Abbasi, Q. H., El Sallabi, H., Chopra, N., Yang, K., Qaraqe, K.

A. and Alomainy, A. (2016) Terahertz channel characterization inside the

human skin for nano-scale body-centric networks. IEEE Transactions on
Terahertz Science and Technology, 6(3), pp. 427-434.
(doi:10.1109/TTHZ.2016.2542213)

This is the author’s final accepted version.

There may be differences between this version and the published version.
You are advised to consult the publisher’s version if you wish to cite from
it.

http://eprints.gla.ac.uk/141057/

Deposited on: 03 July 2017

Enlighten – Research publications by members of the University of Glasgow

http://eprints.gla.ac.uk33640
SUBMITTED TO IEEE TRANSACTIONS ON TERAHERTZ SCIENCE AND TECHNOLOGY 1

Terahertz Channel Characterisation Inside the

Human Skin for Nano-scale Body-Centric Networks
Qammer H. Abbasi, Senior Member, IEEE, Hassan El Sallabi, Member, IEEE, Nishtha Chopra, Student
Member, IEEE, Ke Yang, Student Member, IEEE, Khalid Qaraqe, Senior Member, IEEE, Akram Alomainy, Senior
Member, IEEE

Abstract—This paper focuses on the development of novel radio environmental, industrial and military [2], which can also be
channel model inside the human skin at the terahertz range, extended into other fields like consumer electronics, life style
which will enable the interaction among potential nano-machines and home appliances. Its application in medical diagnostics
operating in the inter cellular areas of the human skin. Thorough and treatment has a quite bright future, because of its ability
studies are performed on the attenuation of electromagnetic waves to access small and delicate body sites non invasively, where
inside the human skin, while taking into account the frequency
of operation, distance between the nano-machines and number
conventional medical devices fall short [3]. Electromagnetic
of sweat ducts. A novel channel model is presented for communi- based communications, handled in the terahertz band, are
cation of nano-machines inside the human skin and its validation considered a viable technique for supporting data exchange
is performed by varying the aforementioned parameters with a in the nano-machines [4]. While almost all other portions of
reasonable accuracy. The statistics of error prediction between spectrum are already deployed in the medical applications, the
simulated and modeled data are: mean (µ) = 0.6 dB and properties of Terahertz (THz) band open a new era and interest,
standard deviation (σ) = 0.4 dB, which indicates the high because of its non ionization hazards for biological tissues and
accuracy of prediction model as compared with measurement less susceptibility to some of propagation effects (i.e., Rayleigh
data from simulation. In addition, the results of proposed channel scattering etc) [5].
model are compared with Terahertz Time Domain Spectroscopy
based measurement of skin sample and the statistics of error Body-centric wireless communication (BCWC) has been ex-
prediction in this case are: µ = 2.10 dB and σ = 6.23 dB, tensively studied in the past for a range of frequencies [6], [7],
which also validates the accuracy of proposed model. Results however the need to reduce the size of the devices makes nano-
in this paper highlight the issues and related challenges while scale technologies attractive for future applications. This opens
characterizing the communication in such a medium, thus paving up opportunities of applying nano-devices made of the novel
the way towards novel research activities devoted to the design materials, like carbon nano tubes (CNT), graphene, [8] which
and the optimization of advanced applications in the health-care operate at THz frequencies and probably inside human bodies.
domain. Since, terahertz radiation can detect the fine variations in the
Keywords—Nanoscale communications, Body Area Nano Net- water contents and other biomaterial tissues as the molecular
work, channel characterisation, Internet of Nano-things resonances are existing and can be detected in this frequency
range. Consequently, one of the emerging areas of research is
analysing the propagation of terahertz electromagnetic waves
I. I NTRODUCTION
through the skin tissues to develop a diagnostic tools for early
With the growing demand of nanotechnology in all the detection and treatment of abnormalities with the skin tissues
domains of science and engineering, its gaining enormous as a sign of skin cancer [9]. Several studies on the applicability
attention from the various disciplines and domains. The con- of THz communication for biomedical applications are per-
nectivity of the nano-devices to conduct complex tasks lead to formed in [10]–[13]. The optical parameters of human tissues
the proposal of the nano-networks, followed by the concept of up to 2.5 THz have been empirically characterized in [10],
nano-communication [1]. Main applications of nano-netwroks while the possibility of applying EM waves in nano-networks
are categorised as, but are are not limited to biomedical, is presented in [13]. Jornet el al. [12] studied the THz channel
model in the air by varying water vapor concentration and also
This publication was made possible by NPRP grant # 7-125-2-061 from proposed a new physical- layer aware medium access control
the Qatar National Research Fund (a member of Qatar Foundation). The
statements made herein are solely the responsibility of the authors. The HPC (MAC) protocol [14]. The characteristics of electromagnetic
resources used in this work were provided by the IT Research Computing waves propagating inside human body at terahertz frequencies
group in Texas A & M University at Qatar, which is funded by Qatar have been studied in [15], which shows that path loss has
Foundation for Education, Science and Community Development. a relation with the dielectric loss of human tissues. In [16],
Qammer, Khalid and Hassan are with the Dep. of Electrical and
Computer Engineering, Texas A & M University at Qatar. Qammer is
the electrical properties (conductivity and permittivity values)
also with Queen Mary University of London and UET, Lahore.; e-mail: of each layer have been investigated depending on the water
{qammer.abbasi;hassan.el sallabi,k.qaraqe}@tamu.edu. content of each layer.
Nishtha Chopra, Ke Yang and A. Alomainy are with Anten-
nas & Electromagnetics School of Electronic Engineering and Com- Though there are some initial studies presented in literature
puter Science Queen Mary University of London, London; e-mail: about THz channel modeling, but to the best of authors’
{n.chopra,ke.yang;a.alomainy}@qmul.ac.uk. knowledge, there are no explicit investigation performed about
 (a) Skin model with one sweat duct. (b) Skin model with two sweat ducts. (c) Skin model with three sweat ducts.

