directions may be used as well.
To determine these update directions, the gradient of F (χ̃[k−1] , w̃) with respect to w̃ at w̃ = w̃[k−1] is
required. Explicit expressions for this gradient and corresponding step length α[k] can be found in
Reference [23], for example.
Note also that, with Equation (5), the object residual can be written as r̂o = χ̃Ẽ − w̃ and, in Steps 2
and 3, we find minimum-norm contrast function χ̃ for which r̂o 2body is minimized. This contrast
function is generally sensitive to small perturbations in w̃ at locations where the magnitude of the
electric field strength is “small.” To suppress this effect, we can alternatively update the contrast
function at every iteration according to update formula
χ̃[k] = χ̃[k−1] + β[k] u[k] , (12)
with u[k]
the Polak–Ribière update direction for the contrast function, and β[k]
its corresponding
update coefficient. Such an approach usually has a regularizing effect and typically leads to smoother
reconstructions.
2.1. Object and Data Operators in Three-Dimensional CSI–EPT

In three dimensions and with air as a background medium, the integral representations for the
scattered fields, as given by Equation (2), take on form
107
J. Imaging 2019, 5, 25
ω
B̂sc (x) = j ∇ × Âsc (x) and Êsc (x) = (k20 + ∇∇·)Âsc (x), (13)
c20
where c0 is the electromagnetic-wave speed in vacuum, k0 = ω/c0 the wave number in vacuum,
and Âsc is the vector potential given by

Âsc (x) = Ĝ (x − x )ŵ(x ) dV, (14)
x ∈Dbody
with Ĝ the three-dimensional Green’s function of the vacuum background domain given by
exp(−jk0 |x|)
Ĝ (x) = . (15)
4π |x|
Note that the nabla operators act on position vector x and not on integration variable x .
Three-dimensional object operator Gbody can easily be identified from the second part in Equation (13).
For data operator Gdata , however, we have to substitute the x and y components of the scattered
magnetic flux density in Equation (3) to obtain
ω # + sc $
B̂1+;sc = ∂ Âz − ∂z Â+;sc , (16)
c20
+;sc
where ∂+ = 12 (∂ x + j∂y ) and Â+;sc = 12 ( Âsc
x + j Ây ). From the above expression for the scattered
+
B̂1 -field, the 3D data operator Gdata can be identified. Note the particular structure of this operator:
scattered B̂1+ -field originates from a difference between the transverse variations of the longitudinal
vector potential (∂+ Âsc
z ) and the longitudinal variations of the transverse vector potential (∂z Â
+;sc ).
2.2. Object and Data Operators in Two-Dimensional CSI–EPT

In various papers (see Reference [25], for example) it has been reported that the radio-frequency
(RF) field in the midplane of a birdcage coil is essentially E-polarized, meaning that the electric-field
strength only has a longitudinal component (Ê = Êz iz ), while magnetic -flux density only has x and
y components (B̂ = B̂x ix + B̂y iy ). Additionally, in a two-dimensional configuration that is invariant
in the z direction, external electric current densities with longitudinal components only generate
E-polarized fields. Identifying the currents in the rungs of the birdcage coil with these z-directed
external current sources and denoting the slice through the object that coincides with the midplane of
the birdcage coil by Sbody , it makes sense to assume that within this midplane the RF field is essentially
two-dimensional and E-polarized with integral representations for the scattered fields given by
ω
B̂sc (ρ) = j ∇T × Âsc (ρ), and Êsc (ρ) = k20 Âsc (ρ), (17)
c20
where ρ is the position vector in the midplane of the birdcage coil, ∇T = ix ∂ x + iy ∂y is the transverse
nabla operator, and
Âsc (ρ) = Ĝ (ρ − ρ )ŵ(ρ ) dS (18)
ρ ∈Sbody
is the vector potential in two dimensions (and is thus expressed as a two-dimensional integral as
opposed to the three-dimensional integral in the 3D case) with
j (2)
Ĝ (ρ) = − H0 (k0 |ρ|) (19)
4
(2)
the Green’s function of the two-dimensional homogeneous background medium (air) and H0
is the Hankel function of the second kind and order zero. In this two-dimensional case, object
operator Gbody can easily be identified from the second part of Equation (17) and does not contain
108
J. Imaging 2019, 5, 25
a gradient-divergence operator as in the three-dimensional case. For 2D data operator Gdata , we

have to substitute the x and y components of the magnetic-flux density as given by the first part of
Equation (17) in the definition of the B̂1+ -field to obtain
ω + sc
B̂1+;sc = ∂ Âz . (20)
c20
From this expression, 2D data operator Gdata can now easily be identified. Comparing the
two-dimensional field representation of Equation (20) with its three-dimensional counterpart of
Equation (16), we observe that longitudinal spatial variations are absent in the two-dimensional case.
Moreover, the vector potentials in both expressions are different, since this quantity is computed using
Equation (18) in the two-dimensional case, while the three-dimensional vector potential is given by
Equation (14). Differences between two- and three-dimensional CSI–EPT reconstructions are discussed
further in Section 3.
2.3. Simplified Two-Dimensional CSI–EPT—foIC-EPT

In two dimensions, the CSI–EPT algorithm can be simplified by exploiting the particular structure
of E-polarized RF fields. To make this simplification explicit, we first introduce differentiation operator
∂− = 12 (∂ x − j∂y ) and note that operators ∂− and ∂+ essentially factor two-dimensional Laplacian
Δ = ∂2x + ∂2y as
Δ = 4∂− ∂+ = 4∂+ ∂− . (21)
Now, as a first step, we substitute the second part of Equation (17) in Equation (20) to obtain
1 + sc
B̂1+;sc (ρ) = ∂ Êz (ρ). (22)
ω
Subsequently, we use the definition of the scattered fields to write the above expression as
1 + 1
B̂1+ (ρ) = B̂1+;b (ρ) + ∂ Êz (ρ) − ∂+ Êzb (ρ) (23)
ω ω
and, since B̂1+;b (ρ) = 1 + b

ω ∂ Êz ( ρ ), this simplifies to
1 +
B̂1+ (ρ) = ∂ Êz (ρ). (24)
ω
If we now act with the ∂− operator on this equation, we obtain
1
∂− B̂1+ = Δ Êz (25)
4ω
and since Êz satisfies Δ Êz − jωμ0 Ĵzind = 0 with
Ĵzind = (σ + jωε) Êz , (26)
we arrive at
4 − +
Ĵzind = ∂ B̂1 . (27)
jμ0
This last equation shows that, in two dimensions, induced current density is obtained (accounting for
multiplication by 4/jμ0 ) by acting with the ∂− operator on the total B1+ -field. The simplified CSI–EPT
method is therefore called a first-order-induced current EPT method, since a first-order differentiation
of the B1+ -field essentially immediately results in an image of the induced current density.
As shown in Reference [20], after the induced current density is obtained, the corresponding
electric-field strength can be computed by solving a specific integral equation defined on Sbody .
With the electric-field strength now known, the conductivity and permittivity profiles within the slice
109
J. Imaging 2019, 5, 25
can be obtained from Equation (26). The overall first-order induced current density EPT algorithm can
be summarized as presented in Listing 2.
Listing 2. First-Order Induced Current EPT Algorithm (foIC-EPT).
• Given the measured B̂1+ -field in the midplane of the birdcage coil:
1. Determine the induced current density using Equation (27).

2. Determine the corresponding electric-field strength by solving a specific integral equation
(Equation (12) in Reference [20]).
3. Knowing the induced current density and the electric-field strength, determine conductivity
and permittivity profiles using Equation (26).
Further details about this algorithm can be found in Reference [20]. Finally, we note that the
above algorithm is a direct noniterative EPT method and, as opposed to CSI–EPT, requires the solution
of a system of equations (Step 2) to arrive at the reconstructed conductivity and permittivity profiles.
Fortunately, as demonstrated in Reference [20], this system of equations can efficiently be solved using
iterative solvers such as the generalized minimal residual (GMRES) method [29] and, typically, only a
small number of iterations is required to reach a prescribed error.
3. Methods and Results

To illustrate the performance of foIC-EPT and two- and three-dimensional CSI–EPT,
we reconstructed the conductivity and permittivity profiles of the head of anatomical human-body
model Duke from Virtual Family [27] (see Figure 1a,b), from noisefree B̂1+ -data. The head model
consists of 124 × 100 × 109 isotropic voxels with side lengths of 2 mm. The model was placed
inside an ideal high-pass birdcage coil (see Figure 1a) consisting of 16 rungs, each having a width of
25 mm. The coil has a radius of 150 mm, is 195 mm long, and is driven in quadrature at 128 MHz,
which corresponds to the operating frequency of a 3 T MRI system. The shield surrounding the coil
has a radius of 180 mm and length of 200 mm. Commercial EM simulation software (XFdtd, v.7.5,
Remcom State College, PA, USA) was used to obtain the background field {Êb , B̂b } as generated by
the high-pass birdcage coil. Finally, to investigate the difference between two- and three-dimensional
conductivity and permittivity reconstructions, we also considered a longitudinally uniform “head
model” in which the center slice was simply repeated in the longitudinal direction, thereby creating a
model with no variations in the longitudinal z direction within the head (see Figure 1c).
(a) (b) (c)

Figure 1. Birdcage coil and head models. (a) High-pass birdcage coil with the head model placed
inside, (b) Duke head model from Virtual Family [27], and (c) longitudinally uniform head model
obtained by repeating the center slice in the longitudinal direction.
110
J. Imaging 2019, 5, 25
3.1. Two-Dimensional CSI–EPT and foIC-EPT

The CSI–EPT method was originally implemented for two-dimensional configurations in
Reference [14] to study its potential as an EPT reconstruction method and to test if the method can
handle strongly inhomogeneous tissue profiles. Let us therefore start with a purely two-dimensional
reconstruction problem in which we attempt to reconstruct conductivity and permittivity profiles
within the center slice of the head model shown in Figure 2a. In this two-dimensional setting, we took
the background field in the midplane of the realistic birdcage coil shown in Figure 1a as the 2D
background field. The reconstructed conductivity and permittivity profiles obtained after 5000
iterations of the two-dimensional CSI–EPT method are shown in Figure 2b. It took the algorithm
approximately 86 s on an Intel i7-6700 CPU (Intel, Santa Clara, CA, USA) operating on Windows 7 with
Matlab 2016a (Mathworks, Natick, MA, USA) to arrive at these reconstructions, and we terminated the
algorithm after 5000 iterations since the objective function had already dropped below a 1.53 × 10−5
tolerance level at that point and essentially no significant improvements were obtained. In addition,
the foIC-EPT reconstruction profiles of conductivity and permittivity are shown in Figure 2c, and the
errors of CSI–EPT and foIC-EPT conductivity and permittivity reconstructions are shown in Figure 2d,e,
respectively. We observed that the quality of the foIC-EPT reconstructions was similar to CSI–EPT
even though it took foIC-EPT only a fraction of a second to produce these reconstructions (see Table 1
for details).
Table 1. Mean and standard deviation of reconstructed electrical properties in different tissue that are
apparent in the center slice of the head models using 2D CSI–EPT and foIC-EPT in a two-dimensional
setting. Units of σ and ε r are in siemens per meter and permittivity of free space, respectively.
Conductivity (σ) Relative Permittivity (εr )

True 2D CSI–EPT foIC-EPT True 2D CSI–EPT foIC-EPT
Fat 0.07 0.13 ± 0.10 0.18 ± 0.10 12.37 17.23 ± 7.89 19.27 ± 7.15
Red marrow 0.16 0.11 ± 0.06 0.15 ± 0.03 13.54 11.08 ± 2.52 13.68 ± 1.47
Bone 0.07 0.14 ± 0.15 0.21 ± 0.14 14.72 19.09 ± 8.05 21.73 ± 6.90
Eye lens 0.31 0.54 ± 0.11 0.78 ± 0.09 42.79 48.86 ± 5.26 51.52 ± 1.44
Nerve 0.35 0.74 ± 0.30 0.77 ± 0.26 44.07 51.49 ± 8.74 47.33 ± 7.42
Connective tissue 0.50 0.47 ± 0.12 0.44 ± 0.13 51.86 46.48 ± 7.56 40.54 ± 7.06
White matter 0.34 0.36 ± 0.04 0.38 ± 0.04 52.53 54.33 ± 3.12 55.37 ± 3.54
Muscle 0.72 0.63 ± 0.11 0.55 ± 0.13 63.49 56.76 ± 7.28 48.30 ± 8.72
Eye sclera 0.92 0.89 ± 0.15 0.87 ± 0.13 65.00 56.78 ± 6.67 50.26 ± 5.95
Skin 0.52 0.48 ± 0.08 0.36 ± 0.09 65.44 59.00 ± 8.95 42.29 ± 8.55
Hypothalamus 0.80 0.88 ± 0.11 0.91 ± 0.11 66.78 59.48 ± 4.91 54.12 ± 4.20
Eye vitreous humor 1.51 1.46 ± 0.13 1.40 ± 0.15 69.06 66.19 ± 4.86 61.03 ± 4.95
Cornea 1.06 0.92 ± 0.13 0.83 ± 0.11 71.46 61.52 ± 9.32 52.81 ± 5.84
Gray matter 0.59 0.61 ± 0.15 0.63 ± 0.15 73.52 72.68 ± 4.18 70.54 ± 4.92
Midbrain 0.83 0.84 ± 0.17 0.88 ± 0.18 79.74 78.58 ± 10.24 81.93 ± 12.33
Cerebrospinal fluid 2.14 1.90 ± 0.29 1.75 ± 0.29 84.04 80.66 ± 8.57 76.46 ± 11.31
Mucosa 2.28 1.50 ± 0.03 1.01 ± 0.02 116.00 78.07 ± 5.17 53.83 ± 2.77
111
J. Imaging 2019, 5, 25
Reconstruction True - Reconstruction

0 12 3 0 12 3
True 2-D CSI-EPT foIC-EPT 2-D CSI-EPT foIC-EPT
[S/m]
[S/m]
2
0.5
1.5 0.25
1 0
σ
0.5 −0.25
0 −0.5
100 30
80 15
60
0
εr
40
−15
20
1 −30
(a) (b) (c) (d) (e)
Figure 2. Reconstruction results from 2D reconstruction methods. (a) True model, (b) reconstruction
obtained after 5000 iterations of 2D CSI–EPT, and (c) reconstruction from foIC-EPT. Respective errors
are shown in (d,e). Top row shows conductivity and bottom row shows relative permittivity.
3.2. Three-Dimensional CSI–EPT

In a two-dimensional approach, the RF field is E-polarized with electric-field strength that is
longitudinal (Ê = Êz iz ) and magnetic-flux density that is transverse (B̂ = B̂x ix + B̂y iy ). Such an
approach was shown to be reasonable for a homogeneous cylindrical phantom in a central region of
a body coil consisting of elementary center-fed dipole antennas in Reference [25]. Indeed, when the
longitudinally uniform head model of Figure 1c was placed within our birdcage coil we also observed
that the x and y components of the electric-field strength in the central transverse slice were small
compared to its z component, as illustrated in the top rows of Figure 3a–c and Figure 4a,b.
Center +5 cm
0 12 3 0 12 3
| Êx | | Êy | | Êz | | Êx | | Êy | | Êz |
[V/m]
1,200
Homogeneous
1,000
800
600
Heterogeneous
400
200
0
(a) (b) (c) (d) (e) (f)
Figure 3. Magnitude of electric-field-strength components. x, y, and z components at (a–c) the
transversal midplane, and (d–f) the slice 5 cm higher, respectively. Top and bottom row show fields in
the case of a longitudinal homogeneous and heterogeneous object, respectively.
112
J. Imaging 2019, 5, 25
Center +5 cm
0 12 3 0 12 3
| Êx / Êz | | Êy / Êz | | Êx / Êz | | Êy / Êz |
1.5
Homogeneous
1
Heterogeneous
0.5
0
(a) (b) (c) (d)
Figure 4. Ratios of x and y components of the electric-field strength relative to its z component. Relative
field components at (a,b) the center slice and (c,d) the slice 5 cm higher. Top row is for the longitudinally
uniform object, bottom row for the object with longitudinal variations.
However, as we move away from the center slice in the longitudinally uniform head model
of Figure 1c, the magnitude of the x and y components of the electric-field strength starts to
increase, as illustrated in the top rows of Figure 3d–f and Figure 4c,d, where the magnitude of the
electric-field-strength components is shown in a slice located 5 cm above the central slice. We observed
that even though the transverse components of the electric-field strength were negligible within the
center slice, they could no longer be neglected when 5 cm away from it.
Furthermore, for the realistic heterogeneous head model of Figure 1b, a two-dimensional
E-polarized field assumption completely failed, as shown in the bottom rows of Figures 3a–f and 4a–d.
In the slice 5 cm above the central slice, and even within the central slice itself, the x and y components
of the electric-field strength could no longer be neglected and had to be taken into account in the full
Maxwell equations to properly describe RF field behavior within the head model.
To study the effects of longitudinal spatial variations of the tissue parameters on the B̂1+ -field,
we considered Equation (16) again and write it in form
B̂1+;sc = B tra + B lon , (28)

ω + sc
where B tra = c20
∂ Âz and B lon = − cω2 ∂z Â+;sc . Longitudinal variation term B lon is absent in a 2D
0
approach (see Equation (20)), since, in a 2D setting, the configuration is assumed to be invariant
in the longitudinal z-direction (∂z = 0). Figure 5, however, shows that, for both the longitudinally
homogeneous and realistic heterogeneous head model, the longitudinal variation term is significant
and cannot be ignored. Especially near the periphery of both head models, B lon contributes to the
scattered B̂1+ -field. More specifically, within a 1 cm outer boundary layer located in the center slice,
the mean of fraction |B lon / B̂1+;sc | is 1.18 and 1.25 for the homogeneous and inhomogeneous head
model, respectively; in the inner region, these means are 0.51 and 0.60, and similar averages were
obtained for the slice located 5 cm above the center slice. From these observations, it is clear that
longitudinal variations of transverse vector potential Â+;sc contribute to the scattered B̂1+ -field and
cannot be ignored.
113
J. Imaging 2019, 5, 25
Center +5 cm
0 12 3 0 12 3
|B tra | |B lon | |B lon / B̂1+;sc | |B tra | |B lon | |B lon / B̂1+;sc |
Homogeneous
Heterogeneous
0 0.5 1 0 0.5 1 0 0.5 1 1.5 0 0.5 1 0 0.5 1 0 0.5 1 1.5

·10−5 [T] ·10−5 [T] ·10−5 [T] ·10−5 [T]
(a) (b) (c) (d) (e) (f)
Figure 5. Magnitude of scattered B̂1+ terms. (a) Transverse variation and (b) longitudinal variation term
of the scattered B̂1+ , and (c) contributions of B lon with regard to B̂1+;sc at the center slice. (d–f) Same,
respectively, at a slice 5 cm higher in the head domain. Top row is in the case of a longitudinally
uniform object, the bottom row for the head model with longitudinal variations. (b) and (e) were
neglected in the 2D approach.
Up to this point, we compared 3D RF field structures with their 2D counterparts for a

longitudinally uniform and a realistic heterogeneous head model. In a two-dimensional configuration,
however, sources are invariant in the longitudinal direction as well, and we expect that, due to the
finite extent of the birdcage coil, additional deviations in the B̂1+ fields will be observed.
To further investigate this issue, we first determine the two-dimensional B̂1+ -field in the central
slice, as described in Section 3.1. The magnitude and phase of this field are shown in the top and
bottom row of Figure 6a, respectively. Subsequently, we consider RF excitation by the 3D birdcage coil,
but assume that the birdcage coil, including its currents, does not vary in the longitudinal direction.
For the longitudinally uniform head model, a B̂1+ -field as shown in Figure 6b is then obtained and
we observe that this field strongly resembles the 2D B̂1+ -field pattern of Figure 6a. Replacing the
longitudinal invariant currents in the rungs by the exact current, but keeping the homogeneous head
model, we obtain the B̂1+ -field pattern shown in Figure 6c. Agreement with the 2D field B̂1+ -field
pattern clearly deteriorates, and this correspondence becomes even worse for the realistic longitudinal
heterogeneous head model as shown in Figure 6d. Since the B̂1+ -field is used as an input for the
CSI–EPT method, accurate correspondence is obviously necessary for proper reconstruction. The 2D
CSI–EPT algorithm expects a 2D B̂1+ -field, as shown in Figure 6a for the center head slice, but in 3D
the B̂1+ -field from Figure 6d is present. Providing this 3D field as an input to a 2D CSI–EPT algorithm
would lead to inaccurate reconstructions in general.
To illustrate how these differences in actual fields (3D) and expected fields (2D) translate to
reconstruction errors, both two-dimensional algorithms were applied to quasi-three-dimensional data
using either 3D amplitudes or phases. Note that, in order to match 2D and 3D data, the maximum
absolute value of the B̂1+ -field of both datasets was taken to be equal. The results are depicted in
Figure 7b–e, from which it can be observed that permittivity is particularly sensitive to 2D violations.
This reconstruction difference between conductivity and permittivity is due to the fact that conduction
currents (σÊ) influence the B̂1+ -field to a much larger extent than displacement currents (jωεÊ) at 3 T.
114
J. Imaging 2019, 5, 25
3-D 3-D 3-D

longitudinal longitudinal longitudinal
homogeneous homogeneous heterogeneous
2-D object and coil object object [T]
·10−5
2
1.5
| B̂1+ |
π/2
0
φ+
−π/2
−π
(a) (b) (c) (d)

Figure 6. B̂1+ field comparison. (a) Total B̂1+ field assumed in the 2D setting, (b) total B̂1+ field obtained
in a 3D setting with longitudinal homogeneity of the object and coil, (c) of longitudinal homogeneity
of only the object, and (d) with longitudinal variations also of the object. Top row shows the B̂1+
magnitude, bottom row the B̂1+ phase.
2-D using 3-D magnitude 2-D using 3-D phase

0 12 3 0 12 3
True 2-D CSI-EPT foIC-EPT 2-D CSI-EPT foIC-EPT
[S/m]
2
1.5
1
σ
0.5
0
100
80
60
εr
40
20
1
(a) (b) (c) (d) (e)
Figure 7. Reconstruction results from 2D reconstruction methods using parts of 3D B̂1+ data. (a) True
model, (b,c) reconstruction results assuming 2D phase with 3D magnitude, and (d,e) reconstruction
results assuming 2D magnitude with 3D phase of the B̂1+ -field in the central transverse slice from
simulations with the longitudinal invariant head model. Top row shows conductivity and the bottom
row shows relative permittivity.
The reconstructions of the conductivity and relative permittivity profiles for the full 3D case
without any further assumptions, using 3D magnitude as well 3D phase B̂1+ data are shown for
the longitudinally uniform model in Figure 8a–f and for the realistic heterogeneous head model in
Figure 8g–l. Reconstructions are shown for the central slice profiles as well as for the profiles located
within the slice positioned 5 cm above the central slice. For comparison, 2D CSI–EPT reconstructions
based on 3D B̂1+ data are also presented. Relative residual error (norm of the difference between
the exact and reconstructed profile normalized by the norm of the exact profile, where the norm is
115
J. Imaging 2019, 5, 25
taken over the center slice) of Figure 8h is 0.7339 and 0.8263 for the conductivity and permittivity,
respectively, while the relative residual error of the conductivity and permittivity of Figure 8i is
0.3358 and 0.1587, respectively. Clearly, 2D CSI–EPT is unable to accurately reconstruct conductivity
and permittivity profiles. The 2D and 3D permittivity reconstructions are also less accurate than
conductivity reconstructions, indicating that B̂1+ -field data acquired at 3 T are less sensitive to
permittivity variations.
Center +5 cm
0 12 3 0 12 3
True 2-D 3-D True 2-D 3-D
[S/m]
2
1.5
1
σ
0.5
0
100
80
60
εr
40
20
1
(a) (b) (c) (d) (e) (f)
[S/m]
2
1.5
1
σ
0.5
0
100
80
60
εr
40
20
1
(g) (h) (i) (j) (k) (l)
Figure 8. Reconstruction comparison of 2D and 3D CSI–EPT on 3D B̂1+ fields after 5000 and
50,000 iterations, respectively. (a–c) true object, reconstruction with 2D CSI–EPT, and reconstruction
with 3D CSI–EPT for a homogeneous object and for the center slice. (d–f) Same as above, but for a
slice 5 cm higher. (g–l) same as (a–f), but in the case of a longitudinal inhomogeneous object. Top row
depicts the conductivity, bottom row shows permittivity.
Finally, to emphasize that 3D CSI–EPT is a fully three-dimensional volumetric reconstruction

method, we present a full 3D CSI–EPT reconstruction of the realistic head model obtained after
50,000 iterations based on 3D B̂1+ data in Figure 9. This number of iterations was chosen due to time
constraints, since it takes approximately 110 h on an Intel i7-6700 CPU operating on Windows 7 with
Matlab 2016a.
116
J. Imaging 2019, 5, 25
σ εr
0 12 3 0 12 3
True 3-D CSI-EPT True 3-D CSI-EPT
0 0.5 1 1.5 2 1 20 40 60 80 100

[S/m]
(a) (b) (c) (d)
Figure 9. Three-dimensional visualization of a section of the 3D CSI–EPT reconstruction of the
heterogeneous Duke head model (Figure 1b) after 50,000 iterations. (a,b) True and reconstructed
conductivity, and (c,d) true and reconstructed relative permittivity. Top slice that is visible is the slice
5 cm above the transverse midplane.
4. Discussion
We investigated the performance of two- and three-dimensional CSI–EPT in reconstructing
dielectric tissue profiles based on B̂1+ data collected inside the reconstruction slice or domain of interest.
Since these data has their support inside the reconstruction domain, EPT belongs to the class of so-called
hybrid inverse problems [8]. In CSI–EPT, reconstructing tissue parameters is posed as an optimization
problem in which an internal objective function, that is, an objective function that measures both
field and model discrepancies within the domain of interest, is minimized in an iterative manner.
Field discrepancies are measured by considering the L2 -norm of the difference between modeled
and measured data, while model discrepancies are measured by an L2 -norm that tells us how well a
conductivity and permittivity tissue profile, and the corresponding contrast source, satisfy Maxwell’s
equations. Including model discrepancies in the objective function is crucial to the performance of
CSI–EPT, since it has been shown that, without this term, unsatisfactory reconstruction results may be
obtained [24]. In addition to the tissue profiles, CSI–EPT reconstructs electric-field strength as well,
and may therefore also be used to predict the SAR that is induced inside the body or a body part of
interest [30], which is important for MR safety and hyperthermia-treatment planning, for example.
Finally, we also showed that, in two dimensions, an alternative noniterative and integral-based
reconstruction algorithm called foIC-EPT may be employed. This method is significantly faster
than 2D and 3D CSI–EPT, and reconstructs tissue profiles and corresponding electric-field strength
essentially in real time on a present-day standard laptop or PC (Intel i5-i7 or similar). However,
foIC-EPT is restricted to two-dimensional configurations since it exploits two-dimensional E-polarized
field structures. CSI–EPT, on the other hand, does not exploit any particular field structure and
can be extended to the vectorial three-dimensional case, turning CSI–EPT into a volumetric EPT
reconstruction method.
We carried out several comparisons between reconstructions obtained with 2D CSI–EPT, foIC-EPT,
and 3D CSI–EPT. Our simulations show that care needs to be exercised when a 2D reconstruction
approach is followed or reconstruction artefacts are obtained in the reconstructed dielectric tissue
profiles. Specifically, we showed that, by using 2D methods, erroneous reconstructions could be
obtained, since the longitudinal variations of the transverse vector potential are completely ignored
in the data model for the B̂1+ field. Moreover, vector potential itself is computed differently in 2D
and 3D since longitudinal invariance is assumed in the 2D case. In fact, the transverse electric field
and the longitudinal magnetic field vanish in 2D as a consequence of the (assumed) invariance of the
object and external sources along the longitudinal direction. In 3D, however, all components of the
electromagnetic field are present and their contributions to the measured data and object equations
117
J. Imaging 2019, 5, 25
have to be taken into account. Of course, in some situations, an E-polarized field structure may be
present in the midplane of a birdcage coil, but the scattered B̂1+ -field is also influenced by longitudinal
variations of transverse vector potential ∂z Âz . These equations can only be simplified to 2D if we can
guarantee that longitudinal invariance or the smoothness of certain field components can be imposed
before any reconstruction algorithm is applied to the measured data. Therefore, cylindrical body parts,
such as legs or arms, could be reliably reconstructed via 2D CSI–EPT, but this at least requires further
validation through simulations and measurements using cylindrical phantom models with known
dielectric characteristics.
No assumptions on the fields are imposed in 3D CSI–EPT and reconstruction errors due to such
assumptions are therefore avoided. Moreover, 3D CSI–EPT is a volumetric reconstruction method,
and is not restricted to a specific plane within the configuration. Reliable reconstructions can be
obtained within any desired domain of interest provided that B̂1+ data are available within this domain.
Unfortunately, computation times significantly increase when applying 3D CSI–EPT. Depending on
the number of unknowns in the EPT reconstruction problem, 3D CSI–EPT may take many iterations
to converge to the desired error tolerance, with total computation times of hours or days, even on
dedicated computers or servers. In future research, we focus on accelerating the convergence rate
of 3D CSI–EPT by including preconditioning techniques in CSI–EPT (as described in Reference [23],
for example) that exploit all a priori knowledge we have about the object or body part that needs to be
reconstructed. This knowledge can also be used to construct an accurate initial guess, thereby possibly
further accelerating CSI–EPT.
In our experiments, we used simulated B̂1+ -field data to test the performance of 2D and 3D
CSI–EPT on strongly inhomogeneous structures, and to study the differences between two- and
three-dimensional CSI–EPT approaches. In real-world measurements, the data obviously differ from
simulated data, and CSI–EPT should be adapted so that it can handle measured B̂1+ data. In this
respect, we identified three practical issues that need to be addressed, which are part of our current
CSI–EPT research.
First, in practice, the B̂1+ -field is obtained in polar form through separate amplitude and phase
measurements. In both cases, the collected data are contaminated with noise. Therefore, filtering
or regularization techniques that suppress the effects of noise should be incorporated in CSI–EPT.
Initial studies show that filtering of the data allows us to handle measured data in foIC-EPT [20] and,
as demonstrated in Reference [14], total-variation (TV) regularization may suppress noise effects in
CSI–EPT. However, due to the many possible choices for the regularization parameter in this method, it
is presently not clear for which parameter or parameter range the TV CSI–EPT scheme is most effective.
Second, the phase that is measured in practice is not the phase of the B̂1+ -field, but the so-called
transceive-phase from which the B̂1+ -phase can be extracted. To this end, the tranceive-phase
approximation is often applied, but the validity of this approximation is not fully understood, and
may lead to reconstruction errors in conductivity and permittivity profiles [28]. Fortunately, it is
shown in Reference [31] that improved B̂1+ phase approximations can be obtained from the tranceive
phase by incorporating an iterative-phase correction scheme in the CSI–EPT reconstruction algorithm.
This correction scheme seems to reliably retrieve B̂1+ phase maps from the measured tranceive phase,
and leads to improved conductivity and permittivity reconstructions compared with reconstructions
that are obtained when tranceive-phase approximation is applied. We will include this phase-correction
mechanism in future CSI–EPT implementations as well. Another option is to opt for phaseless
approaches as, for example, proposed in References [32,33].
Finally, in practice, the current densities in the transmit coil that generate the incident field
depend on the present object, and we must account for this loading effect as well. Specifically,
integral representations for the fields in CSI–EPT are obtained using scattered-field formalism in
which it is assumed that current density in the transmitting antenna is impressed and independent
of the scatterer that may be present in the configuration. In practice, however, these currents do
depend on the object, and, consequently, care must be taken when we compute the background
118
J. Imaging 2019, 5, 25
field in CSI–EPT. One approach is, therefore, to simulate this loading effect using a suitable coil and
body model in a commercial field solver and to extract current densities in the coil from this solver.
The background field in CSI–EPT (the field without any load) can then be computed using these
extracted currents. In this way, the loading effect encountered in practice can be incorporated in our
CSI–EPT reconstruction algorithm.
Our final aim is, of course, to turn CSI–EPT into a practical reconstruction method to obtain
accurate and reliable conductivity and permittivity tissue maps of an interior part of the human body
at MR operation frequencies. Reconstruction results based on simulated data are very promising,
and we think that, by addressing the practical issues discussed above, we will indeed make significant
progress towards a reliable EPT reconstruction method that provides us with accurate dielectric tissue
maps in practice.
Author Contributions: Conceptualization, R.R.; methodology, R.R.; software, R.L., P.F., W.B. and R.R.; validation,
R.L and P.F.; formal analysis, R.R.; investigation, R.L., P.F., W.B. and R.R; resources, R.L., P.F., W.B., A.W. and R.R.;
data curation, R.L., P.F., W.B. and R.R.; writing—original draft preparation, R.R., R.L. and P.F.; writing—review
and editing, R.L., P.F., W.B., A.W. and R.R.; visualization, R.L. and P.F.; supervision, R.R. and A.W.; project
administration, R.R. and A.W.; funding acquisition, R.R. and A.W.
Funding: The research of R. Leijsen was funded by the European Research Council Advanced NOMA MRI under
grant number 670629. The research of P. Fuchs was funded in part through a collaboration between the Delft
University of Technology and the Indian Institute of Science.
Acknowledgments: The authors thank Nico van den Berg of the Imaging Division of the University Medical
Center Utrecht for many valuable discussions and support.
References
1. Hartwig, V. Engineering for safety assurance in MRI: Analytical, numerical and experimental dosimetry.
Magn. Reson. Imaging 2015, 33, 681–689. [CrossRef] [PubMed]
2. Lagendijk, J.J.W. Hyperthermia treatment planning. Phys. Med. Biol. 2000, 45, R61–R76. [CrossRef] [PubMed]
3. Katscher, U.; van den Berg, C.A.T. Electric properties tomography: Biochemical, physical and technical
background, evaluation and clinical applications. NMR Biomed. 2017, 30, 1–15. [CrossRef] [PubMed]
4. Gabriel, S.; Lau, R.W.; Gabriel, C. The dielectric properties of biological tissues: II. Measurements in the
frequency range 10 Hz to 20 GHz. Phys. Med. Biol. 1996, 41, 2251–2269. [CrossRef] [PubMed]
5. Brown, B.H. Electrical impedance tomography (EIT): A review. J. Med. Eng. Technol. 2003, 27, 97–108.
[CrossRef] [PubMed]
6. Chandra, R.; Zhou, H.; Balasingham, I.; Narayanan, R. On the Opportunities and Challenges in Microwave
Medical Sensing and Imaging. IEEE Trans. Biomed. Eng. 2015, 62, 1667–1682. [CrossRef]
7. Zhang, X.; Liu, J.; He, B. Magnetic-Resonance-Based Electrical Properties Tomography: A Review. IEEE Rev.
8. Bal, L. Hybrid inverse problems and internal functionals. In Inverse Problems and Applications: Inside Out II;
Uhlmann, G., Ed.; Cambridge University Press: New York, NY, USA, 2012; pp. 325–368, ISBN 978-1-107-03201-9.
9. Haacke, E.M.; Petropoulos, L.S.; Nilges, E.W.; Wu, D.H. Extraction of conductivity and permittivity using
magnetic resonance imaging. Phys. Med. Biol. 1991, 36, 723–734. [CrossRef]
10. Katscher, U.; Voigt, T.; Findeklee, C.; Vernickel, P.; Nehrke, K.; Doessel, O. Determination of Electric
Conductivity and Local SAR Via B1 Mapping. IEEE Trans. Med. Imaging 2009, 28, 1365–1374. [CrossRef]
11. Voigt, T.; Katscher, U.; Doessel, O. Quantitative Conductivity and Permittivity Imaging of the Human Brain
Using Electric Properties Tomography. Magn. Reson. Med. 2011, 66, 456–466. [CrossRef]
12. Van Lier, A.L.H.M.W.; Brunner, D.O.; Pruessmann, K.P.; Klomp, D.W.J.; Luijten, P.R.; Lagendijk, J.J.W.;
van den Berg, C.A.T. B1+ Phase Mapping at 7T and its Application for In Vivo Electrical Conductivity
Mapping. Magn. Reson. Med. 2012, 67, 552–561. [CrossRef] [PubMed]
13. Hafalir, F.S.; Oran, O.F.; Gurler, N.; Ider, Y.Z. Convection-Reaction Equation Based Magnetic Resonance
Electrical Properties Tomography (cr-MREPT). IEEE Trans. Med. Imaging 2014, 33, 777–793. [CrossRef]
[PubMed]
119
J. Imaging 2019, 5, 25
14. Balidemaj, E.; van den Berg, C.A.T.; Trinks, J.; van Lier, A.L.H.M.W.; Nederveen, A.J.; Stalpers, L.J.A.;
Crezee, H.; Remis, R.F. CSI-EPT: A Contrast Source Inversion Approach for Improved MRI-Based Electrical
Properties Tomography. IEEE Trans. Med. Imaging 2015, 34, 1788–1796. 2015.2404944. [CrossRef] [PubMed]
15. Serrallés, J.E.C.; Daniel, L.; White, J.K.; Sodickson, D.K.; Lattanzi, R.; Polimeridis, A.G. Global Maxwell
Tomography: A novel technique for electrical properties mapping based on MR measurements and volume
integral equation formulations. In Proceedings of the 2016 IEEE International Symposium on Antennas and
Propagation (APSURSI), Fajardo, Puerto Rico, 26 June–1 July 2016; pp. 1395–1396. [CrossRef]
16. Hong, R.; Li, S.; Zhang, J.; Zhang, Y.; Liu, N.; Yu, Z.; Liu, Q.H. 3-D MRI-Based Electrical Properties
Tomography Using the Volume Integral Equation Method. IEEE Trans. Microw. Technol. 2017, 65, 4802–4811.
[CrossRef]
17. Arduino, A.; Zilberti, L.; Chiampi, M.; Bottauscio, O. CSI-EPT in Presence of RF-Shield for MR-Coils.
IEEE Trans. Med. Imaging 2017, 36, 1396–1404. [CrossRef] [PubMed]
18. Rahimov, A.; Litman, A.; Ferrand, G. MRI-based electric properties tomography with a quasi-Newton
approach. Inverse Probl. 2017, 33, 105004. [CrossRef]
19. Gurler, N.; Ider, Y.Z. Gradient-Based Electrical Conductivity Imaging Using MR Phase. Magn. Reson. Med.
2017, 77, 137–150. [CrossRef]
20. Fuchs, P.S.; Mandija, S.; Stijnman, P.R.S.; Brink, W.M.; van den Berg, C.A.T.; Remis, R.F. First-Order Induced
Current Density Imaging and Electrical Properties Tomography in MRI. IEEE Trans. Comput. Imaging 2018, 4,
624–631. [CrossRef]
21. Van den Berg, P.M.; Abubakar, A. Contrast source inversion method: State of art. Prog. Electromagn. Res.
2001, 34, 189–218. [CrossRef]
[CrossRef]
23. Van den Berg, P.M.; van Broekhoven, A.L.; Abubakar, A. Extended contrast source inversion. Inverse Probl.
1999, 15, 1325–1344. [CrossRef]
24. Van den Berg, P.M.; Haak, K.F.I. Profile inversion by error reduction in the source type integral equations.
In Wavefields and Reciprocity–Proceedings of a Symposium Held in Honour of Professor dr. A.T. de Hoop;
van den Berg, P.M., Blok, H., Fokkema, J.T., Eds.; Delft University Press: Delft, The Netherlands, 1996;
pp. 87–98, ISBN 90-407-1402-9.
25. Van den Bergen, B.; Stolk, C.C.; van den Berg, J.B.; Lagendijk, J.J.W.; Van den Berg, C.A.T. Ultra fast
electromagnetic field computations for RF multi-transmit techniques in high field MRI. Phys. Med. Biol. 2009,
54, 1253–1264. [CrossRef] [PubMed]
26. Leijsen, R.L.; Brink, W.M.; van den Berg, C.A.T.; Webb, A.G.; Remis, R.F. 3-D Contrast Source
Inversion-Electrical Properties Tomography. IEEE Trans. Med. Imaging 2018, 9, 2080–2089. [CrossRef]
[PubMed]
27. Christ, A.; Kainz, W.; Hahn, E.G.; Honegger, K.; Zefferer, M.; Neufeld, E.; Rascher, W.; Janka, R.; Bautz, W.;
Chen, J.; et al. The Virtual Family–development of surface-based anatomical models of two adults and two
children for dosimetric simulations. Phys. Med. Biol. 2010, 55, N23–N38. [CrossRef] [PubMed]
28. Van Lier, A.L.H.M.W.; Raaijmakers, A.; Voigt, T.; Lagendijk, J.J.W.; Luijten, P.R.; Katscher, U.;
van den Berg, C.A.T. Electrical Properties Tomography in the Human Brain at 1.5, 3, and 7T: A Comparison
Study. Magn. Reson. Med. 2014, 71, 354–363. [CrossRef] [PubMed]
29. Saad, Y. Iterative Methods for Sparse Linear Systems; SIAM: Philadelphia, PA, USA, 2003; ISBN 0-89871-534-2.
30. Balidemaj, E.; van den Berg, C.A.T.; van Lier, A.L.H.M.W.; Nederveen, A.J.; Stalpers, L.J.A.; Crezee, H.;
Remis, R.F. B1-based SAR reconstruction using contrast source inversion–electric properties tomography
(CSI-EPT). Med. Biol. Eng. Comput. 2017, 55, 225–233. [CrossRef] [PubMed]
31. Stijnman, P.R.S.; Mandija, S.; Fuchs, P.S.; Remis, R.F.; van den Berg, C.A.T. Transceive Phase Corrected
Contrast Source Inversion-Electrical Properties Tomography. In Proceedings of the ISMRM Joint Annual
Meeting, Paris, France, 15–18 June 2018; Volume 5087.
120
J. Imaging 2019, 5, 25
32. Arduino, A.; Bottauscio, O.; Chiampi, M.; Zilberti, L. Magnetic resonance-based imaging of human electric
properties with phaseless contrast source inversion. Inverse Probl. 2018, 34, 084002. [CrossRef]
33. Bevacqua, M.T.; Bellizzi, G.G.; Isernia, T.; Crocco, L. A method for quantitative imaging of electrical
properties of human tissues from only amplitude electromagnetic data. Inverse Probl. 2019, 35, 025006.
[CrossRef]
121
Journal of
Imaging
Article
Novel Stopping Criteria for Optimization-Based
Microwave Breast Imaging Algorithms
Cameron Kaye *, Ian Jeffrey and Joe LoVetri
Electrical and Computer Engineering, University of Manitoba, Winnipeg, MB R3T 5V6, Canada;
[email protected] (I.J.); [email protected] (J.L.)
Received: 25 April 2019; Accepted: 14 May 2019; Published: 22 May 2019
Abstract: A discontinuous Galerkin formulation of the Contrast Source Inversion algorithm