Fig. 1. Numerical skin model (based on CST Microwave StudioTM ) representing the layers and their thickness, while including different number of sweat
ducts (dimensions for each layer are shown as well [17]).

the EM channel modeling inside the skin at THz, while taking al. in [19]. In this study, the sweat duct was modeled in the
into account the number of sweat ducts, frequency and distance epidermis layer [19] by a helix [17] as shown in Fig. 1 with
dependant characteristics. In this paper, a novel channel model a height of 265 µm and diameter equals to 40 µm. As, only
is proposed for inside the skin, which takes into account all when the sweat duct contains sweat, containing 99% water and
of the aforementioned parameters and the blind testing is 1% salt and amino acid [20], it can be regarded as a conductor;
performed on the proposed model for validating the analytical and hence the permittivity of water at THz band was used to
results with the simulated data. In addition, the proposed model model sweat duct. In real skin tissues, the thickness of each
is also validated by comparing it with the measurement results layer is differed according to the position all over the body,
of skin sample using Terahertz Time Domain Spectroscopy ranging from very thin, when covering sensitive areas of the
(THz TDS). body, such as eyelids, to very thick when covering the tough
The rest of paper is organized as follows. Section II high- areas, such as the hand palm. The skin tissue dimensions in the
lights the details of skin model used in this study. Section model were taken to be average between the skin dimensions of
III details about the proposed channel model, in addition to the sensitive areas and the tough areas. Two dipoles optimised
its validation by blind testing and THz TDS data and Finally, to ensure that the impedance matching is better than −10 dB
conclusions are drawn in section IV. (modeled as nano antennas working at THz band) were used
as transmit (Tx) and receive (Rx) antennas, where one was
II. S IMULATION AND M EASUREMENT S ETTINGS placed in the epidermis and other was located in the dermis
layer. Two discrete ports were applied to the dipoles and the
A. Numerical Human Skin Model tranmission coefficient, S21, was recorded from 0.8 - 1.2 THz.
The skin is a complex heterogeneous and anisotropic Fig. 1 represents the skin layer model studied in this paper
medium, where the small parts, like blood vessel and pigment with different spacing between the Tx and Rx by varying
content, are spatially distributed in depth [18]. In this study, number of sweat ducts. The simulations are performed in CST
the skin tissues were represented by three layer model (i.e., Microwave StudioTM and due to large simulation domain and
stratum corneum (SC), epidermis, and dermis. The dimensions large number of mesh cells for the human skin, simulations
are shown in Fig. 1) and the roughness of boundaries between were run on a high performance computing cluster named
the SC and epidermis was considered in the skin model as ”RAAD”. RAAD is a 42+ TFLOP system with blade-based
the order of magnitude for THz wavelength and roughness Linux HPC cluster containing upwards of 2,200 traditional
dimensions were comparable [17]. Since, sweat glands are Intel CPU cores with 12 NVIDIA TESLA K20 GPUs, each
distributed almost all over the human body and represent with 2,496 CUDA cores. RAAD is running portable batch
a form of cooling in humans. Hence their consideration in system (PBS) for job scheduling. In addition, FAT nodes offer
the model was essential to get a closer insight of EM wave 32 CPU cores coupled with 256 GB of RAM on a single
propagation inside the real human tissue and their effect on motherboard, catering to a certain class of parallel workloads,
the signal propagation losses through the skin. The Optical or workloads with unusually large memory requirements. The
coherent tomography imaging [19] of human skin shows that majority of the cluster’s compute nodes are comprised of 16-
the sweat ducts are of spiral shape. The detailed model of core servers with 64 GB of RAM and a single 128 GB SSD on
human skin with sweat duct is initially presented by Hayut et each server. In this study simulations were performed using 32
CPUs with 32 GB of memory for each individual simulation mixed and separated into sections of 6-well plate. The gels
and more than 50 such simulations were run on the RAAD are incubated (5% CO2) at 37◦ C to support the growth of
cluster. the cells. Fig. 3 also shows the microscopic view of growth
of collagen layer. As mentioned above the artificially cultured
collagen samples were wedged between TPX window with a
B. Terahertz Time Domain Spectroscopy of The Human Skin spacer of 0.2 mm, thus keeping the thickness of the sample
The Terahertz Time Domain Spectroscopy (THz-TDS) at constant for THz TDS measurement.
Queen Mary University of London (QMUL) [21] has a typical
range of 0.1- 4 THz, which provides access to broader spectral
analysis (schematic of this system is shown in Fig. 2). The
THz-TDS system is a carefully designed assembly of the
following components - Ti:Sapphire is the femtosecond pulsed
laser with adjustable wavelength range of 750 - 850 nm;
pulse repetition rate of 80 MHz and peak power is about
1 W. The delay stage has maximum travel distance of 15
cm. THz emitter is LT-GaAs photoconductive antenna with
a biased voltage of 200 V and a gap size of approximately
0.5 mm, which makes the laser beam positioning relatively
easier. A ZnTe crystal is employed as an electro-optic detector
with thickness of 2 mm, which allows enough interaction
length of probe beam and THz wave in the crystal. The
sample consist of two layers: epidermis and dermis, how- Fig. 2. Schematic of Terahertz Time Domain Spectroscopy setup at Queen
Mary University of London, which is used for skin sample measurement [24].
ever, epidermis is very thin in comparison to dermis. The
samples are wedged between two TPX (Polymethylpentene)
slabs with a spacer of known thickness. TPX is used as a
sample holder, since it is transparent to THz radiation; the
light is mostly transmitted through the sample with minimal
absorption in TPX [22]. These measurements are performed
applying the aforementioned THz TDS system and therefore
2.1 cm
phase and amplitude information are collected and later on
post processed using transfer equation based algorithm for
calculating pathloss and other material parameters [23]. The
TDS pulses are generated and detected via mode locked laser.
The lock-in amplifier locks and records the detected THz data
aided by a computer-generated program written in LabViewTM .
In this study, the incoming THz signal consists of total 1024 Fig. 3. Artificially synthesized collagen layer at the Blizard Institute, QMUL
consecutive samples, which are recorded at each position of which was used for measurement (left) & Fibroblast cells assisting the growth
the delay stage. THz-time domain data takes the form of a of collagen layer (right) [24].
single pulse, with a period of 1 picosecond, followed by a
series of attenuated pulses arising either from reflections at the
interface of components within the TDS system, or from etalon III. T ERAHERTZ C HANNEL M ODELING I NSIDE THE S KIN
reflections within the sample itself. Measurements are repeated
several times with same sample to ensure repeatability. A modified Friis equation has been proposed by Jornet et al.
in [12] to calculate the path loss of the THz channel in water
Analogous to the human skin tissue, artificial synthesis of
vapor, which can be divided into two parts: the spreading path
collagen was done in a controlled environment at Subcutaneous
loss P Lspr , which is due to the expansion of waves inside the
Group Blizard Institute, Queen Mary University of London.
tissue and the absorption path loss P Labs (due to absorption
Collagen forms the second layer of human skin (i.e., dermis)
of waves in tissue). Similarly, the path loss in human tissues
and is thickest most layer of the skin. It provides scaffolding
can also be divided into two parts:
for essential cells and structures including blood vessels, lymph
vessels, hair follicles and sweat glands. Collagen type 1 is P Ltotal [dB] = P Lspr (f, d)[dB] + P Labs (f, d)[dB] (1)
extracted from the rat-tail is a heterotrimer consisting of two 4πd
(alpha-1) chains and one (alpha-2) chains [24]. Collagens can = 20 log + 10αd log e (2)
λg
be readily isolated from animal tissues, stored in solution or
as a solid, and used as biocompatible coatings or soft-tissue where, f stands for the frequency, d is the path length, α =
substitutes that can protect and support damaged tissues . The 4πK/λ, where K is extinction coefficient, which measures the
synthesis of collagen (as shown in Fig. 3) was done with a amount of absorption loss. In this study, the electromagnetic
help of standard recipe with total stock solution of 3.35 mg/ml power is considered to spread spherically with distance. This
out of which 80% is collagen. The solution was thoroughly path loss model takes into account only distance and frequency
and it doesn’t explicitly discuss about inside the skin. This 1
paper presents a novel model inside the human skin by Data
Model (dα)
considering additional parameters, which also plays a vital role 0.98