(DGM-CSI) for microwave breast imaging employing a frequency-cycling reconstruction technique
has been modified here to include a set of automated stopping criteria that determine a suitable time to
shift imaging frequencies and to globally terminate the reconstruction. Recent studies have explored
the use of tissue-dependent geometrical mapping of the well-reconstructed real part to its imaginary
part as initial guesses during consecutive frequency hops. This practice was shown to improve
resulting 2D images of the dielectric properties of synthetic breast models, but a fixed number of
iterations was used to halt DGM-CSI inversions arbitrarily. Herein, a new set of stopping conditions
is introduced based on an intelligent statistical analysis of a window of past iterations of data error
using the two-sample Kolmogorov-Smirnov (K-S) test. This non-parametric goodness-of-fit test
establishes a pattern in the data error distribution, indicating an appropriate time to shift frequencies,
or terminate the algorithm. The proposed stopping criteria are shown to improve the efficiency of
DGM-CSI while yielding images of equivalent quality to assigning an often liberally overestimated
number of iterations per reconstruction.
Keywords: breast imaging; microwave imaging; discontinuous Galerkin method (DGM); contrast
source inversion (CSI); stopping criteria; Kolmogorov-Smirnov (K-S) test
1. Introduction
Microwave imaging (MWI) has made steady progress towards widespread clinical application
in breast cancer detection and monitoring over the past decade, offering a safety advantage over
established x-ray-based modalities due to its use of non-ionizing radiation and a significant cost benefit
over magnetic resonance imaging (MRI). While its currently attainable spatial resolution does not
match that of cancer screening tools like mammography, concurrent use of MWI of the breast in its
current state could lead to decreased false positive rates among certain population groups, particularly
pregnant women or those with radiographically dense (“Category C” and “Category D”) breasts,
as classified by the American College of Radiology’s Breast Imaging Reporting and Data System
(BI-RADS) [1,2].
Methods of utilizing MWI for biomedical applications have been investigated for almost half
a century [3,4]. Some recent representative research can be found in [5]. MWI algorithms can be
split into those producing qualitative images, where the reconstructed value at each pixel is a
relative quantity, and those producing quantitative reconstructions of a physical property, typically
the complex-valued permittivity. Examples of qualitative algorithms are those utilized in so-called
radar imaging techniques [6], and those retrieving only qualitative aspects of the target, such as
its support [7,8]. The interest in this paper are quantitative imaging methods that solve the full
electromagnetic inverse problem.

J. Imaging 2019, 5, 55
It is well known that the electromagnetic inverse problem associated with MWI, where one
attempts to reconstruct the complex-valued permittivity of an object-of-interest (OI), is ill-posed [9].
The ill-posedness stems from the non-uniqueness of the inverse-source problem where one attempts
to reconstruct the electromagnetic sources responsible for a remotely measured field. In the
electromagnetic inverse problem, these sources are the contrast-sources that are illuminated by the
interrogating field. Various regularization techniques have been developed over several decades
of research to deal with the ill-posedness of the problem [10]. The electromagnetic inverse problem
is not only ill-posed but also non-linear with repect to the two unknowns within the inaccessible
imaging domain, the electromagnetic field and the permittivity. The Contrast Source Inversion (CSI)
technique effectively linearizes the problem by casting the data-error norm in terms of contrast-source
variables, which vary with the interrogating field, and regularization is performed by introducing the
so-called Maxwellian regularizer which is written in terms of the contrast-sources and the contrast [11].
Much research has been performed on advancing the CSI algorithm since it was first reported,
including the addition of a multiplicative regularizer [12].
For biomedical imaging applications, which are of interest in the context of the present work,
the incorporation of discretized numerical inversion models into CSI that are based on either
the finite-element method (FEM) or the discontinuous Galerkin method (DGM) forward solvers
have provided much flexiblity and several advantages for the overall inversion process [13–15].
Most importantly, these forward solvers have allowed one to incorporate prior information into the
inversion model that has the effect of regularizing the inversion, enabling high-contrast reconstructions
such as are required for breast imaging [16–19]. The flexibility of these partial differential equation
(PDE) based solvers has also allowed advancements in data-acquisition systems utilized to acquire
scattered-field data for MWI [20,21]. Other potential advancements in MWI for breast tumor detection
include the use of contrast agents [22]
Recent work in multi-frequency imaging has shown promise in increasing the spatial resolution
and correspondingly, the amount of distinguishable anatomic detail in MWI reconstructions.
For instance, while the use of “frequency hopping” has long been shown to be an effective means of
obtaining images from high-frequency data by feeding low-frequency solutions as “initial guesses”
into inversion algorithms [23], a so-called “frequency cycling” and “tissue-dependent mapping”
framework has been adopted in recent studies to perform multi-frequency imaging demonstrating
significant enhancement in the quality of resulting 2D images [24,25]. In those works, an arbitrary
fixed number of iterations was used as the stopping condition for the imaging algorithm, but it was
noted that further exploration into more intelligent stopping criteria was justified, particularly based
on the observation that running the algorithm longer than necessary could deteriorate the imaging
results. In fact, to date, there has been little exploration into optimizing the conditions to which
a microwave imaging algorithm should adhere when determining the appropriate point to move
on from intermediate inversions to the next frequency in a multi-frequency sequence, or when to
terminate reconstructions entirely.
While monitoring the convergence of data error has been the most common objective approach
to halting microwave imaging algorithms, as it is often difficult to ascertain whether a chosen error
limit will undershoot, overshoot, or consistently meet the attainable image reconstruction quality for
a given dataset, several early studies focusing on synthetic and experimental imaging performance
fell into the same habit of simply pre-assigning an arbitrary number of iterations for the algorithm to
complete, chosen for convenience or through trial and error [11,26–28].
Although there have been previous attempts to develop more intelligent stopping criteria for
imaging algorithms in recent literature [17], a more thorough statistical analysis of the data error is
undertaken here and employed as part of a logical multi-variable framework of stopping criteria
for a frequency-cycling tissue-dependent mapping formulation of Contrast Source Inversion (CSI).
For comparison purposes, some 2D imaging results from a recent study are included in this work to
demonstrate the significant gains in efficiency and image quality granted through the incorporation of
123
J. Imaging 2019, 5, 55
the described stopping criteria [25]. Reconstruction results are subjected to quantitative analysis using
common image error metrics to provide an objective measurement of these improvements.
2. Materials and Methods
2.1. DGM-CSI Algorithm

An implementation of CSI using a high-order frequency-domain formulation of Maxwell’s curl
equations has been used for all 2D breast images presented here, employing the discontinuous Galerkin
method (DGM) as a forward solver. Nodal coefficients in high-order polynomial expansions represent
unknown field and contrast quantities in the resulting DGM-CSI algorithm. Independent from
this choice of implementation, among the benefits of CSI is its use of operators that are functions
of frequency and the material properties of the background medium only, which do not change
from iteration to iteration [11]. As aforementioned, further improvement in performance has been
demonstrated through the use of a tissue-dependent mapping process and the practice of cycling back
to low-frequency reconstructions [25], and a brief overview of this process is provided in Section 2.2.
Further details specific to DGM-CSI can be found in [14,29], but a summary of the relevant quantities
and definitions involved in the CSI cost functional is included here, since it pertains to the discussion
of data and domain error that follows.
The microwave imaging problem in Figure 1 depicts an object of interest (OI) with an unknown
dielectric contrast χ(r ), a function of position vector r, typically defined as
ε̂ r (r ) − ε̂ b
χ(r ) = , (1)
ε̂ b
where ε̂ r is the complex relative permittivity and ε̂ b is the complex background permittivity. In breast
imaging applications, the background permittivity is usually represented as the homogeneous
immersion medium in which the breast is submerged and the imaging domain D is an area typically
chosen to be contained by the outer skin layer of the breast. An array of t = 1, 2, . . . , T transmitters is
used to generate probing incident fields; the resulting scattered fields are measured by an array of R
receivers that make up a discrete measurement surface S .
Figure 1. Two-dimensional representation of the imaging problem. The object of interest (OI) has an
unknown contrast χ, where D is the imaging domain and S is the surface containing the transmitters
(Tx) and receivers (Rx).
It is assumed that the reader is familiar with the standard CSI algorithm, and only the aspects
important to the proposed stopping criteria will be reviewed. The CSI algorithm solves for the contrast
χ(r ) of breast tissue within D from this scattered field data by minimizing the following two-part
cost functional:
F CSI (χ, wt ) = F S (wt ) + F D (χ, wt ), (2)
124
J. Imaging 2019, 5, 55
where both the contrast χ(r ) and the transmitter-dependent contrast sources wt (r ) = χ(r )utot
t
t ) are unknown and confined to the domain D .
(a product of the contrast and the total field utot
The first term of Equation (2) represents the “data error”, given by
" "2
∑t "usct "
t − L S { wt } S
F S ( wt ) = " sct "2 , (3)
∑t u" "
t S
and correspondingly, the second term represents the “domain error”,
" "2
∑t "χuinc "
t − wt + χ LD { wt } D
F D (χ, wt ) = " inc "2 , (4)
∑t "χu " t D
with ·D and ·S denoting the L2 norms on D and S . The measured scattered field data is represented
t at R receiver locations per transmitter t, and the incident field inside D is ut . The forward
above by usct inc
operator LS converts contrast source estimates wt in D to scattered field values at R receiver points per
transmitter t on the measurement surface S ; LD performs a similar function, but transforms contrast
sources to scattered field values within D . For further simplicity, the “data error” per iteration i can be
defined from the numerator of Equation (3) as
(i ) (i )
ρt = dt − L S { wt } (5)
and correspondingly, the “domain error” from the numerator of Equation (4) as
(i ) (i ) (i )
rt = χ(i) uinc (i )
t − w t + χ L D { w t }. (6)
These error values, particularly the data error of Equation (5), are subject to analysis for evaluating
convergence behaviour of DGM-CSI in this study, described in Section 2.3.
2.2. Frequency-Cycling Tissue-Dependent Mapping Technique

The well-documented dielectric properties of breast tissues lead to the observation that the
geometry of the real and imaginary parts of tissues’ permittivity profiles should be qualitatively
similar [24,25]. The quantitative permittivity values of different types of breast tissue also form
reasonably discrete ranges that are highly correlated between real and imaginary components, making
it easy to stratify these values into expected tissue types (i.e., fat, transitional, fibroglandular, and
cancer) [30]. Through a tissue-dependent mapping of the imaginary part based on geometric and
quantitative reconstructions of the dielectric constant, an “artificial” imaginary component was
proposed as part of the initial guess for the next frequency dataset in the imaging sequence, to
good effect [24,25].
In addition to tissue-dependent mapping, this previous work invoked “frequency cycling”
as opposed to conventional frequency hopping, where frequency cycling simply continues the
reconstruction process past the highest frequency in a multi-frequency sequence by returning to
the lowest frequency dataset (with an initial guess based on the highest frequency solution) [25].
The initial studies using DGM-CSI have shown that this practice preserves the fine detail acquired
from high-frequency data, and produces better results than a single incremental frequency-hopping
sequence operating for the same total number of iterations. In a practical microwave imaging scenario
without the benefit of foreknowledge of this total number of iterations, however, a potential drawback
to frequency cycling would be the increased number of iterations required to produce images when
arbitrary or overly-stringent stopping conditions are applied.
125
J. Imaging 2019, 5, 55
2.3. Stopping Criteria for Single-Frequency Reconstructions

As described in Section 2.1, iterative inversion methods such as DGM-CSI attempt to solve the
inverse scattering problem by minimizing the cost functional of Equation (2), which includes a data
(i )
error ρt (Equation (5)) represented by the difference between the measured and computed fields on S
for a given transmitter t. If U is the total number of data points across all receivers R and sources T,
then the U × 1 data-error vector ρ̄(i) can be represented as a concatenation of every transmitter’s data
errors, such that
(i ) (i ) (i )
ρ̄(i) = [ρ1 ; ρ2 ; . . . ; ρ T ] (7)
(i )
After all contrast sources wt are updated for each transmitter t, the inversion process by
convention performs a convergence check that calculates the new data-error vector ρ̄(i+1) . The goal is
to reduce this new data error as much as possible, and aside from simply assigning a maximum number
of iterations, one could monitor the data error until it falls below an arbitrarily-chosen tolerance level δ
such that some form of a termination condition examining its sum, maximum value, or norm applies,
commonly implemented as
if (ρ̄(i+1) < δ) stop.
Alternatively, the difference between the normalized data-error vector at two successive iterations,
ρ̄(i) and ρ̄(i+1) , can be calculated until it falls within a prescribed tolerance to check for convergence [26].
Such conditions would seem to be reasonable criteria by which termination of the inversion
algorithm should take place. However, the total data error often has little bearing on the accuracy of
the resulting reconstruction, despite properly converging to a preset minimum. While liberally large
values of error tolerance will logically terminate the imaging algorithm prematurely, it has also been
observed that conservatively small choices of error tolerance will result in a poorer quality image.
This outcome is likely due to the supposition that the minimization of a cost functional towards zero
arguably means accounting for any and all error in modeling, calibration, and measurements (i.e.,
noise) through non-physical changes in the reconstructions. The apparent inadequacy of having
the data error converge to within an arbitrary tolerance alone served as motivation to develop a
multifaceted approach to stopping criteria that improves the reconstructions of a frequency-cycling
reconstruction technique through statistical analysis of data and domain error of each iteration.
This proposed framework stops the imaging algorithm based primarily on the statistical
distribution of the contributing terms to the data-error vector ρ̄. Referring the reader back to Equation (7),
assuming a best-case scenario where the forward operator LS is able to perfectly model the field
behaviour of the imaging system and the contrast sources wt had been iteratively solved to the exact
solution, the data error ρt would be reduced to the difference between the field values dt experimentally
measured at the receivers for all given transmitters and the ideal noise-free field values generated
(i )
by LS {wt }. This difference would amount to the noise of the experimental system, which would
presumably converge to a known pattern, such as a Gaussian distribution.
However, it is clear that even if one could acquire completely noise-free measurement data, it is
not practical, or even possible, to find a numerical model that could exactly match it. This so-called
modeling error, attributed to mismatches between the approximate representation of the imaging
environment LS (which varies by formulation, mesh granularity, order number, etc.) and the true,
unknowable functions that perfectly predict field behaviours within the system, invalidates the
assumption that the remaining data-error vector ρ̄ (upon convergence of the imaging algorithm to a
solution) consists simply of noise. Modeling error is the primary reason why data calibration must be
performed on experimental data, and it is not a trivial issue to address in MWI [31].
Correspondingly, since the contributions to the data error are multifactorial in nature and vary
with each experimental set-up and possibly with each target, it will not necessarily converge to a known
family of statistical distributions. Use of any of the myriad of statistical normality tests available (such
as the Anderson–Darling or Shapiro–Wilk test) on the data error as part of the proposed algorithmic
126
J. Imaging 2019, 5, 55
stopping criteria would therefore be unreliable. However, the profile of the data error distribution
(regardless of its final form) should reach a steady state when the algorithm has sufficiently converged
to a particular solution. To confirm that this trend occurs during a reconstruction requires comparison
of the current iteration’s data error distribution to those of multiple previous iterations, which is
accomplished in this work with an implementation of the two-sample Kolmogorov-Smirnov (K-S) test.
The two-sample K-S test is a non-parametric goodness-of-fit hypothesis test that evaluates the
maximum absolute difference D between the cumulative distribution functions (CDF) of two sample
data vectors over a range of x in each data set. Recall that the CDF (FX ) of a continuous random variable
X can be expressed as the integral of its probability density function (PDF) f x (that is, a function whose
value at any given sample can be interpreted as providing a relative likelihood that the value of the
random variable would be equal to that sample), such that
x
FX ( x ) = f X (t) dt. (8)
−∞
In the discrete case, FX ( x ) would be equivalent to the proportion of values in f X ( x ) less than or
equal to a specified value of x.
Suppose the data-error vector PDFs from two consecutive iterations of the DGM-CSI imaging
algorithm, ρ̄(i) and ρ̄(i+1) , each have some observed CDF, say P̂(i) ( x ) and P̂(i+1) ( x ); that is, for any
specified value of x, the value of P̂( x ) is the proportion of error values less than or equal to x in the
corresponding dataset. Consider the quantity
D = max(| P̂(i+1) ( x ) − P̂(i) ( x )|), (9)