in varying channel characteristics.

0.96

Correlation Level
A. Proposed Channel Model 0.94

In order to infer the path loss model from measurement data, 0.92
investigation on the correlation level between measurement
data and explanatory variables of propagation environment is 0.9

needed, i.e., distance, number of ducts and frequency. The

relationship between distance and measured path loss varies 0.88
0 0.5 1 1.5 2 2.5 3 3.5 4
with distance model and different models are investigated to α

find out the model that has highest correlation level. This (a) Correlation level fitting of α
indicates highest level of information captured by model from
measured path loss. As a first step, the correlation with 0.97
Data
distance in logarithmic scale is tested. The correlation level 0.965 Model ( fγ )

between path loss (P L) in dB scale and distance in logarithmic

0.96
domain is 0.98. In order to find out, if the P L follows other
distance models, the correlation of measured P L with dα is

Correlation Level
0.955

investigated. Figure 4(a) shows variation of correlation level 0.95

with exponent of distance model α. Result shows that there
are some values of α, where the distance model dα has higher 0.945

correlation level than the logarithmic distance model, i.e., 0.94

log10 (d) . In order to find the value of α that has highest 0.935
correlation level, a model of correlation level (C) has been
inferred from correlation level data as a function of α. The 0.93
0 1 2 3 4 5 6 7 8 9 10
model is given as: γ