x
that represents the maximum error between two CDFs. The K-S test makes use of this D value to
return a probability, in the form of a scalar asymptotic p-value in the range of [0, 1], which represents
the likelihood that the two samples were drawn from the same distribution. More precisely, the p-value
is defined as the probability of observing a test statistic (D) as extreme or more extreme than the
observed value, for the null hypothesis that the data in these two vectors are from the same continuous
distribution [32–35]. In the implementation of the test, the decision to reject the null hypothesis is
based on comparing the returned p-value with a preset significance level (commonly 5% for many
applications). The test becomes very accurate for large sample sizes, and is believed to be reasonably
accurate for sample sizes U1 and U2 , such that (U1 × U2 )/(U1 + U2 ) ≥ 4, which is the case for
a collection of data error values in a system of 24 transmitters and receivers operating at a single
frequency (representing a total of 576 data points).
While the strength of the two-sample K-S test lies in determining when two samples are from
differing distributions, the nature of the test guarantees that the smaller the value of D, the higher the
returned p-value will be. This feature allows the test’s p-value to be used as a numerical indicator
for how likely the two samples were drawn from the same distribution (i.e., how closely their CDFs
match), as opposed to its primary use in rejecting the null hypothesis.
Empirical studies of the real and imaginary parts of the data error of DGM-CSI reconstructions
indeed revealed that the shape of the probability distribution curves fluctuate significantly with large
variances in early iterations (Figure 2) but gradually reach a steady state in later iterations as expected,
once all remarkable subjective change in the resulting images had ceased (Figure 3). Since any two
neighboring iterations’ data error may only change slightly and be interpreted as being drawn from
the same distribution by the two-sample K-S test, it was decided that a sliding window of past
iterations’ data errors should be examined and compared to that of the current iteration. Moreover, the
threshold value that the returned p-value should surpass for any given K-S test within this sliding
window needed to be high enough to similarly reject subtler variations in two iterations of data
error that would otherwise pass the goodness-of-fit test. This wider breadth of analysis ensured that
127
J. Imaging 2019, 5, 55
more gradual changes were detected so that any corresponding stopping criteria would not halt the
reconstruction prematurely.
Figure 2. Examples of histograms demonstrating estimates of the probability distribution of the real
and imaginary parts of the data error (ρE ) during a DGM-CSI inversion of an arbitrary 2D synthetic
breast model early in the reconstruction process (iteration 20).
Figure 3. Examples of histograms demonstrating estimates of the probability distribution of the real
and imaginary parts of the data error (ρE ) during the same DGM-CSI inversion as Figure 2, later in the
reconstruction process (iteration 190).
Therefore, the window size of iterations under scrutiny, the p-value threshold (assumed constant
over the window), and the percentage of that window’s iterations reaching or surpassing that p-value
threshold, are all parameters to the stopping criteria for the algorithm based on the K-S test, which
is applied separately for the real and imaginary parts of each iteration’s data error. A preliminary
study attempting to establish the effects of these parameters was undertaken for the frequency-cycled
reconstruction of a 2D breast model, described in Section 3.1.
It should be noted that the statistical analysis governing the aforementioned stopping criteria
only makes use of the data error, which may be sufficient for certain MWI algorithms (such as the
distorted Born iterative method) whose minimized cost functional consists solely of a data error
term. However, as shown in Section 2.1, CSI-based iterative inversion algorithms make use of cost
functionals containing a second domain error term dealing with total field values inside the imaging
domain [11,26]. Therefore, another ad hoc stopping condition based on the domain error is also
prudently employed in the algorithm, as described in the next subsection.
128
J. Imaging 2019, 5, 55
2.4. Global Termination of Multi-Frequency Reconstructions

The stopping conditions described in Section 2.3 are appropriate for halting single-frequency
inversions, but multiple-frequency reconstructions have thus far not been addressed. In a
frequency-hopping scenario, the stopping criteria based on K-S tests alone can be easily employed to
determine when each individual inversion should terminate and move on to the next frequency, with
global termination occurring following reconstruction of the highest frequency.
However, in a frequency-cycling reconstruction scheme, the cycle of reconstructions could
theoretically continue ad infinitum without another explicit set of rules in place governing the global
termination of the imaging process. A separate, secondary ad hoc global termination criterion was
therefore implemented, based on a previous study [17]; once each frequency’s dataset in the cycle is
subjected to the K-S test stopping criteria at least once, if the percentage of relative change in domain
error between two successive iterations falls below 0.1%, then it is deemed appropriate to globally
terminate the frequency cycle and end the reconstruction. This provision could also serve as a fallback
contingency in the event that the data error-based conditions turn out to be too strict to satisfy for a
given individual frequency.
2.5. Full Description of Multi-Frequency Imaging Procedure

The key concepts involved in the proposed multi-frequency imaging procedure,
while implemented herein for the DGM-CSI algorithm, can easily be applied (with minor
modifications) to other iterative algorithms widely used in MWI, including Gauss-Newton Inversion
(GNI) and distorted Born Iterative Method (DBIM). As such, a descriptive approach to the technique
is undertaken here in order to keep its applicability as general as possible. Taking into account
the stopping criteria described throughout Section 2.3 and using the procedural description of the
frequency-cycling reconstruction approach with tissue-dependent mapping described in [25], the new
consolidated multi-frequency imaging procedure is proposed as follows:
1. The real and imaginary parts of the complex permittivity are reconstructed using the lowest
frequency data available (e.g., 1.0 GHz). The termination point of this reconstruction is dictated
by the results of successive two-sample K-S tests performed on both the real and imaginary
parts of the data error separately, comparing the current iteration to those of a sliding window of
past iterations, governed by a choice of parameters for p-value, window size, and percentage
of windowed iterations reaching this p-value threshold. For robustness, a back-up termination
condition may be implemented, either related to the relative change in domain error for CSI-based
algorithms as described in Section 2.4, or a maximum number of iterations.
2. A point-by-point search through the reconstructed real part of each nodal basis coefficient in
the DGM-CSI mesh (or more generally, each mesh element or pixel of the reconstructed image)
classifies the type of breast tissue. This classification is based solely on the range of expected
values of dielectric constant at that frequency, as outlined in [25].
3. An initial guess for the next imaging frequency (e.g., 2.0 GHz) is generated using the
tissue-dependent mapping process [24,25]. It consists of the unmodified real parts of the
reconstructed ε r at the mesh nodal points, and a new imaginary part created from a simple
linear interpolation of the expected range of dielectric loss values, based on the appropriate
Cole-Cole models of tissues classified in Step 2. This technique preserves the geometry of the real
and imaginary parts of the solution.
4. This new initial guess for the complex permittivity is used to run the inversion algorithm at the
next frequency (e.g., 2.0 GHz). As per the procedure outlined in [25], the user may choose to keep
the imaginary part constant during this inversion and update only the real part to converge to a
new solution. This “anchoring” process has been shown to improve overall imaging results due to
the tendency of CSI-based inversion algorithms to cause significant deterioration of the imaginary
129
J. Imaging 2019, 5, 55
part at high-frequency reconstructions. Again, the aforementioned parameterized stopping criteria

would be primarily employed to determine the appropriate point to halt this reconstruction.
5. If more than two frequencies are used in the frequency hop, steps 2–4 are repeated as necessary
until the reconstruction of the final frequency of the succession is complete (e.g., 3.0 GHz).
This succession may include “frequency cycling”; that is, returning the inversion algorithm to the
lowest frequency data and incrementally stepping through each frequency again [25].
If a simple frequency-hopping scheme is employed, the reconstruction will terminate after the
highest-frequency inversion, again as decided by the parameterized stopping criteria based on K-S
tests of the data error. However, if a frequency-cycling scheme is used, and if the imaginary part of
the solution has been “anchored” in place following the first inversion in Step 1, there are further
considerations that come into play for global termination of the imaging process, and to address
concerns of full CSI optimization:
• When each available dataset in the frequency cycle has been used at least once to contribute to the
overall image reconstruction, a global termination criterion will become active, which will monitor the
relative change in the domain error between successive iterations (Section 2.4). If this relative change
falls below 0.1% at any point, the current reconstruction is halted and the frequency cycle is broken.
• Regardless of the frequency at which the algorithm was halted by this relative domain error
threshold, if the imaginary part of the solution has been continuously held constant during the
frequency cycle after Step 1, one last initial guess is generated as in Step 3 and a final reconstruction
is run at the lowest frequency available (e.g., 1.0 GHz) with both the real and imaginary parts
allowed to converge to a solution (i.e., the imaginary part is no longer “anchored”). This final
inversion is terminated by the parameterized stopping criteria or a relative change of domain
error between successive iterations falling below 0.1%, whichever occurs first. The purpose of this
final run is to demonstrate the stability of the final solution and ensure that its imaginary part,
despite being originally based on the geometry and tissue properties of the real part, does indeed
satisfy full CSI optimization.
2.6. Synthetic Breast Models

Initial testing of the described stopping criteria and comparison to fixed-iteration scenarios
were carried out on synthetic transverse magnetic (TM) data collected from a two-dimenstional
MRI-derived BI-RADS Category C (heterogeneously dense) 2D breast model supplied by the University
of Calgary, derived from an MRI slice of a cancer patient with a breast tumour visible at the
“3 o’clock” position. To demonstrate the robustness of the technique across different breast phantoms,
an additional two-dimensional MRI-derived BI-RADS Category D (“extremely dense”) healthy breast
model supplied by the University of Wisconsin’s public database (ID: 070604PA2) was employed. Both
models are depicted at three different frequencies (1.0 GHz, 2.0 GHz, and 3.0 GHz) in Figures 4 and 5,
illustrating the similarities between breast tissues’ dielectric permittivity at the low-gigahertz band.
The values of the tissue-dependent complex permittivity in all cases were calculated for every frequency
using an appropriate fitted single-pole Cole-Cole model (or equivalently, a single-pole Debye model as
the exponent parameter α = 0), and for the University of Calgary model, then subjected to random
perturbations of ±10% [36].
Data was collected for both models in a low-loss background of r = 23 − 1.13i with
24 transmitters and 24 receivers evenly distributed at a radius of 10 cm, employing a finite-element
method (FEM) forward solver with a finely-discretized mesh independent from that used for the
DGM-CSI inversion. The synthetic data was collected for open boundary conditions at 1.0 GHz,
2.0 GHz, and 3.0 GHz. To test the robustness of the stopping criteria to problems with PEC boundaries,
and to better reflect the capabilities of a recent University of Manitoba imaging system prototype [37],
data was collected a second time with the Category C model surrounded by a circular PEC boundary
with a radius of 15 cm at a reduced bandwidth. The frequencies used for the PEC-bounded case were
130
J. Imaging 2019, 5, 55
1.0 GHz, 1.25 GHz, and 1.5 GHz. Uniformly-distributed noise at 5% of the maximum field magnitude
was used to corrupt the Ez scattered electric field data for the Category C model, and several noise
levels (3%, 5%, 7.5%, and 10%) were tested for the Category D model to evaluate the effect of noise
on the stopping criteria’s performance. The skin thickness and inner skin boundary both remained
unknown; the only prior information used during inversions was the outer skin boundary.
Figure 4. Complex dielectric properties of University of Calgary 2D synthetic breast model at

1.0 GHz (top), 2.0 GHz (middle) and 3.0 GHz (bottom).
Figure 5. Complex dielectric properties of University of Wisconsin 2D synthetic breast model at

1.0 GHz (top), 2.0 GHz (middle) and 3.0 GHz (bottom).
131
J. Imaging 2019, 5, 55
2.7. Error Calculation

An objective method of evaluating the DGM-CSI algorithm’s performance for differing imaging
scenarios was accomplished by comparing the reconstruction results to the original Category C
and Category D synthetic models of the current frequency being used in the reconstruction cycle,
unless otherwise specified. The models and reconstructions each employed different underlying
meshes, so their complex permittivity values were interpolated onto a common 250 × 250-pixel square
grid (N = 62,500) that encompassed an area covering ( x1 , y1 ) = (−0.07, −0.07) [m] to ( x2 , y2 ) =
(0.07, 0.07) [m]. The relative error of L1 and L2 norms of the difference between the model (actual)
complex permittivity ε̂ ract and the reconstructed complex permittivity ε̂ rrec over each pixel k were
used as primitives to evaluate the performance of the imaging algorithm. For convenience, these will
hereafter be referred to as the relative error norms REN1 and REN2 , and are calculated and expressed
as percentages, where
ε̂ ract − ε̂ rrec 1

REN1 (ε̂ ract , ε̂ rrec ) = × 100%
ε̂ ract 1
N
∑ |ε̂ ract (k ) − ε̂ rrec (k)|
k =1
= N
× 100%,
∑ |ε̂ ract (k )|
k =1 (10)
ε̂ ract − ε̂ rrec 2

REN2 (ε̂ ract , ε̂ rrec ) = × 100%
ε̂ ract 2
%
& N
&
& ∑ [ε̂ ract (k) − ε̂ rrec (k )]2
& k =1
= &
& × 100%.
' N
∑ [ε̂ ract (k)]2
k =1 (11)
The raw numerical value of the norms will be slightly reduced since the square grid covers an area
outside of the imaging domain of the problem, which will consist of common unaltered background
permittivity in both the model and reconstruction and thus represent pixels of zero error. Since the
REN values serve primarily as a means of monitoring iterative trends in solution convergence and to
demonstrate comparative improvement between imaging scenarios with equally-sized interpolated
grids and imaging domains, the impact of this artificial error reduction is of little significance.
3. Results and Discussion
3.1. Imaging with Open Boundaries

As mentioned in the introduction of the stopping criteria in Section 2.3, the approach had three
variables: the p-value threshold that any given two-sample K-S test would use to conclude the two
data error samples were drawn from the same distribution, the window size of past iterations’ data
errors compared to that of the current iteration, and the percentage of that window’s iterations that
would need to to reach or surpass the chosen p-value threshold in order to halt the reconstruction at the
current frequency, deemed the “pass percentage”. For the same Category C breast model in Figure 4,
DGM-CSI was run in a frequency-cycled configuration using the rules described in Section 2.5 for
every combination of p-value thresholds of 0.90, 0.95 and 0.99, window sizes of 10, 30, and 50, and pass
percentages of 80%, 90%, and 100%. There were 27 simulations whose total number of iterations would
vary based on the laxity or stringency of stopping criteria associated with low and high variable values
respectively. Error norms (REN1 and REN2 ) for each scenario were plotted for every frequency jump
132
J. Imaging 2019, 5, 55
(Figure 6), and the best variable combination (resulting in the lowest error metric magnitudes upon
termination) was found to be a p-value threshold of 0.99, a window size of 30, and a pass percentage
of 80%. This combination of values was subsequently used for all future simulations employing the
stopping criteria. Note that in Figure 6 and in all subsequent plots of REN values, the x-axis’ number
of iterations represents the cumulative total throughout all frequency-hopping or frequency-cycling
steps of the reconstruction.