(b) Correlation level fitting of γ

C(α) = 0.00181α3 − 0.018α2 + 0.021α + 0.984 (3)
Fig. 4. Correlation level fitting of eq. 3 and eq. 4, which is used to calculate
The distance model that has highest correlation is the model the value of α (distance model) and γ (frequency model), which is the
that has α, which fulfills dC
dα = 0. The solution of this equation
maximum value in fitting.
is α = 0.65, which leads to correlation value of 0.9903. The
correlation level fitting of eq. 3 is shown in Figure. 4(a).
An interim distance model (d0.65 ) is used to extract the intensive experiments, it turns out that the model should be of
information of PL data that are not function of distance model the form:
as the error between the PL and the interim distance model.
Then, this extracted residual is tested against frequency as an P L = A(N ) + B(N )d0.65 + C(N )f 4.07 (5)
explanatory model variable to check, if it has any correlation. Where the regression technique is used to find the functions
This correlation between the residual of PL after excluding of A(N ), B(N ), and C(N ), which are the constant offset,
the effect of distance model and frequency model is tested coefficients of distance and frequency as functions of number
against frequency models of log10 (f ) and f γ . The correlation of ducts, respectively. The modeling process is conducted in
level with log10 (f ) is 0.94 and with frequency models different two steps: (i) the fitting process is done for fixed values of N ,
values of γ is shown in Fig. 4(b). It is clear from the figure and model fitting is performed for corresponding data of P L.
that there are some values of γ, where frequency model has (ii) a series of data for A, B and C corresponding to various
higher correlation than the model log10 (f ). In order to find values of N , is modeled as affine functions of N . The fitting
the value that has the highest correlation value, the correlation process is based on least squares fit for particular value of N ,
level is modeled as function of γ and is given as: which can be formulated as:
Cf (γ) = 3.97 ∗ 10−5 γ 3 − 0.00164γ 2 + 0.0114γ + 0.942 (4)
P L = Xβ (6)
The value of γ, which gives highest correlation for fre-
dC
quency model is obtained from dγf = 0, which is γ = 4.07.  
1 d0.65
1 f14.07
In order to find P L model as function of multi variables, 1 d0.65 f24.07 
X= 
2
a multivariable regression technique has been invoked. This . . . (7)
is to find the relationship between two or more explanatory  .. .. .. 
variables by fitting a linear equation to the collected data. After 0.65 4.07
1 dn fn
 TABLE I. F ITTING STATISTICS OF THE PROPOSED MODEL WITH
RESPECT TO MEASUREMENT
B. Performance Evaluation of Proposed Model

Correlation Level R-squared Adjusted R-squared RMSE In order to test the performance of the model against
0.9993 0.9987 0.9987 0.3407 blind data (where number of ducts, distance and frequency
was unknown) that has not been used in fitting data, a new
experiment was conducted to test further of level of blind
][ measurement data that has both new value of number of duct
A1 at different measurements distances. Figure 5(a) shows the
β = B1 (8) comparison of modeled path loss with the measured data for
C1 different values of number of ducts for different distances and
  frequency ranges. Figs. 5(b) and 5(c) shows the comparison
P L1
of modeled and measured data for different variables of the
 P L2 
PL = 
 ... 
 (9) model distance and frequency, respectively. Figure 5(d) shows
the prediction error between blind path loss data (i.e., data
P Ln which was not used in fitting and derivation process of path
loss model) and that predicted by the path loss model and are
The goal here is to find the best values of Ai , Bi and Ci for given as: mean error is 0.6 dB, and its standard deviation is 0.4
each value of N = i, that minimizes the difference between dB. This all indicates the high accuracy of prediction model as
left and right sides such that: compared with the measured data from the simulation, which
have not been used in formulating the proposed model.
β̂ = minβ S(β), (10)
S(β) = ||P L − Xβ|| (11)
C. Comparison of Proposed Model with Measurement Studies
S is the objective function that has to be solved using quadratic
minimization problem since eq. 5 has no solution. The solution The human skin sample absorption is measured using the
of this optimization problem (eq. 11) is well-known and is THz TDS system (the details about measurement are given
given as [25]: in Section II-B). The material properties i.e., refractive index
and absorption coefficient are calculated using details given
β̂ = (X T X)−1 X T P L (12) in [24]. The total path loss is calculated by converting the
measured time domain data to frequency domain by applying
Then, each coefficients (Ai , Bi and Ci ) for different values of fast fourier transform and then averaging it across the band of
N is fitted with affine function of N as follows: interest [12], [15] for THz channel inside the skin. The mea-
A(N ) = a1 + b1 N, (13) sured sample pathloss results are compared with the modeled
B(N ) = a2 + b2 N, (14) pathloss as per eq. 16 (Section III-A). Fig. 6 shows comparison
of measured and modeled path loss for inside the skin. The
C(N ) = a3 + b3 N (15) statistics of prediction error between the THz TDS based skin
The final form of the path loss model inferred from measure- sample measurement and modeled pathloss are given as: mean
ment data after the two described fitting process is given as: error is 2.10 dB, and its standard deviation is 6.23 dB. This
indicates the high accuracy of prediction model as compared
PL = −0.2 ∗ N + 3.98 + (0.44 ∗ N + 98.48)d0.65 to measurement data.
+ (0.068 ∗ N + 2.4)f 4.07 (16)
Where, f is frequency in THz, d is distance in mm , and IV. C ONCLUSIONS
N is number of sweat ducts, respectively.
The fitting statistics of the proposed model is shown in Table In this paper a novel path loss model is proposed inside the
I using R-squared, coefficient of determination, which is a human skin at terahertz frequencies by considering different
measure of how close a statistical model to the measurement distances, number of sweat ducts and frequencies. A blind
data. However, the R-squared does not determine whether the testing is performed on the proposed model and comparison
estimates of the prediction coefficients are biased. It also may is performed with the data from simulation with a very
mislead the goodness of fit as a result of overfitting process. good agreement. Also the comparison of proposed model
These problems are addressed in Adjusted R-squared measure with the THz TDS measurement results show a reasonable
which adjusts the number of explanatory variables to the agreement with the statistics of error prediction in this case
measurement data. This is important measure in multi-variable are: µ = 2.10 dB and σ = 6.23 dB. This all validates
linear regression since it penalizes the model for adding more the accuracy of proposed channel model inside the human
nonsense variables. The root mean squared error (RMSE) is an skin. The reported analysis highlights the novel method to
absolute measure of close the model to the measurement data. cater the communication challenges in such environment and
It can be interpreted as the standard deviation of unexplained paved a way for further studies in this harsh environment, even
and un-modeled variation in data. The lower value of RMSE combined with hybrid molecular communication techniques
is the better fit. for health-care applications.
 60 75
Model
Data 70 Measured
55 Model
65