íLWHUDWLRQZLQGRZV
íLWHUDWLRQZLQGRZV
íLWHUDWLRQZLQGRZV
5(1
5(1

5(1

SDVVUDWHIRUíLWHUDWLRQZLQGRZZLWK
SíYDOXHWKUHVKROGRI

&XPXODWLYHLWHUDWLRQQXPEHU
Figure 6. Relative error norms of open-boundary 2D DGM-CSI reconstructions across several choices
of stopping criteria parameter values. Each curve in the figure corresponds to a frequency-cycled
inversion for a particular choice of the three parameters (Section 3.1), with the window sizes coded
by color for convenience. Data points on each curve correspond to REN values for the reconstructed
model at each frequency change. Arrows point to the final REN values of simulations that terminated
with the lowest relative error norms, indicating the best combination of parameters among those tested
for this imaging scenario.
Having established an appropriate set of variables to govern the stopping criteria, a series of
simulations adhering to the full description of the technique in Section 2.5 could be carried out to
demonstrate its benefit in efficiency over conventional methods. The REN values of three scenarios (A,
B, and C from Table 1) comparing fixed-iteration reconstructions (with and without tissue-dependent
mapping) to the use of both tissue-dependent mapping and the stopping criteria, are shown in Figure 7.
The final imaging reconstructions of these three cases are also depicted in Figure 8. As the proposed
stopping conditions introduce an unpredictable halt to the reconstruction cycle that could occur at
any frequency and also includes a mandatory return to the lowest frequency (in this case, 1.0 GHz) for
the final optimization, each REN data point may not represent consecutive frequencies of the cycle.
For consistency, the REN values shown in Figure 7 are therefore all calculated based on the same
2.0 GHz model from Figure 4. For clarity, the frequency associated with each data point plotted in
Figure 7, along with the triggering condition that caused a change in frequency or termination in the
reconstruction (when relevant), is shown in Table 1.
133
J. Imaging 2019, 5, 55

6FHQDULR$
6FHQDULR%
6FHQDULR&
5(1
5(1

5(1

Figure 7. Relative error norms of frequency-hopping and frequency-cycled reconstructions: Scenario A

(solid line)—without tissue-dependent mapping at 400 iterations per frequency terminating after first
inversion of 3.0 GHz, Scenario B (dashed line)—with tissue-dependent mapping at 200 iterations per
frequency, cycling through reconstruction frequencies once (with imaginary component “anchored”
following initial 1.0 GHz inversion), Scenario C (dotted line)—with tissue-dependent mapping and
stopping criteria in place, terminating after two consecutive inversions of 1.0 GHz data (one with the
imaginary component “anchored” and the final run with the imaginary component freely optimized
according to the guidelines of Section 2.5). See Table 1 for further details.
Table 1. Reconstruction progression of open boundary scenarios (Figure 7).
Frequency:
Scenario TM SC No. of Iterations (Stopping Condition) Total
Components Reconstructed
1.0 GHz: 2.0 GHz: 3.0 GHz:
A No No 400 (F) 400 (F) 400 (F) 1200
Re, Im Re, Im Re, Im
1.0 GHz: 2.0 GHz: 3.0 GHz: 1.0 GHz: 2.0 GHz: 3.0 GHz:
B Yes No 200 (F) 200 (F) 200 (F) 200 (F) 200 (F) 200 (F) 1200
Re, Im Re Re Re Re Re
C Yes Yes * 140 (KS) 190 (KS) 255 (KS) 53 (KS) 89 (DE) 41 (KS) 768
Re, Im Re Re Re Re Re, Im
TM = tissue mapping; SC = stopping criteria; F = fixed; KS = Kolmogorov-Smirnov test window;

DE = Domain error; Re = Real part; Im = Imaginary part. * 80% of a 30-iteration data error window
must reach or surpass a K-S test p-value threshold of 0.99.
134
J. Imaging 2019, 5, 55
Although DGM-CSI, like other iterative microwave imaging algorithms, performs reasonably
well in recovering the real component of breast models it reconstructs, it has difficulty producing an
accurate profile of the imaginary part, most notably at higher frequencies. This deficiency is illustrated
from the Scenario A inversion of the Category C breast model in the top image of Figure 8. Subjectively,
the benefit of employing the tissue-dependent mapping in the fidelity of the imaginary component is
obvious in both the middle and bottom images. It has been already shown elsewhere that the described
mapping technique and “anchoring” of the imaginary component at each subsequent frequency stage,
along with the practice of frequency cycling, imparts a benefit to the quality of the real part of the
reconstructions following the same number of iterations [24,25]. However, these other studies only
tested reconstructions that used fixed numbers of iterations as termination conditions; the addition of
the stopping criteria for the latter case here (Scenario C) does not appreciably degrade the quality of
the final image, while reducing the total number of iterations by over 35%.
Figure 8. Final results of DGM-CSI frequency-hopping and frequency-cycled complex dielectric

property reconstruction of synthetic breast model using 1.0–3.0 GHz data, without modification of
intermediate initial guesses (Scenario A—top), using tissue-dependent mapping at fixed 200 iterations
per frequency (Scenario B—middle), and employing stopping criteria and tissue-dependent mapping
(Scenario C—bottom).
3.2. Imaging with PEC Boundaries

A set of simulations similar to those carried out in the previous section were performed using
PEC boundaries and data from frequencies of 1.0 GHz, 1.25 GHz and 1.50 GHz, to demonstrate
this method’s robustness to smaller frequency bandwidths and different boundary conditions, as
mentioned in Section 2.6. Scenarios D and E are imaging simulations recreated from an earlier study
that employed fixed-iteration reconstructions exclusively [25]; they are included in order to emphasize
the benefit of adding stopping criteria in Scenario F. The imaging results are shown in Figure 9 with a
breakdown of the reconstruction progression documented in Table 2. A plot of the relative error norms
135
J. Imaging 2019, 5, 55
again similar to the open boundary cases, employing REN calculations based on the original 1.25 GHz
model, is also provided in Figure 10.
As explained in [25], it was expected that the results would not be as impressive as open-boundary
scenarios since PEC-bounded problems are generally more difficult to solve, and data was taken at
lower frequencies. However, it is again clear from Figure 9 that there is an improvement in the recovery
of the imaginary component when tissue-dependent mapping is used (middle and bottom images),
more so for Scenario F. A modest improvement in the real part of ε r for Scenarios E and F is also noted.
Without the use of stopping criteria, however, Scenario E overshoots the expected values of dielectric
constant in several areas. Its REN values depicted in Figure 10 correspondingly suffer increases relating
to the fact that the reconstruction cycle has perhaps been allowed to run too long, as pointed out in [25].
The introduction of the stopping criteria has effectively prevented this degradation by terminating the
reconstruction cycle close to the observed nadir of these error values from Scenario E, while once again
demonstrating a sizable increase in computational efficiency by reducing the number of iterations by
an impressive 66%.

property reconstruction of PEC-bounded synthetic breast model using 1.0–1.5 GHz data, without
modification of intermediate initial guesses (Scenario D—top), using tissue-dependent mapping
at fixed 200 iterations per frequency (Scenario E—middle), and employing stopping criteria and
tissue-dependent mapping (Scenario F—bottom).
136
J. Imaging 2019, 5, 55
Table 2. Reconstruction progression of PEC-bounded scenarios (Figure 10).
Frequency:
Scenario TM SC No. of Iterations (Stopping Condition) Total
1.0 GHz: 1.25 GHz: 1.5 GHz:
D No No 400 (F) 400 (F) 400 (F) 1200
E Yes No 200 (F) 200 (F) 200 (F) 200 (F) 200 (F) 200 (F) 1200
Re, Im Re Re Re Re Re
1.0 GHz: 1.25 GHz: 1.5 GHz: 1.0 GHz: 1.0 GHz:
F Yes Yes * 123 (KS) 124 (KS) 97 (KS) 14 (DE) 48 (DE) 406
Re, Im Re Re Re Re, Im

6FHQDULR'
6FHQDULR(
5(1
6FHQDULR)
5(1
5(1

Figure 10. Relative error norms of reconstructions with PEC boundaries: Scenario D (solid
line)—without tissue-dependent mapping at 400 iterations per frequency terminating after first
inversion of 1.5 GHz data, Scenario E (dashed line)—with tissue-dependent mapping at 200 iterations
per frequency, cycling through reconstruction frequencies once (with imaginary component “anchored”
following initial 1.0 GHz inversion), Scenario F (dotted line)—with tissue-dependent mapping and
stopping criteria in place, terminating after two consecutive inversions of 1.0 GHz data (one with the
imaginary component “anchored” and the final run with the imaginary component freely optimized
according to the guidelines of Section 2.5). See Table 2 for further details.
3.3. Effect of Noise Levels

To validate that the stopping criteria developed and tested on the BI-RADS Category C 2D model
would perform similarly on different breast models and be robust to varying levels of noise, another
full set of simulations were carried out on the aforementioned Category D model from the University
of Wisconsin database using open boundary data at 1.0 GHz, 2.0 GHz, and 3.0 GHz (Figure 5).
The uniformly-distributed noise used to corrupt the Ez scattered electric field data was generated at
137
J. Imaging 2019, 5, 55
3%, 5%, 7.5%, and 10% of the maximum field magnitude. The imaging results are shown in Figure 11
with a breakdown of the reconstruction progression documented in Table 3. A plot of the relative
error norms again similar to the Category C cases, employing REN calculations based on the original
2.0 GHz Category D model, is also provided in Figure 12.
property reconstruction of open-boundary Category D synthetic breast model using 1.0–3.0 GHz data:
(a) Scenario G—without tissue-dependent mapping at 400 iterations per frequency terminating after
first inversion of 3.0 GHz (5% noise), (b) Scenario H—with tissue-dependent mapping and stopping
criteria in place at 3% noise, (c) Scenario I—with tissue-dependent mapping and stopping criteria in
place at 5% noise, (d) Scenario J—with tissue-dependent mapping and stopping criteria in place at 7.5%
noise, (e) Scenario K—with tissue-dependent mapping and stopping criteria in place at 10% noise. See
Table 3 for further details.
As shown in the final images of Figure 11, there is an obvious improvement using
tissue-dependent mapping and stopping criteria over the simple frequency-hopping approach
(Scenario G), especially in the imaginary part, whose geometry is not reconstructed properly and
138
J. Imaging 2019, 5, 55
contains a large overshooting boundary artifact. However, remarkably, there is very little subjective
difference in the images reconstructed from data corrupted with different levels of noise (from 3% to
10% for Scenarios H through K), with subtle changes in the real and imaginary parts visible only with
careful scrutiny.
The relative error norms of all scenarios depicted in Figure 12 and the reconstruction break-down
in Table 3 offer more objective measurements of the effect of noise. Specifically, lower final REN values
are observed for images inverted from data containing lower noise levels, as would be expected. Also,
the stopping criteria transitions frequencies and halts the entire reconstruction cycle sooner for higher
noise levels. These trends are consistent for all reconstructions employing tissue-dependent mapping
and stopping criteria in each level of noise used in the simulations. For instance, data containing 3%
noise (Scenario H) runs longest, for a total of 759 iterations; this total is observed to gradually decrease
at the 5% and 7.5% noise levels, and data containing 10% noise (Scenario K) runs for only 448 iterations.
These observations are congruous with the theory introduced in Section 2.3 that data containing
higher magnitudes of noise would have less useful information for the optimization algorithm to
extract before the distribution of data error reached a steady state (i.e., before the algorithm began to
reconstruct noise) and continuing the inversion would be superfluous, or perhaps even detrimental
to the final image. Overall, while these additional 2D simulations do not necessarily represent
a fully comprehensive investigation across all possible breast models and imaging scenarios, the
performance of the proposed stopping criteria appears satisfactory on the Category D model, and
behaves as expected on noisier datasets, even with parameter choices derived from the Category C
model reconstructions.
Table 3. Reconstruction progression of Category D model noise-variant scenarios (Figure 12).
Frequency :
Scenario (Noise %) TM SC No. of Iterations (Stopping Condition) Total
1.0 GHz: 2.0 GHz: 3.0 GHz:
G (5%) No No 400 (F) 400 (F) 400 (F) 1200
H (3%) Yes Yes * 155 (KS) 140 (KS) 265 (KS) 50 (KS) 52 (DE) 43 (DE) 759
Re, Im Re Re Re Re Re, Im
I (5%) Yes Yes * 139 (KS) 140 (KS) 265 (KS) 34 (DE) 8 (DE) 586
J (7.5%) Yes Yes * 91 (KS) 112 (KS) 234 (KS) 37 (DE) 13 (DE) 487
K (10%) Yes Yes * 87 (KS) 90 (KS) 228 (KS) 11 (DE) 32 (DE) 448
139
J. Imaging 2019, 5, 55

6FHQDULR*
6FHQDULR+
6FHQDULR,
6FHQDULR-

6FHQDULR.
5(1
5(1
5(1

Figure 12. Relative error norms of frequency-hopping and frequency-cycled reconstructions of

Category D model: Scenario G (blue solid line)—without tissue-dependent mapping at 400 iterations
per frequency terminating after first inversion of 3.0 GHz (5% noise), Scenario H (red dashed line)—with
tissue-dependent mapping and stopping criteria in place at 3% noise, Scenario I (black dotted
line)—with tissue-dependent mapping and stopping criteria in place at 5% noise, Scenario J (magenta
solid line)—with tissue-dependent mapping and stopping criteria in place at 7.5% noise, Scenario K
(green dashed line)—with tissue-dependent mapping and stopping criteria in place at 10% noise. See
Table 3 for further details.
4. Conclusions
It has been demonstrated that a stopping condition based on the similarities of the statistical
distribution of data error at successive iterations has the capability of significantly speeding up the
DGM-CSI algorithm while having no appreciable detrimental effect on the final imaging results, even
yielding superior solutions than fixed-iteration scenarios. This boost of efficiency is granted through
novel use of sequential two-sample K-S tests on the data error to determine appropriate times to
halt the imaging algorithm, truncating likely unnecessary (and in some cases counterproductive)
reconstruction iterations.
Whether or not the best set of parameters used in this analysis represents the true optimal values
for the stopping criteria for other breast models, or if these same values are applicable to 3D inversions,
is open to debate. However, given the promising imaging results obtained from this initial set of
tests, and the gains in efficiency provided through significant reduction of the number of iterations
needed to converge to these solutions, it was judged appropriate to keep the best parameters from this
introductory exploration of the method and leave a more comprehensive examination of parameter
optimization for future investigations.
The most interesting form of the proposed multi-frequency imaging procedure employs
tissue-dependent mapping and “anchoring” of the imaginary part along with the new stopping criteria.
This ad hoc procedure also returns the reconstruction cycle to stable low-frequency data after recovering
anatomical detail from higher frequencies, allowing both real and imaginary parts to properly converge
before global termination. The advantages gained by this use of both tissue-dependent mapping and
novel stopping criteria have also been demonstrated to be robust for different breast models, frequency
bandwidths, boundary conditions, and noise levels.
140
J. Imaging 2019, 5, 55
In fact, the stopping criteria have a more dramatic effect on the imaging results obtained from the
more complex PEC-bounded problem and data with higher noise levels, suggesting this technique
may be especially useful for avoiding image degradation and reconstruction artifacts in difficult cases
such as full 3D reconstructions in semi-resonant enclosures, including experimental measurements
that may be excessively noisy or contain poor frequency data unsuitable for inversion. To explore
this possibility, preliminary extension of this technique to the 3D DGM-CSI algorithm developed at
the University of Manitoba is a logical next step, along with parameter optimization studies of the
stopping criteria for data from experimental systems.
Author Contributions: Conceptualization, C.K., I.J. and J.L.; Methodology, C.K., I.J. and J.L.; Resources, C.K.,
I.J. and J.L.; Software, C.K. and I.J.; Supervision, J.L.; Validation, C.K. and I.J.; Writing—original draft, C.K.;
Writing—review & editing, C.K., I.J. and J.L.
Funding: This research is partially supported by the Natural Sciences and Engineering Research Council of
Canada (NSERC).
Acknowledgments: The authors would like to thank the University of Calgary for providing the 2D Category C
synthetic breast model used for the reconstructions presented in this work.
References
1. Trentham-Dietz, A.; Kerlikowske, K.; Stout, N.K.; Miglioretti, D.L.; Schechter, C.B.; Ergun, M.A.;
van den Broek, J.J.; Alagoz, O.; Sprague, B.L.; van Ravesteyn, N.T.; et al. Tailoring breast cancer screening
intervals by breast density and risk for women aged 50 years or older: Collaborative modeling of screening
outcomes. Ann. Intern. Med. 2016, 165, 700–712. [CrossRef] [PubMed]
2. D’Orsi, C.J.; Mendelson, E.B.; Morris, E.A. (Eds.) ACR BI-RADS Atlas: Breast Imaging Reporting and Data
System, 5th ed.; American College of Radiology: Reston, VA, USA, 2012.
3. Boerner, W.M.; Brand, H.; Cram, L.A.; Gjessing, D.T.; Jordan, A.K.; Keydel, W.; Schwierz, G.; Vogel, M. (Eds).
Inverse Methods in Electromagnetic Imaging; D. Reidel Publishing: Dordrecht, The Netherlands, 1985.
4. Larsen, L.E.; Jacobi, J.H. (Eds.) Medical Applications of Microwave Imaging; IEEE Press: New York, NY, USA, 1986.
5. Kosmas, P.; Crocco, L. Introduction to Special Issue on “Electromagnetic Technologies for Medical
Diagnostics: Fundamental Issues, Clinical Applications and Perspectives”. Diagnostics 2019, 9, 19. [CrossRef]
6. Sill, J.M.; Fear, E.C. Tissue sensing adaptive radar for breast cancer detection—Experimental investigation of
simple tumor models. IEEE Trans. Microw. Theory Tech. 2005, 53, 3312–3319. [CrossRef]
7. Delbary, F.; Brignone, M.; Bozza, G.; Aramini, R.; Piana, M. A visualization method for breast cancer detection
using microwaves. SIAM J. Appl. Math. 2010, 70, 2509–2533. [CrossRef]
8. Cakoni, F; Colton, D.; Monk, P. Qualitative Methods in Inverse Electromagnetic Scattering Theory: Inverse
Scattering for Anisotropic Media. IEEE Antennas Propag. Mag. 2017, 59, 24–33. [CrossRef]
9. Colton, D.; Kress, R. Inverse Acoustic and Electromagnetic Scattering Theory; Springer Science & Business Media:
Berlin, Germany, 2012; Volume 93.
10. Engl, H.W.; Hanke, M.; Neubauer, A. Regularization of Inverse Problems; Kluwer Academic Publishers:
Dordrecht, The Netherlands, 1996.
[CrossRef]
12. Abubakar, A.; Van den Berg, P.M.; Mallorqui, J.J. Imaging of biomedical data using a multiplicative
regularized contrast source inversion method. IEEE Trans. Microw . Theory Tech. 2002, 50, 1761–1771.
[CrossRef]
13. Zakaria, A.; Gilmore, C.; LoVetri, J. Finite-element contrast source inversion method for microwave imaging.
14. Jeffrey, I.; Geddert, N.; Brown, K.; LoVetri, J. The Time-Harmonic Discontinuous Galerkin Method as a
Robust Forward Solver for Microwave Imaging Applications. Prog. Electromagn. Res. 2015, 154, 1–21.
[CrossRef]
141
J. Imaging 2019, 5, 55
15. Brown, K.G.; Geddert, N.; Asefi, M.; LoVetri, J.; Jeffrey, I. Hybridizable Discontinuous Galerkin Method
Contrast Source Inversion of 2-D and 3-D Dielectric and Magnetic Targets. IEEE Trans. Microw. Theory Tech.
2019, 67, 1766–1777. [CrossRef]
16. Baran, A.; Kurrant, D.; Zakaria, A.; Fear, E.; LoVetri, J. Breast Imaging Using Microwave Tomography with
Radar-Based Tissue-Regions Estimation. Prog. Electromagn. Res. 2014, 149, 161–171. [CrossRef]
17. Kurrant, D.J.; Baran, A.; Fear, E.C.; LoVetri, J. Integrating prior information into microwave tomography
Part 1: Impact of detail on image quality. Med. Phys. 2017, 44, 6461–6481. [CrossRef]
18. Kurrant, D.; Baran, A.; LoVetri, J.; Fear, E. Integrating prior information into microwave tomography part
2: Impact of errors in prior information on microwave tomography image quality. Med. Phys. 2017, 44,
6482–6503. [CrossRef]
19. Abdollahi, N.; Kurrant, D.; Mojabi, P.; Omer, M.; Fear, E.; LoVetri, J. Incorporation of Ultrasonic Prior
Information for Improving Quantitative Microwave Imaging of Breast. IEEE J. Multiscale Multiphys.
Comput. Tech. 2019, 4, 98–110. [CrossRef]
20. Nemez, K.; Baran, A.; Asefi, M.; LoVetri, J. Modeling Error and Calibration Techniques for a Faceted Metallic
Chamber for Magnetic Field Microwave Imaging. IEEE Trans. Microw. Theory Tech. 2007, 65, 4347–4356.
[CrossRef]
21. Asefi, M.; Baran, A.; LoVetri, J. An Experimental Phantom Study for Air-Based Quasi-Resonant Microwave
Breast Imaging. IEEE Trans. Microw . Theory Tech. 2019. [CrossRef]
22. Shea, J.D.; Kosmas, P.; Van Veen, B.D.; Hagness, S.C. Contrast-enhanced microwave imaging of breast tumors:
A computational study using 3D realistic numerical phantoms. Inverse Probl. 2010, 26, 074009. [CrossRef]
23. Chew, W.; Lin, J. A frequency-hopping approach for microwave imaging of large inhomogeneous bodies.
IEEE Microw. Guided Wave Lett. 1995, 5, 439–441. [CrossRef]
24. Kaye, C.; Jeffrey, I.; LoVetri, J. Enhancement of multi-frequency microwave breast images using a
tissue-dependent mapping technique with discontinuous Galerkin contrast source inversion. In Proceedings
of the AES 2016—4th Advanced Electromagnetics Symposium, Malaga, Spain, 26–28 July 2016; pp. 37–39.
25. Kaye, C.; Jeffrey, I.; LoVetri, J. Improvement of Multi-Frequency Microwave Breast Imaging through
Frequency Cycling and Tissue-Dependent Mapping. IEEE Trans. Antennas Propag. 2019, submitted.
26. Abubakar, A.; van den Berg, P.M.; Habashy, T.M. Application of the multiplicative regularized contrast
source inversion method on TM- and TE-polarized experimental Fresnel data. Inverse Probl. 2005, 21, S5–S13.
[CrossRef]
27. Van den Berg, P.M.; van Broekhoven, A.L.; Abubakar, A. Extended contrast source inversion. Inverse Probl.
1999, 15, 1325–1344. [CrossRef]
28. Bloemenkamp, R.F.; Abubakar, A.; van den Berg, P.M. Inversion of experimental multi-frequency data using
the contrast source inversion method. Inverse Probl. 2001, 17, 1611–1622. [CrossRef]
29. Jeffrey, I.; Zakaria, A.; LoVetri, J. Microwave Imaging by Mixed-Order Discontinuous Galerkin Contrast
Source Inversion. In Proceedings of the 2014 XXXIst URSI General Assembly and Scientific Symposium
(URSI GASS 2014), Beijing, China, 16–23 August 2014; pp. 255–258.
30. Lazebnik, M.; Popovic, D.; McCartney, L.; Watkins, C.B.; Lindstrom, M.J.; Harter, J.; Sewall, S.; Ogilvie, T.;
Magliocc, M.A.; Breslin, T.M.; et al. A large-scale study of the ultrawideband microwave dielectric properties
of normal, benign and malignant breast tissues obtained from cancer surgeries. Phys. Med. Biol. 2007, 52,
6093–6115. [CrossRef] [PubMed]
31. Kaye, C. Development and Calibration of Microwave Tomography Imaging Systems for Biomedical
Applications Using Computational Electromagnetics. Master’s Thesis, University of Manitoba,
Winnipeg, MB, Canada, 2009.
32. Massey, F.J. The Kolmogorov-Smirnov Test for Goodness of Fit. J. Am. Stat. Assoc. 1951, 46, 68–78. [CrossRef]
33. Miller, L.H. Table of Percentage Points of Kolmogorov Statistics. J. Am. Stat. Assoc. 1956, 51, 111–121.
[CrossRef]
34. Marsaglia, G.; Tsang, W.; Wang, J. Evaluating Kolmogorov’s Distribution. J. Stat. Softw. 2003, 8, 1–4.
[CrossRef]
35. MathWorks. Two-Sample Kolmogorov-Smirnov Test (R2018a). Available online: www.mathworks.com/
help/stats/kstest2.html (accessed on 16 March 2018).
142
J. Imaging 2019, 5, 55
36. Kurrant, D.J.; Fear, E.C. Regional estimation of the dielectric properties of inhomogeneous objects using
near-field reflection data. Inverse Probl. 2012, 28, 075001. [CrossRef]
37. Nemez, K.; Asefi, M.; Baran, A.; LoVetri, J. A faceted magnetic field probe resonant chamber for 3D breast
MWI: A synthetic study. In Proceedings of the 2016 17th International Symposium on Antenna Technology
and Applied Electromagnetics (ANTEM), Montreal, QC, Canada, 10–13 July 2016; pp. 322–324.
143
Journal of
Imaging
Article
Comparing Radar-Based Breast Imaging Algorithm
Performance with Realistic Patient-Specific
Permittivity Estimation
Declan O’Loughlin *, Bárbara L. Oliveira, Martin Glavin, Edward Jones and Martin O’Halloran
Electrical and Electronic Engineering, National University of Ireland Galway, Galway H91CF50, Ireland ;
[email protected] (B.L.O); [email protected] (M.G.); [email protected] (E.J.);
[email protected] (M.O.)
Received: 2 September 2019; Accepted: 11 November 2019; Published: 19 November 2019
Abstract: Radar-based breast imaging has shown promise as an imaging modality for early-stage
cancer detection, and clinical investigations of two commercial imaging systems are ongoing.
Many imaging algorithms have been proposed, which seek to improve the quality of the reconstructed
microwave image to enhance the potential clinical decision. However, in many cases, the radar-based
imaging algorithms have only been tested in limited numerical or experimental test cases or
with simplifying assumptions such as using one estimate of permittivity for all patient test
cases. In this work, the potential impact of patient-specific permittivity estimation on algorithm
comparison is highlighted using representative experimental breast phantoms. In particular, the case
studies presented help show that the permittivity estimate can impact the conclusions of the
algorithm comparison. Overall, this work suggests that it is important that imaging algorithm
comparisons use realistic test cases with and without breast abnormalities and with reconstruction
permittivity estimation.
Keywords: radar-based breast imaging; microwave imaging; breast cancer
1. Introduction
Radar-based imaging is an emerging modality for breast cancer detection [1]. In recent years,
a number of clinical studies have been published [2–7], including the commercially available MARIA R
system (Micrima Ltd., Bristol, UK). A competing system developed by Microwave Vision Group
(Villebon-sur-Yvette, France) is being used in clinical investigations at the National University of
Ireland Galway (Ireland) [8–10]. The increasing amount of clinical studies of microwave breast
imaging systems is motivating further research into the optimal imaging algorithms and system design
which can maximize the clinical efficacy of radar-based breast imaging [11].
Several imaging reconstruction algorithms for radar-based imaging have been developed
(known as beamformers) and many introductory books, comprehensive reviews and open-source
implementations of these algorithms have been published [12–15]. These beamformers have also
been compared using a variety of test cases with both numerical and experimental breast phantoms
and patient images [16–22]. However, many comparative studies to date have used a limited set of
beamformers [16], simplified numerical models [17,20], idealized artefact removal algorithms [18],
or considered only test situations with abnormalities [21]. A more exhaustive comparative study using
five clinical case studies identified that the Delay-Multiply-and-Sum (DMAS) algorithm achieved
the highest signal-to-clutter ratios (SCRs) for the five clinical case studies, but did not consider the
potential impact of variations in breast dielectric properties between the five patients [22].

J. Imaging 2019, 5, 87
However, it is known that an incorrect estimate of the breast dielectric properties can impair
the image quality of radar-based breast images [23]. In particular, assuming the same estimate of
breast dielectric properties for each test case has been shown to impair the sensitivity of radar-based
breast imaging where tumors are not correctly detected [24]. Several methods to identify the
optimal patient-specific dielectric properties have been proposed [25,26], many of which rely on the
properties of the reconstructed image. Similarly, some preliminary studies have suggested that DMAS
beamformer may “improve” the image quality in cases where no abnormalities exist [22,27]. However,
to date, no study has analyzed the impact of different reconstruction algorithms on patient-specific
permittivity estimation algorithms, nor has the potential benefit of improved imaging algorithms been
examined with patient-specific estimation algorithms.
The goal of this study is to examine the impact of patient-specific permittivity estimation
algorithms on beamformer comparisons in healthy and tumor experimental models. Although
patient-specific permittivity estimation and sophisticated beamformers have been shown to improve
image quality independently in certain cases, the impact of patient-specific permittivity estimation
on beamformer comparative studies is unknown. The remainder of this paper is structured as
follows: Section 2 reviews the beamformers which have been proposed in the literature as well as
patient-specific permittivity estimation algorithms. Section 3 describes the experimental set-up used to
test the algorithms. The results are presented in Section 4 and, finally, Section 5 concludes the paper.
2. Background
Many reviews, books and editorials have been published about microwave imaging and
microwave breast imaging in recent years [1,11–15,28–37] which describe the mathematical background
and existing literature of the method. Broadly speaking, microwave breast imaging algorithms attempt
to infer information about the dielectric contrast of the breast, and correspondingly, information
about any tumors or other abnormalities present. The subject of this work is radar-based imaging,
a qualitative technique which reconstructs an energy map of the backscattered energy within the breast.
As tumor tissues often have the highest dielectric properties [1], large reflections occur at tumor tissue
boundaries. In areas of high dielectric contrast (i.e., tumors), these reflections cohere and result in high
image energy, whereas in areas of lower contrast, these reflections do not cohere and result in lower
image energy.
The Delay-and-Sum (DAS) imaging algorithm was the first proposed for radar-based breast
imaging and has been used in both mono- and multistatic configurations [38,39]. Mathematically, DAS
can be represented as follows:

I (r) = w(r) P(ω ) wa,a (r) Ea,ω (a ) exp jωτa,a (r, ω ) da da dω (1)
Ω AA
where the image intensity, I (r), at the point r is calculated from the scattered energy, Ea,ω (a ), recorded
at antenna a ∈ A while the incident signal, P(ω ), is transmitted from antenna a ∈ A at angular
frequency ω ∈ Ω. The propagation time for a signal from the transmitting antenna, a, to the point of
interest, r, and back to the recording antenna, a , is represented by τa,a (r, ω ). DAS has been used in
most patient imaging studies to date [3–5,40], and most comparative studies on imaging algorithm
performance have used numerical or experimental data only [1,22].
Many extensions to DAS have been proposed which aim to improve the image quality.
The extensions can be broadly classified into three types:
1. calculating a separate weighting factor to reward points where tumors are more likely based either
on epidemiological studies or on characteristics of the scattered signals (w(r) from Equation (1));
2. prioritizing scattered signals collected at certain locations based on the relative locations of the
antennas and points of interest or on the antenna radiation patterns (wa,a (r) in Equation (1));
145
J. Imaging 2019, 5, 87
3. by improving the quality of the input signals prior to imaging through improved artefact removal
algorithms or other noise reduction techniques.
In most cases, these modified beamformers have been tested in limited numerical or experimental
test cases, and most radar-based operational systems have used traditional DAS [1]. In this work,
the DAS beamformer is compared to the Delay-Multiply-and-Sum (DMAS) beamformer which has
been identified as improving image quality in a series of clinical case studies [22]. In the literature,
DMAS is increasingly being used in new experimental studies without comparison to DAS or
evaluation in imaging scenarios without targets [41,42].
The DMAS beamformer was first proposed in [43]. DMAS aims to reduce noise in the scattered
signals and hence, reduce clutter in the image. After synthetic focusing and prior to summation in
Equation (1), the scattered signals are pairwise multiplied, such that:

I (r) = P(ω ) Ea1 ,ω (a1 ) exp jωτa1 ,a1 (r, ω ) Ea2 ,ω (a2 ) exp jωτa2 ,a2 (r, ω ) da1 da1 da2 da2 dω (2)
Ω AA AA
Although this does not increase the amount of independent information, this pairwise
multiplication artificially increases the number of channels in the summation. The DMAS beamformer
assumes that signals with a high degree of coherence should be rewarded by multiplication, whereas
incoherent signals should not increase in energy after multiplication.
A fundamental assumption of radar-based imaging is that knowledge of propagation within the
breast can be used to synthetically focus the recorded electromagnetic energy to points within the
breast. This is represented by the propagation delay, τa,a (r, ω ), in Equation (1). The propagation delay
can be calculated from the dielectric properties of the breast, however, in a realistic imaging scenario,
the dielectric properties of the breast are not known. Thus, most practical implementations assume
a single estimate of the dielectric properties for the entire breast, known as the effective permittivity in
this work.
Several patient-specific permittivity estimation methods have been proposed which are reviewed
in detail in [24]. Patient-specific permittivity estimation methods based on analyzing images
reconstructed within a range of potential permittivity estimates have shown promising results in
experimental and clinical test cases [23,26,44,45]. These patient-specific permittivity estimation
methods have been shown to improve the sensitivity in 110 experimental test cases including diverse
breast phantoms and tumor phantoms [46]. However, in [23,26], the original DAS beamformer was
used, and the impact of other beamformers on the method is unknown.
Evaluation Metrics
To assess the performance of beamformers, it is necessary to identify a set of evaluation metrics
for the radar-based images. In many cases, image quality has been defined using the signal-to-mean
ratio (SMR) and the signal-to-clutter ratio (SCR). Typically, the SMR is calculated by dividing the mean
energy of the tumor area (the area surrounding the maximum intensity of the image) by the mean
energy of the image of the ipsilateral breast, where the tumor area is twice the full-width half-maximum
(FWHM) of the response of highest energy in the image. Similarly, the SCR is calculated by dividing
the maximum energy of the tumor area (i.e., the maximum intensity of the image) by the maximum
energy in the background of the image. Additionally, in images of test scenarios with abnormalities,
the localization error, or the distance between the maximum response in the image and the location of
the abnormality, is used as a measure of quality. Lower localization error means higher image quality.
In this work, the SCR and the FWHM of the images using different beamformers is compared.
Although recent studies have suggested that these metrics may not be sufficient to distinguish images
of modelled healthy and cancerous cases [47], the SCR and the FWHM are still the standard metrics
for image quality quantification.
Additionally, the images are analyzed quantitatively and annotated as containing a tumor or not
according to the criteria described in detail in [24]. Broadly speaking, images where the main response
146
J. Imaging 2019, 5, 87
has a SCR greater than 1.5 dB are annotated as a positive and considered to be detection. Where the
main response is located within the physical extent of the tumor, the image is marked as a true positive,
otherwise, it is considered a false positive.
3. Experimental Methods
In this section, the experimental methods are described. The BRIGID breast and tumor phantom
set was used to assess the performance of the imaging algorithms which is fully presented in [48].
The breast and tumor phantoms are fabricated from a polyurethane mixture combined with graphite
and carbon black powders to control the dielectric properties [49]. Once cured, the mixture is solid,
and can be used to create a modular and stable test platform [48,50].
Breast phantoms were hemispherical with diameter of 14 cm and consisted of three tissue types:
• a hemispherical skin which varied between 1 to 3 mm in thickness with relative permittivity of

ε r ≈ 30 and electrical conductivity of σ ≈ 2 S m−1 at 3 GHz;
• conical glandular structures to model breast lobes radiating from the areola with relative
permittivity of ε r ≈ 45 and electrical conductivity of σ ≈ 2.5 S m−1 at 3 GHz;
• an adipose layer with relative permittivity of ε r ≈ 8 and electrical conductivity of σ ≈ 0.1 S m−1 .
In each breast phantom, a hole was left in which a tumor “plug” could be inserted. Each tumor
phantom was embedded in a plug of adipose tissue, and the tumor phantom had a mean relative
permittivity of ε r ≈ 70 and mean electrical conductivity of σ ≈ 6 S m−1 . The contrast between the
dielectric properties of tumor and glandular tissues in these breast phantoms is similar to the contrast
between healthy and glandular tissues in the literature [1,51–53].
The BRIGID phantom set contains breast phantoms which vary according to the volume of
glandular tissue present within the phantom, known as the volume glandular fraction (VGF). Recent
advances in three-dimensional imaging using breast tomosynthesis have allowed the VGF to be
quantified, such as in a study of 219 women in [54]. This study suggests that breast density in terms of
VGF may be overestimated from the two-dimensional slices used in X-ray mammography, and that
the average VGF of breasts even in BI-RADS Class D maybe be as low as 20%.
In this work, images of four breast phantoms with VGF varying from 0% to 20% are analyzed as
the healthy cases, specifically 0%, 10%, 15% and 20% as shown in Figure 1. These breast phantoms are
representative of the variation in VGF present in the population in the sense that breast phantoms with
VGF covering more than 80% of the population are included [54]. Each of the four breast phantoms
can also be combined with a tumor phantom to model a cancerous case, the location of the tumor
phantom is shown in Figure 1 by the dashed white circle. Five spherical tumor phantoms ranging
in diameter from 5 mm to 20 mm are used, representing plugs 1 to 5 from [48], which are shown in
Figure 1f. In total, there are 20 test cases with tumor phantoms (five tumor phantoms in four breast
phantoms) and four test cases without tumors (four breast phantoms with an adipose tissue in place of
the tumor phantom).
Signals are acquired from the breast phantoms using experimental hardware developed at NUI
Galway, Ireland which is described fully in [23]. Experimental data between 2 GHz and 4 GHz were
acquired using a ZNB40 2-port VNA and ZN-Z84 24-port switching matrix (Rohde & Schwartz GmbH,
Munich, Germany) in the frequency domain. 24 flexible microstrip antennas were used to collect the
multistatic scan. The antennas were first developed and used at McGill University in experimental
and healthy clinical case studies [55,56]. The antennas housed in the radome are shown in Figure 1e.
The antennas were designed to be in contact with a material with relative permittivity close to skin and
were in direct contact with the skin of the phantom in this system. The imaging system was calibrated
to the antenna using open/short/load and through measurements for all channels.
147
J. Imaging 2019, 5, 87
70 70
35 35
Head–Toe (mm)
Head–Toe (mm)
0 0
−35 −35
−70 −70
−70 −35 0 35 70 −70 −35 0 35 70
Left–Right (mm) Left–Right (mm)
(a) (b)
70 70
35 35
Head–Toe (mm)
Head–Toe (mm)
0 0
−35 −35
−70 −70
−70 −35 0 35 70 −70 −35 0 35 70
Left–Right (mm) Left–Right (mm)
(c) (d)
>2/ĜhQ2 UmmV
0
0
0
0 0 0 0 0 00 0
G27iĜ_B;?i UmmV
(e) (f)
Figure 1. The four breast phantoms are shown in (a–d) containing 0%, 10%, 15% and 20% VGF,
respectively. The tumor location is shown by the dashed, white circle. The 24 flexible microstrip
antennas shown housed in the radome are shown in (e) and the 5 tumour phantoms are showin in (f).
148
J. Imaging 2019, 5, 87
A reference scan with a homogeneous breast phantom was used to account for differences between
the antennas. This reference scan was used to calculate a “calibration factor” which was then applied
to neighboring antennas such that the response of each antenna was the same. While effective in these
experimental studies, this calibration is not optimized for use with patients without optimization
to ensure the breast contacts the antennas correctly. Many approaches have been proposed in the
literature to overcome these issues:
• inclusion of a suction system to ensure the breast contacts the radome [57];
• undersizing the radome compared to the reported breast size to improve the fit [56];
• use of a coupling medium to help reduce air gaps between the breast and the antennas [2,56,58];
• automated metrics to ensure the breast is well-situated prior to the scan [58].
After acquisition, all signal processing and imaging was performed in the frequency domain
using MERIT, an open-source toolbox for microwave imaging [15]. First, rotational subtraction was
performed to remove the large skin reflection and other artefacts [23]. In this experimental set-up,
a rotational angle of 36 deg was used corresponding to the angle between neighboring antennas.
Although rotational subtraction has also been used with the MARIA R
system and hundreds of
women [2,58], the optimal angle of rotation and the impact of rotational subtraction on image quality
are not well studied.
Next, an assumed estimate of the relative permittivity of the breast was used to synthetically
focus all signals to a point of interest within the breast [23]. Considering the generic radar-based
beamformer shown in Equation (1), this is discretized and simplified to become the DAS beamformer
in the frequency domain:
8 :
jω 9 # " "$
I (r, εr ) =
∑ ∑ Ea,a ( r ) exp ε r a − r + "a − r" (3)
A A
c0
where I (r, εr ) is the beamformer with the average permittivity of the breast, εr , as a parameter;
the antenna locations are the 24 antenna locations shown in Figure 1e; and the propagation paths are
assumed to be the straight-line paths between the antennas and the points of interest. The DMAS
beamformer is implemented similarly; however, prior to the beamformer being applied, the signals
are pairwise multiplied.
For each test case, a cost function is used to estimate the “quality” of that image. Specifically,
the Absolute Gradient, ΦDMA G , from [23] is applied. For each set of artefact-removed signals
corresponding to a single test case, 50 images are reconstructed: specifically, both DAS and DMAS
images are reconstructed at 25 permittivity values between εr = 8 and εr = 14. For each of the 24 test
cases (5 tumor phantoms in 4 breast phantoms and the 4 breast phantoms without tumor phantoms),
the 50 images per case are analyzed qualitatively and quantitavely below.
4. Results
The results of this paper are presented in a three stages:
1. first, the impact of permittivity estimation in phantoms without abnormalities is discussed in
Section 4.1, considering both beamformers and the four breast phantoms;
2. secondly, the ability of DMAS to compensate for errors in the permittivity estimation process is
considered in Section 4.2;
3. thirdly, the permittivity estimation algorithms are applied to images of the same scene
reconstructed with different beamformers to estimate if the characteristics of the images are
different in Section 4.3.
4.1. Algorithm Performance in Test Cases without Abnormalities