50 60

55
Path Loss (dB)

Pathloss (dB)
45
50

40 45

40
35
35

30 30

25
25
0 1 2 3 4 5 20
Number of Ducts 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2 1.3 1.4 1.5
Frequency (THz)

(a) Path loss with respect to number of ducts (N)

Fig. 6. Comparison of measured (THz TDS) and modeled path loss inside
60
the skin at THz frequencies.
55 Data
Model

50
R EFERENCES
Path Loss (dB)

45

40
[1] S. F. Bush, Nanoscale Communication Networks. Artech House, 2010.
[2] I. F. Akyildiz, F. Brunetti, and C. Blázquez, “Nanonetworks: A new
35
communication paradigm,” Computer Networks, vol. 52, no. 12, pp.
30
2260–2279, 2008.
[3] M. Sitti, H. Ceylan, W. Hu, J. Giltinan, M. Turan, S. Yim, and E. Diller,
25
0.1 0.15 0.2 0.25 0.3 0.35 “Biomedical applications of untethered mobile milli/microrobots,” Pro-
Distance (mm) ceedings of the IEEE, vol. 103, no. 2, pp. 205–224, 2015.
(b) Path loss with respect to distance at N = 4 [4] S. Balasubramaniam and J. Kangasharju, “Realizing the internet of nano
things: Challenges, solutions, and applications,” Computer, vol. 46,
60 no. 2, pp. 62–68, Feb 2013.
Data
Model [5] G. Piro, K. Yang, G. Boggia, N. Chopra, L. Grieco, and A. Alo-
55
mainy, “Terahertz communications in human tissues at the nano-scale
50 for healthcare applications,” IEEE Transactions on Nanotechnology,
vol. 14, no. 3, pp. 404–406, 2015.
Path Loss (dB)

45
[6] P. S. Hall and Y. Hao, Antennas and Propagation for Body-Centric
40
Wireless Communication. Artech House, 2012.
[7] Q. H. Abbasi, A. Sani, A. Alomainy, and Y. Hao, “Numerical charac-
35 terization and modeling of subject-specific ultrawideband body-centric
radio channels and systems for healthcare applications,” Information
30
Technology in Biomedicine, IEEE Transactions on, vol. 16, no. 2, pp.
25
221–227, 2012.
0.8 0.85 0.9 0.95 1 1.05 1.1 1.15 1.2
Frequency (THz) [8] S. Luryi, J. Xu, and A. Zaslavsky, Future Trends in Microelectronics:
Frontiers and Innovations. John Wiley & Sons, 2013.
(c) Path loss with respect to frequency for different distances at N =
4 [9] T. Binzoni, A. Vogel, A. H. Gandjbakhche, and R. Marchesini, “De-
tection limits of multi-spectral optical imaging under the skin surface,”
350
Physics in medicine and biology, vol. 53, pp. 617–636, 2008.
[10] E. Berry, A. J. Fitzgerald, N. N. Zinov’ev, G. C. Walker, S. Homer-
300 Vanniasinkam, C. D. Sudworth, R. E. Miles, J. M. Chamberlain, and
M. A. Smith, “Optical properties of tissue measured using terahertz-
250
pulsed imaging,” in Medical Imaging 2003. International Society for
Frequency of Error