In this section, the impact of permittivity estimation on beamformer selection in breast phantoms
without abnormalities is discussed. First, the images of the four breast phantoms without abnormalities
149
J. Imaging 2019, 5, 87
using both the DAS and the DMAS beamformers are investigated in Figure 2. These images are
reconstructed at ε r = 8.75, which is the optimal permittivity based on maximizing the sensitivity for
all test cases with abnormalities. Figure 2a,c,e,g are reconstructed with DAS, whereas Figure 2b,d,f,h
are reconstructed with DMAS.
The images in Figure 2 highlight the potential challenges in distinguishing scenarios containing
tumor phantoms from scenarios without tumor phantoms. Due to the reflections from other tissue
boundaries within the breast (i.e., between the skin and the adipose tissue not fully removed by artefact
removal and between the glandular structures and the adipose tissue), images of breast phantoms
without tumor phantoms can have characteristics similar to images of tumor phantoms embedded in
breast phantoms.
Considering the least dense breast phantom initially (Figure 2a,b), it can be seen that the DMAS
image shows a distinct response in the middle of the breast whereas the DAS image of the same
scene does not. A similar result is visible in Figure 2c,d where the DMAS image shows a more
distinct response than the DAS image. These images suggest that DMAS may make the specificity of
radar-based imaging worse even in breast phantoms with minimal dielectric contrast in the breast
phantom interior.
However, in the denser breast phantoms, Figure 2e,f, both DAS and DMAS images show responses,
as is expected as the fibroglandular structures have significant dielectric contrast to the adipose
background. Interestingly, in the case of the 15% VGF breast phantom in Figure 2e,f, Where, as the
DAS image highlights a response in the middle of the breast, the DMAS image highlights a response
very close to the chest wall for the same scene. For the densest breast phantom in Figure 2g,h, it can
be seen that the DMAS image shows a response in the same location as the DAS image, but with
higher SCR.
Considering the images selected by using the permittivity estimation method described in [23],
the parameter search algorithm selects ε r = 8 for images of the 15% and 20% breast phantoms for
both the DAS and DMAS images. In both cases, the responses are in the same location (within 5 mm)
and could both be interpreted as containing a response that could be a tumor. However, the DMAS
images have SCR of 2.62 dB and 3.06 dB for the 15% and 20% breast phantoms respectively, compared
to 0.56 dB and 0.29 dB for the DAS images. These are shown quantitively in Table 1.
1 1
6 6
0.8 0.8
3 3
>2/ĜhQ2
>2/ĜhQ2
0.6 0.6
0 0
0.4 0.4
−3 −3
0.2 0.2
−6 −6
0 0
0 0 3 6 0 0 3 6
6`QMiĜ"+F
6`QMiĜ"+F
"+FĜ6`QMi "+FĜ6`QMi
3 3
6 6
−6 −3 0 3 6 −6 −3 0 3 6
G27iĜ_B;?i G27iĜ_B;?i
(a) (b)
Figure 2. Cont.
150
J. Imaging 2019, 5, 87
1 1
6 6
0.8 0.8
3 3
>2/ĜhQ2
>2/ĜhQ2
0.6 0.6
0 0
0.4 0.4
−3 −3
0.2 0.2
−6 −6
0 0
0 0 3 6 0 0 3 6
6`QMiĜ"+F
6`QMiĜ"+F
3 3
6 6
−6 −3 0 3 6 −6 −3 0 3 6
(c) (d)
1 1
6 6
0.8 0.8
3 3
>2/ĜhQ2
>2/ĜhQ2
0.6 0.6
0 0
0.4 0.4
−3 −3
0.2 0.2
−6 −6
0 0
0 0 3 6 0 0 3 6
6`QMiĜ"+F
6`QMiĜ"+F
3 3
6 6
−6 −3 0 3 6 −6 −3 0 3 6
(e) (f)
1 1
6 6
0.8 0.8
3 3
>2/ĜhQ2
>2/ĜhQ2
0.6 0.6
0 0
0.4 0.4
−3 −3
0.2 0.2
−6 −6
0 0
0 0 3 6 0 0 3 6
6`QMiĜ"+F
6`QMiĜ"+F
3 3
6 6
−6 −3 0 3 6 −6 −3 0 3 6
(g) (h)
Figure 2. Images using DAS and DMAS of four breast phantoms with between 0% and 20% VGF.
The images reconstructed using DAS are shown in the leftmost column (a, c and e) in order of increasing
VGF (10%, 15% and 20% respectively). The corresponding image of the same test case reconstructed
using DMAS is shown in the rightmost column (b, d and f). As a response that is very close to the skin
would likely be regarded as an artefact, in the less dense breast phantoms, (a,c) reconstructed with
DAS could be considered to be no detection, whereas the DMAS for the same breast phantom shows a
response which is more in the center of the breast and more likely to be interpreted as a false positive.
All dimensions are in cm, and the three slices shown (coronal, sagittal and axial) are cross-sections at
the location of the maximum amplitude of the image in each case.
151
J. Imaging 2019, 5, 87
For the less dense breast phantoms, the parameter search algorithm would select images
reconstructed at different permittivity values, ε r = 10.5 and ε r = 9 for DAS and DMAS respectively for
the 0% breast phantom and ε r = 8.75 and ε r = 8 for the 10% breast phantom. No clear trend is visible:
for 0% VGF the DAS image selected by the parameter search algorithm has higher SCR compared to
the DMAS image selected by the parameter search algorithm (2.6 dB compared to 1.8 dB), whereas
the reverse is true for 10% (1.6 dB compared to 2.0 dB). However, even in the simple and limited case
studies shown in this section, the choice of beamformer may impact the specificity, suggesting that this
should be evaluated in all beamformer comparison studies.
Table 1. The SCR and FWHM of images of breast phantoms without abnormalities containing 0%
and 10% VGF, reconstructed with DAS and DMAS and a parameter search algorithm. For 0% VGF:
εDAS
r = 10.5 and εDMAS
r = 9 whereas for 10% VGF, εDAS
r = 8.75 and εDMAS
r = 8.
SCR (dB) FWHM (mm)

0% VGF 10% VGF 0% VGF 10% VGF
εDAS
r εDMAS
r εDAS
r εDMAS
r εDAS
r εDMAS
r εDAS
r εDMAS
r
DAS 2.6 1.5 1.6 0.75 10 12 13 7
DMAS 0.7 1.8 2.6 2.0 7 6 9 10
4.2. Effects of Permittivity Estimation on Performance

In this section, the effect of using DMAS instead of DAS are analyzed. The reconstruction
permittivities are chosen using the parameter search algorithm applied to the DAS images.
Additionally, case studies are presented which show how the beamformer comparison can depend on
the permittivity chosen for reconstruction.
First, a case study is shown in Figure 3 where all six images are reconstructed from the same
experimental case but using a reconstruction permittivity estimate of ε r = 11.75 in Figure 3a,b,
ε r = 12.25 in Figure 3c,d, and ε r = 12.75 in Figure 3e,f. Figure 3a,c,e are reconstructed with DAS,
whereas Figure 3b,d,f are reconstructed with DMAS.
Considering the lowest estimate of ε r = 11.75 in Figure 3a,b, the strongest response is correctly
located in the images reconstructed with both DAS and DMAS. The FWHM is similar in both cases
(approximately 8 mm), but the SCR of the image reconstructed with DMAS is 1.08 dB whereas the SCR
of the image reconstructed with DAS is 0.09 dB.
Overestimating the relative permittivity by 5% in Figure 3c,d, the image reconstructed with DAS
does not identify the tumor in the correct location (a false negative) whereas the image reconstructed
with DMAS does correspond with the tumor location, albeit with a lower SCR of 0.42 dB. This result
may indicate that improving the beamforming algorithm can compensate for errors in the relative
permittivity estimates for image reconstruction.
Finally, overestimating the relative permittivity by 10% in Figure 3e,f, neither the image
reconstructed with DAS nor the image reconstructed with DMAS show a response in the correct
location. However, whereas the image reconstructed with DAS would be regarded as a false negative,
the image reconstructed with DMAS does have a response in the image, but in the wrong location.
Considering all 20 cases (five tumor phantoms different in size in four breast phantoms differing
in VGF), 14 are detected using DAS and permittivity estimation. For these 14, the DAS image is
compared to the DMAS image at the same estimate of the permittivity. For 3 of the 14 images (P5 in
20%; and P3 and P4 in 15%), the DMAS image is “sharper” with SCR at least 1 dB higher and FWHM
approximately 2 mm smaller. For a further 6 of the 14 (all from the 0% and 10% phantoms), the images
were similar where the SCR was within 1 dB and the localization was within 1 cm. However, for 5 of
the 14 cases (P1 from 0%; P1 from 10%; P2 and P5 from 15%; and P4 from 20%), the image deteriorated
such that the tumor was not detected: for all five, the location of the maximum image response was
more than 5 cm away from the location of the tumor phantom.
152
J. Imaging 2019, 5, 87
1 1
6 6
0.8 0.8
3 3
>2/ĜhQ2
>2/ĜhQ2
0.6 0.6
0 0
0.4 0.4
−3 −3
0.2 0.2
−6 −6
0 0
0 0 3 6 0 0 3 6
6`QMiĜ"+F
6`QMiĜ"+F
3 3
6 6
−6 −3 0 3 6 −6 −3 0 3 6
(a) (b)
1 1
6 6
0.8 0.8
3 3
>2/ĜhQ2
>2/ĜhQ2
0.6 0.6
0 0
0.4 0.4
−3 −3
0.2 0.2
−6 −6
0 0
0 0 3 6 0 0 3 6
6`QMiĜ"+F
6`QMiĜ"+F
3 3
6 6
−6 −3 0 3 6 −6 −3 0 3 6
(c) (d)
1 1
6 6
0.8 0.8
3 3
>2/ĜhQ2
>2/ĜhQ2
0.6 0.6
0 0
0.4 0.4
−3 −3
0.2 0.2
−6 −6
0 0
0 0 3 6 0 0 3 6
6`QMiĜ"+F
6`QMiĜ"+F
3 3
6 6
−6 −3 0 3 6 −6 −3 0 3 6
(e) (f)
Figure 3. Six images of a 2 cm diameter tumor in the breast phantom with 20% VGF are shown,
reconstructed with the DAS and DMAS beamformers are three different reconstruction permittivities.
The DAS images are shown in the leftmost column in order of increasing permittivity, similarly,
the corresponding DMAS images are shown in the rightmost column. Depending on the reconstruction
permittivity chosen between ε r = 11.75 and ε r = 12.75, both beamformers perform the same and
correctly localize the tumor in (a) and (b), DMAS outperforms DAS in (c) and (d), or neither DAS
nor DMAS localize the tumor in (e) and (f). All dimensions are in cm, and the three slices shown
(coronal, sagittal and axial) are cross-sections at the location of the maximum amplitude of the image
in each case.
153
J. Imaging 2019, 5, 87
For the 6 of 20 cases where DAS and permittivity estimation did not detect the tumor, in 2 cases,
the DMAS image was very similar to the DAS image but with less clutter. In 3 more cases, the DMAS
image would also not have located the tumor in the right location, but the main response in the image
was in a different location when compared to the DAS image. In only 1 case, P5 in 10% phantom, did
the DMAS image reconstructed at the same permittivity as the DAS image correctly identify the tumor
where the DAS image would not have.
4.3. Parameter Search Performance Using both Beamformers

Finally, in this section, images selected by DMAS and the permittivity estimation algorithm
are compared to the DAS image at the same permittivity as the DMAS images, and to the DAS
images selected by the permittivity estimation algorithm. Overall, the sensitivity from the DMAS
images with permittivity estimation was 10 out of 20 cases compared to the 14 out of 20 for DAS with
permittivity estimation.
For the 20 breast and tumor phantom combinations, in 9 cases the permittivity estimation
algorithm selected a DMAS image reconstructed within Δε r = 0.5 of the DAS image. Of these
9 cases, 7 of the DAS images correctly identified the tumor in the correct location where as 6 of the
DMAS images did. The majority of these nine cases were in the lower density phantoms: 6 of the
9 were images of phantoms with 0% and 10% VGF. This is expected, as images are less sensitive to the
reconstruction permittivity at lower VGF.
For the remaining 11 cases where the reconstruction permittivity selected by the parameter search
algorithm from the DMAS image differed from the DAS image, 7 of the 11 DAS images correctly
identified the tumor whereas only 3 of the 11 DMAS images did. Interestingly, of the four images where
DAS with permittivity estimation outperformed DMAS with permittivity estimation, two (P1 and P5)
were in the breast phantom with 0% VGF.
5. Conclusions
In this work, the potential impact of permittivity estimation on beamformer comparative study
design is highlighted. Although many comparative studies of various beamformers have been
published in the literature, most use the simplifying assumption that the same estimate of the
permittivity can be used for all test cases, although this is known to impair the image quality. Similarly,
the majority include only test cases with tumors, and do not compare the beamformers in test cases
without tumors.
The original DAS beamformer is compared to the DMAS beamformer, which was identified as a
promising beamformer from a recent comparative study using five clinical case studies. The results
in this paper suggest that DMAS may reconstruct an image of higher quality (defined primarily in
terms of SCR) than the original DAS, however, this may result in more false positives such as in the
less dense breast phantoms in this paper.
The importance of including realistic patient-specific estimation when comparing beamformers is
also shown. A case study showing how the conclusion of a beamformer comparison study may vary
depending on the permittivity estimate is presented. In this case study, as the estimate varies by 5%, it
could be concluded that both beamformers perform the same, or that DMAS localizes the tumor where
DAS does not, or that neither DAS nor DMAS localizes the tumor. Hence it is important to consider
realistic and varied test cases when comparing beamformer performance.
Future work in this area should include the development of metrics to quantify the image
quality which are useful to distinguish images of healthy and abnormal breasts. Future studies
should consider the effects of permittivity estimation when designing and comparing beamformer
performance experimentally and clinically.
Author Contributions: D.O. and B.L.O. designed and performed the experiments; D.O. analysed the data
and wrote the paper; and M.G., E.J. and M.O. supervised the work, including experimental design, analysis
and writing.
154
J. Imaging 2019, 5, 87
Funding: Irish Research Council (GOIPG/2014/987 and EPSPD/2019/200); Science Foundation Ireland
(12/IP/1523); European Research Council (grant agreement number 637780).
Acknowledgments: This work was supported by the Irish Research Council (GOIPG/2014/987 and
EPSPD/2019/200), Science Foundation Ireland (12/IP/1523) and the European Research Council (grant agreement
number 637780).
References
1. O’Loughlin, D.; O’Halloran, M.; Moloney, B.M.; Glavin, M.; Jones, E.; Elahi, M.A. Microwave Breast Imaging:
Clinical Advances and Remaining Challenges. Trans. Biomed. Eng. 2018, 65, 2580–2590. [CrossRef] [PubMed]
2. Preece, A.W.; Craddock, I.J.; Shere, M.; Jones, L.; Winton, H.L. MARIA M4: Clinical Evaluation of a Prototype
Ultrawideband Radar Scanner for Breast Cancer Detection. J. Med. Imaging 2016, 3, 033502. [CrossRef]
[PubMed]
3. Porter, E.; Coates, M.; Popović, M. An Early Clinical Study of Time-Domain Microwave Radar for Breast
Health Monitoring. IEEE Trans. Biomed. Eng. 2016, 63, 530–539. [CrossRef] [PubMed]
4. Song, H.; Sasada, S.; Kadoya, T.; Okada, M.; Arihiro, K.; Xiao, X.; Kikkawa, T. Detectability of Breast Tumor
by a Hand-Held Impulse-Radar Detector: Performance Evaluation and Pilot Clinical Study. Sci. Rep. 2017, 7.
[CrossRef] [PubMed]
5. Yang, F.; Sun, L.; Hu, Z.; Wang, H.; Pan, D.; Wu, R.; Zhang, X.; Chen, Y.; Zhang, Q. A Large-Scale Clinical Trial
of Radar-Based Microwave Breast Imaging for Asian Women: Phase I. In Proceedings of the International
Symposium on Antennas and Propagation (APSURSI), San Diego, CA, USA, 9–14 July 2017; IEEE: San Diego,
CA, USA, 2017; pp. 781–783. [CrossRef]
6. Song, H.; Sasada, S.; Masumoto, N.; Kadoya, T.; Shiroma, N.; Orita, M.; Arihiro, K.; Okada, M.; Kikkawa, T.
Detectability of Breast Tumors in Excised Breast Tissues of Total Mastectomy by IR-UWB-Radar-Based Breast
Cancer Imaging. IEEE Trans. Biomed. Eng. 2018, 66, 2296–2305. [CrossRef] [PubMed]
7. Wörtge, D.; Moll, J.; Krozer, V.; Bazrafshan, B.; Hübner, F.; Park, C.; Vogl, T. Comparison of X-ray
Mammography and Planar UWB Microwave Imaging of the Breast: First Results from a Patient Study.
Diagnostics 2018, 8, 54. [CrossRef]
8. Fasoula, A.; Anwar, S.; Toutain, Y.; Duchesne, L. Microwave Vision: From RF Safety to Medical Imaging.
In Proceedings of the 11th European Conference on Antennas and Propagation (EuCAP), Paris, France,
19–24 March 2017; IEEE: Paris, France, 2017; pp. 2746–2750. [CrossRef]
9. Fasoula, A.; Duchesne, L.; Gil Cano, J.; Lawrence, P.; Robin, G.; Bernard, J.G. On-Site Validation of a
Microwave Breast Imaging System, before First Patient Study. Diagnostics 2018, 8, 53. [CrossRef]
10. Duchesne, L.; Fasoula, A.; Kaverine, E.; Robin, G.; Bernard, J.G. Wavelia Microwave Breast Imaging:
Identification and Mitigation of Possible Sources of Measurement Uncertainty. In Proceedings of the 13th
European Conference on Antennas and Propagation (EuCAP), Kraków, Poland, 31 March–5 April 2019.
11. Bolomey, J.C. Crossed Viewpoints on Microwave-Based Imaging for Medical Diagnosis: From Genesis to
Earliest Clinical Outcomes. In The World of Applied Electromagnetics; Lakhtakia, A., Furse, C.M., Eds.; Springer
International Publishing: Cham, Switzerland, 2018; pp. 369–414. [CrossRef]
12. Nikolova, N.K. Microwave Biomedical Imaging. In Wiley Encyclopedia of Electric and Electronics Engineering;
John Wiley & Sons, Inc: New York, NY, USA, 2014 ; pp. 1–22.
13. Conceição, R.C.; Mohr, J.J.; O’Halloran, M. (Eds.) An Introduction to Microwave Imaging for Breast Cancer
Detection, 1st ed.; Biological and Medical Physics, Biomedical Engineering; Springer International Publishing:
Cham, Switzerland, 2016.
14. Nikolova, N.K. (Ed.) Introduction to Microwave Imaging; EuMA High Frequency Technologies Series;
Cambridge University Press: Cambridge, UK, 2017. [CrossRef]
15. O’Loughlin, D.; Elahi, M.A.; Porter, E.; Shahzad, A.; Oliveira, B.L.; Glavin, M.; Jones, E.; O’Halloran, M.
Open-Source Software for Microwave Radar-Based Image Reconstruction. In Proceedings of the 12th
European Conference on Antennas and Propagation (EuCAP), London, UK, 9–13 April 2018.
16. Klemm, M.; Craddock, I.J.; Leendertz, J.A.; Preece, A.W.; Benjamin, R. Radar-Based Breast Cancer Detection
Using a Hemispherical Antenna Array—Experimental Results. IEEE Trans. Antennas Propag. 2009, 57,
1692–1704. [CrossRef]
155
J. Imaging 2019, 5, 87
17. Xie, Y.; Guo, B.; Xu, L.; Li, J.; Stoica, P. Multistatic Adaptive Microwave Imaging for Early Breast Cancer
Detection. IEEE Trans. Biomed. Eng. 2006, 53, 1647–1657. [CrossRef]
18. O’Halloran, M.; Glavin, M.; Jones, E. Effects of Fibroglandular Tissue Distribution on Data-Independent
Beamforming Algorithms. Prog. Electromagn. Res. 2009, 97, 141–158. [CrossRef]
19. Byrne, D.; O’Halloran, M.; Glavin, M.; Jones, E. Data Independent Radar Beamforming Algorithms for
Breast Cancer Detection. Prog. Electromagn. Res. 2010, 107, 331–348. [CrossRef]
20. Moll, J.; Kexel, C.; Krozer, V. A Comparison of Beamforming Methods for Microwave Breast Cancer
Detection in Homogeneous and Heterogeneous Tissue. In Proceedings of the Microwave Conference
(EuMC), Nuremberg, Germany, 6–10 October 2013; pp. 1839–1842.
21. Elahi, M.A.; Lavoie, B.R.; Porter, E.; Glavin, M.; Jones, E.; Fear, E.C.; O’Halloran, M. Comparison of
Radar-Based Microwave Imaging Algorithms Applied to Experimental Breast Phantoms. In Proceedings
of the 32nd International Union of Radio Science (URSI) General Assembly and Scientific Symposium,
Montreal, QC, Canada, 19–26 August 2017; Union of Radio Science (URSI): Montréal, QC, Canada, 2017.
22. Elahi, M.A.; O’Loughlin, D.; Lavoie, B.R.; Glavin, M.; Jones, E.; Fear, E.C.; O’Halloran, M. Evaluation of
Image Reconstruction Algorithms for Confocal Microwave Imaging: Application to Patient Data. Sensors
2018, 18, 1678. [CrossRef] [PubMed]
23. O’Loughlin, D.; Oliveira, B.L.; Elahi, M.A.; Glavin, M.; Jones, E.; Popović, M.; O’Halloran, M. Parameter
Search Algorithms for Microwave Radar-Based Breast Imaging: Focal Quality Metrics as Fitness Functions.
Sensors 2017, 17, 2823. [CrossRef] [PubMed]
24. O’Loughlin, D.; Oliveira, B.L.; Santorelli, A.; Porter, E.; Glavin, M.; Jones, E.; Popović, M.; O’Halloran, M.
Sensitivity and Specificity Estimation Using Patient-Specific Microwave Imaging in Diverse Experimental
Breast Phantoms. IEEE Trans. Med. Imaging 2019, 38, 303–311. [CrossRef]
25. O’Loughlin, D.; Krewer, F.; Glavin, M.; Jones, E.; O’Halloran, M. Focal Quality Metrics for the Objective
Evaluation of Confocal Microwave Images. Int. J. Microw. Wirel. Technol. 2017, 9, 1365–1372. [CrossRef]
26. Lavoie, B.R.; Okoniewski, M.; Fear, E.C. Estimating the Effective Permittivity for Reconstructing Accurate
Microwave-Radar Images. PLoS ONE 2016, 11. [CrossRef]
27. O’Loughlin, D.; Glavin, M.; Jones, E.; O’Halloran, M. Evaluation of Experimental Microwave Radar-Based
Images: Evaluation Criteria. In Proceedings of the Antennas and Propagation Society International
Symposium (APSURSI), Boston, MA, USA, 8–13 July 2018; IEEE: Boston, MA, USA, 2018.
28. Iskander, M.F.; Durney, C.H. Electromagnetic Techniques for Medical Diagnosis: A Review. Proc. IEEE 1980,
68, 126–132. [CrossRef]
29. Larsen, L.; Jacobi, J. Microwaves Offer Promise as Imaging Modality. Diagn. Imaging 1982, 11, 44–47.
30. Fear, E.C.; Hagness, S.C.; Meaney, P.M.; Okoniewski, M.; Stuchly, M.A. Enhancing Breast Tumor Detection
with Near-Field Imaging. IEEE Microw. Mag. 2002, 3, 48–56. [CrossRef]
31. Fear, E.C. Microwave Imaging of the Breast. Technol. Cancer Res. Treat. 2005, 4, 69–82. [CrossRef]
32. Nikolova, N.K. Microwave Imaging for Breast Cancer. IEEE Microw. Mag. 2011, 12, 78–94. [CrossRef]
33. Meaney, P.M. Microwave Imaging: Perception and Reality. Expert Rev. Med. Devices 2013, 10, 581–583.
[CrossRef] [PubMed]
34. Crocco, L. Microwaves for Medical Imaging: Some Possible Pathways for an Accelerated Progress towards
Clinical Practice. New Horizons Transl. Med. 2015, 2, 62. [CrossRef]
35. Chandra, R.; Zhou, H.; Balasingham, I.; Narayanan, R.M. On the Opportunities and Challenges in Microwave
Medical Sensing and Imaging. IEEE Trans. Biomed. Eng. 2015, 62, 1667–1682. [CrossRef] [PubMed]
36. Kwon, S.; Lee, S. Recent Advances in Microwave Imaging for Breast Cancer Detection. Int. J. Biomed. Imaging
2016, 2016, 1–25. [CrossRef] [PubMed]
37. Modiri, A.; Goudreau, S.; Rahimi, A.; Kiasaleh, K. Review of Breast Screening: Towards Clinical Realization
of Microwave Imaging. Med. Phys. 2017, 44, e446–e458. [CrossRef] [PubMed]
38. Li, X.; Hagness, S.C. A Confocal Microwave Imaging Algorithm for Breast Cancer Detection. Microw. Wirel.
Components Lett. IEEE 2001, 11, 130–132.
39. Nilavalan, R.; Gbedemah, A.; Craddock, I.J.; Li, X.; Hagness, S.C. Numerical Investigation of Breast Tumour
Detection Using Multi-Static Radar. Electron. Lett. 2003, 39, 1787–1789. [CrossRef]
156
J. Imaging 2019, 5, 87
40. Fear, E.C.; Bourqui, J.; Curtis, C.F.; Mew, D.; Docktor, B.; Romano, C. Microwave Breast Imaging With a
Monostatic Radar-Based System: A Study of Application to Patients. IEEE Trans. Microw. Theory Tech. 2013,
61, 2119–2128. [CrossRef]
41. Shao, W.; Edalati, A.; McCollough, T.R.; McCollough, W.J. A Time-Domain Measurement System for UWB
Microwave Imaging. IEEE Trans. Microw. Theory Tech. 2018, 66, 2265–2275. [CrossRef]
42. Islam, M.; Samsuzzaman, M.; Islam, M.; Kibria, S. Experimental Breast Phantom Imaging with
Metamaterial-Inspired Nine-Antenna Sensor Array. Sensors 2018, 18, 4427. [CrossRef] [PubMed]
43. Lim, H.B.; Nhung, N.T.T.; Li, E.P.; Thang, N.D. Confocal Microwave Imaging for Breast Cancer Detection:
Delay-Multiply-and-Sum Image Reconstruction Algorithm. IEEE Trans. Biomed. Eng. 2008, 55, 1697–1704.
[CrossRef] [PubMed]
44. O’Loughlin, D.; Krewer, F.; Glavin, M.; Jones, E.; O’Halloran, M. Estimating Average Dielectric Properties for
Microwave Breast Imaging Using Focal Quality Metrics. In Proceedings of the 10th European Conference on
Antennas and Propagation (EuCAP), Davos, Switzerland, 10–15 April 2016; IEEE: Davos, Switzerland, 2016;
pp. 1–5. [CrossRef]
45. O’Loughlin, D.; Glavin, M.; Jones, E.; O’Halloran, M. Optimisation of Confocal Microwave Breast Images
Using Image Focal Metrics. In Proceedings of the 22nd Bioengineering in Ireland (BINI), Galway, Ireland,
22–23 January 2016; Royal Academy of Medicine in Ireland: Galway, Ireland, 2016; p. 39.
46. O’Loughlin, D.; Oliveira, B.L.; Glavin, M.; Jones, E.; O’Halloran, M. Advantages and Disadvantages of
Parameter Search Algorithms for Permittivity Estimation for Microwave Breast Imaging. In Proceedings of
the 13th European Conference on Antennas and Propagation (EuCAP), Krakow, Poland, 31 March–5 April
2019; IEEE: Kraków, Poland, 2019.
47. O’Loughlin, D.; Oliveira, B.L.; Glavin, M.; Jones, E.; O’Halloran, M. Effects of Interpatient Variance on
Microwave Breast Images: Experimental Evaluation. In Proceedings of the 40th Annual International
Conference of the Engineering in Medicine and Biology Society (EMBC), Honolulu, HI, USA, 18–21 July
2018; IEEE: Honolulu, HI, USA, 2018.
48. Oliveira, B.L.; O’Loughlin, D.; O’Halloran, M.; Porter, E.; Glavin, M.; Jones, E. Microwave Breast Imaging:
Experimental Tumour Phantoms for the Evaluation of New Breast Cancer Diagnosis Systems. Biomed. Phys.
Eng. Express 2018, 4, 025036. [CrossRef]
49. Garrett, J.; Fear, E.C. A New Breast Phantom With a Durable Skin Layer for Microwave Breast Imaging.
50. Santorelli, A.; Laforest, O.; Porter, E.; Popović, M. Image Classification for a Time-Domain Microwave Radar
System: Experiments with Stable Modular Breast Phantoms. In Proceedings of the 9th European Conference
on Antennas and Propagation (EuCAP), Lisbon, Portugal, 13–17 April 2015; IEEE: Lisbon, Portugal, 2015;
pp. 1–5.
51. Lazebnik, M.; McCartney, L.; Popović, D.; Watkins, C.B.; Lindstrom, M.J.; Harter, J.; Sewall, S.; Magliocco, A.;
Booske, J.H.; Okoniewski, M.; et al. A Large-Scale Study of the Ultrawideband Microwave Dielectric
Properties of Normal Breast Tissue Obtained from Reduction Surgeries. Phys. Med. Biol. 2007, 52, 2637–2656.
[CrossRef]
52. Lazebnik, M.; Popović, D.; McCartney, L.; Watkins, C.B.; Lindstrom, M.J.; Harter, J.; Sewall, S.; Ogilvie, T.;
Magliocco, A.; Breslin, T.M.; et al. A Large-Scale Study of the Ultrawideband Microwave Dielectric Properties
of Normal, Benign and Malignant Breast Tissues Obtained from Cancer Surgeries. Phys. Med. Biol. 2007,
52, 6093–6115. [CrossRef]
53. Sugitani, T.; Kubota, S.; Kuroki, S.; Sogo, K.; Arihiro, K.; Okada, M.; Kadoya, T.; Hide, M.; Oda, M.;
Kikkawa, T. Complex Permittivities of Breast Tumor Tissues Obtained from Cancer Surgeries. Appl. Phys.
Lett. 2014, 104, 253702. [CrossRef]
54. Huang, S.Y.; Boone, J.M.; Yang, K.; Packard, N.J.; McKenney, S.E.; Prionas, N.D.; Lindfors, K.K.; Yaffe, M.J.
The Characterization of Breast Anatomical Metrics Using Dedicated Breast CT. Med. Phys. 2011, 38,
2180–2191. [CrossRef]
55. Bahramiabarghouei, H.; Porter, E.; Santorelli, A.; Gosselin, B.; Popović, M.; Rusch, L.A. Flexible 16 Antenna
Array for Microwave Breast Cancer Detection. IEEE Trans. Biomed. Eng. 2015, 62, 2516–2525. [CrossRef]
56. Porter, E.; Bahrami, H.; Santorelli, A.; Gosselin, B.; Rusch, L.A.; Popović, M. A Wearable Microwave Antenna
Array for Time-Domain Breast Tumor Screening. IEEE Trans. Med. Imaging 2016, 35, 1501–1509. [CrossRef]
157
J. Imaging 2019, 5, 87
57. Kuwahara, Y. Microwave Imaging for Early Breast Cancer Detection. In New Perspectives in Breast Imaging;
Malik, A.M., Ed.; InTechOpen: London, UK, 2017. [CrossRef]
58. Shere, M.; Lyburn, I.; Sidebottom, R.; Massey, H.; Gillett, C.; Jones, L. MARIA
R
M5: A Multicentre Clinical
Study to Evaluate the Ability of the Micrima Radio-Wave Radar Breast Imaging System (MARIA R
) to Detect
Lesions in the Symptomatic Breast. Eur. J. Radiol. 2019, 116, 61–67. [CrossRef] [PubMed]
158
MDPI
St. Alban-Anlage 66
4052 Basel
Switzerland
Tel. +41 61 683 77 34
Fax +41 61 302 89 18
www.mdpi.com
Journal of Imaging Editorial Office

E-mail: [email protected]
www.mdpi.com/journal/jimaging
MDPI
St. Alban-Anlage 66
4052 Basel
Switzerland
Tel: +41 61 683 77 34
Fax: +41 61 302 89 18
www.mdpi.com ISBN 978-3-03921-951-3

Microwave Imaging and Electromagnetic Inverse Scattering Problems

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Microwave Imaging and Electromagnetic Inverse Scattering Problems

Uploaded by

Copyright:

Available Formats

Microwave

Special Issue Editors

MDPI • Basel • Beijing • Wuhan • Barcelona • Belgrade

ISBN 978-3-03921-950-6 (Pbk)

About the Special Issue Editors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii

Preface to ”Microwave Imaging and Electromagnetic Inverse Scattering Problems” . . . . . . ix

Alessandro Fedeli, Manuela Maffongelli, Ricardo Monleone, Claudio Pagnamenta,

Yu Zhong and Kuiwen Xu

Marco Salucci, Lorenzo Poli and Giacomo Oliveri

Tushar Rajvanshi, Angela Dell’Aversano, and Raffaele Solimene

Martina T. Bevacqua and Roberta Palmeri

Loreto Di Donato, David Mascali, Andrea F. Morabito and Gino Sorbello

Cameron Kaye, Ian Jeffrey and Joe LoVetri

Declan O’Loughlin, Bárbara L. Oliveira, Martin Glavin, Edward Jones and

Loreto Di Donato, Andrea F. Morabito

Received: 1 October 2018; Accepted: 22 November 2018; Published: 26 November 2018

Keywords: microwave imaging; tomography; inverse problems

J. Imaging 2018, 4, 139; doi:10.3390/jimaging4120139 1 www.mdpi.com/journal/jimaging

2. Measurement Setup and Inversion Procedure

Figure 1. Block diagram of the microwave tomographic system architecture. USB—universal

Figure 4. Flow chart of the inversion algorithm. DAS—delay-and-sum; FH-INLW—frequency-hopping

3. Preliminary Experimental Results

Figure 5. Photograph of the experimental set up.

Figure 7. Experimental results. Qualitative indicator map.

Received: 31 December 2018; Accepted: 31 January 2019; Published: 8 February 2019

J. Imaging 2019, 5, 27; doi:10.3390/jimaging5020027 10 www.mdpi.com/journal/jimaging

Figure 1. Schematics of inverse scattering problems.

2. Inversion with CIE-I

2.1. 3-D Modeling

where u, v = 1, 2, 3 represent the x-, y- and z-components of a vector, respectively, and ι1 = x, ι2 = y,

2.2. Objective Function for Inversions

With this deﬁnition, we could write the mismatch of Equation (5) as

2.3. Sketch of the Inversion Method

2.4. Updating Contrast and Contrast Sources with Different β

Figure 2. The values of |(GD;xx

Figure 3. The physical setup for the numerical tests.

(a) The coated cube in Example 1.

Figure 4. Coated cube used in Example 1.

Figure 6. Errors of reconstructions obtained by different inversion methods for Example 1.

Figure 7. True permittivity proﬁle (real part) for Example 2.

Figure 8. Reconstruction results of inversions with LSIE for Example 2.

Figure 9. Errors of reconstructions obtained by different inversion methods for Example 2.

Figure 11. True permittivity proﬁle (real part) for Example 3.

Figure 12. Reconstruction results of inversions with LSIE for Example 3.

CIE-I Contraction integral equation for inversion

Received: 3 December 2018; Accepted: 8 January 2019; Published: 15 January 2019

J. Imaging 2019, 5, 19; doi:10.3390/jimaging5010019 27 www.mdpi.com/journal/jimaging

is used to successively probe the unaccessible domain D ( Figure 1).

Jv (r) = τ (r ) ∑ Ev,p (r) u p (5)

within the investigation domain (r ∈ D ).

of sub-domains D (o) ∈ D , o = 1, ..., O, occupied by scatterers (i.e., dielectric discontinuities of the

3D-CSI BCS-Based Problem Formulation—Starting from the measurement of E v,p (v =

R[τ(r)] [Arbitrary Unit]

Figure 2. Sensitivity Analysis (K = 1, O = 1, τ = 1.0, SNR ∈ [5, 50] [dB])—Actual contrast

K=O=1, τ=1.0 − SNR=50 [dB] K=O=1, τ=1.0 − SNR=20 [dB]

R[τ(r)] [Arbitrary Unit]

R[τ(r)] [Arbitrary Unit]

R[τ(r)] [Arbitrary Unit]

Figure 3. Numerical Assessment (K = 1, O = 1, τ = 1.0) - MT-BCS reconstructions when processing the

Table 1. Numerical Assessment (K = 1, O = 1, τ = 1.0, W = 100, MT-BCS)—Total error statistics.

K=O=1 K=O=1 K=O=1

ξint [Arbitrary Unit] (×10 )

(a) (b) (c)

K=O=1, τ=4.0 K=O=1, τ=4.0 − SNR=10 [dB]

φ[k, ro (s), r] − φ[k, ro (s), r ] = −w[k, ro (s), r, r ] · (r − r ) (12)

w[k, ro (s), r, r ] = w[k, ro (s), r, r] = −∇x φ[k, ro (s), x]|x=r (14)