200
Optics and Photonics, 2003, pp. 459–470.
[11] A. Fitzgerald, E. Berry, N. Zinov’ev, S. Homer-Vanniasinkam, R. Miles,
150
J. Chamberlain, and M. Smith, “Catalogue of human tissue optical
properties at terahertz frequencies,” Journal of Biological Physics,
100
vol. 29, no. 2-3, pp. 123–128, 2003.
50 [12] J. M. Jornet and I. F. Akyildiz, “Channel modeling and capacity
analysis for electromagnetic wireless nanonetworks in the terahertz
0
−0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 band,” Wireless Communications, IEEE Transactions on, vol. 10, no. 10,
Prediction Error (dB)
pp. 3211–3221, 2011.
(d) Prediction Error of proposed model with actual input data [13] I. F. Akyildiz and J. M. Jornet, “Electromagnetic wireless nanosensor
networks,” Nano Communication Networks, vol. 1, no. 1, pp. 3–19,
Fig. 5. Blind testing of proposed model by varying number of ducts, distance 2010.
and frequency, while comparing to measured data and the prediction error. [14] J. M. Jornet, J. C. Pujol, and J. S. Pareta, “Phlame: A physical layer
 aware mac protocol for electromagnetic nanonetworks in the terahertz
band,” Nano Communication Networks, vol. 3, no. 1, pp. 74–81, 2012.
[15] k. Yang, A. Pellegrini, M. O. Munoz, A. Brizzi, A. Alomainy, and
Y. Hao, “Numerical analysis and characterization of THz propagation
channel for body-centric nano-communications,” IEEE Transactions on
Terahertz Science and technology, vol. 5, pp. 419–426, May 2015.
[16] Y. Feldman, A. Puzenko, P. Ishai, A. Caduff, and A. Agranat, “The
electromagnetic response of human skin in the millimetre and submil-
limetre wave range,” Phys. Med. Biol., vol. 54, pp. 3341–3363, Jun.
2009.
[17] K. Yang, Q. Abbasi, N. Chopra, M. Munoz, Y. Hao, and A. Alomainy,
“Effects of non-flat interfaces in human skin tissues on the In-Vivo THz
communication channel,” Journal of Nano Communication Network,
Nov., 2015.
[18] W. Montagna, The Structure and Function of Skin 3E. Elsevier, 2012.
[19] I. Hayut, A. Puzenko, P. B. Ishai, A. Polsman, A. J. Agranat, and
Y. Feldman, “The helical structure of sweat ducts: Their influence on the
electromagnetic reflection spectrum of the skin,” Terahertz Science and
Technology, IEEE Transactions on, vol. 3, no. 2, pp. 207–215, March
2015.
[20] G. Shafirstein and E. G. Moros, “Modelling millimetre wave propaga-
tion and absorption in a high resolution skin model: the effect of sweat
glands,” Physics in medicine and biology, vol. 56, no. 5, p. 1329, 2011.
[21] O. Sushko, “Terahertz dielectric study of bio-molecules using time-
domain spectrometry and molecular dynamics simulations,” PhD The-
sis, Queen Mary University of London, 2014.
[22] J. R. Birch and E. A. Nicol, “The f ir optical constants of the polymer
tpx,” Infrared Phys., vol. 24, pp. 573 – 575, 1984.
[23] T. D. Dorney, R. G. Baraniuk, and D. M. Mittleman, “Material param-
eter estimation with THz time-domain spectroscopy,” J. Opt. Soc. Am.
A, vol. 18, pp. 1562–1571, 2001.
[24] N. Chopra, A. Alomainy, and M. Philpot, “Investigating electromagnetic
material properties of collagen at THz for health monitoring applica-
tions,” in MobiHealth 2015. EAI, 2015.
[25] S. Kay, Fundamentals of statistical signal processing: estimation theory.
Prentice-Hall, 1993.