Developing The Digital Lung From First Lung CT To Clinical Ai John D Newell Full Chapter
Developing The Digital Lung From First Lung CT To Clinical Ai John D Newell Full Chapter
Developing The Digital Lung From First Lung CT To Clinical Ai John D Newell Full Chapter
No part of this publication may be reproduced or transmitted in any form or by any means,
electronic or mechanical, including photocopying, recording, or any information storage
and retrieval system, without permission in writing from the publisher. Details on how
to seek permission, further information about the Publisher’s permissions policies and
our arrangements with organizations such as the Copyright Clearance Center and the
Copyright Licensing Agency, can be found at our website: www.elsevier.com/permissions.
This book and the individual contributions contained in it are protected under copyright by
the Publisher (other than as may be noted herein).
Notice
Practitioners and researchers must always rely on their own experience and knowledge
in evaluating and using any information, methods, compounds or experiments described
herein. Because of rapid advances in the medical sciences, in particular, independent
verification of diagnoses and drug dosages should be made. To the fullest extent of the
law, no responsibility is assumed by Elsevier, authors, editors or contributors for any
injury and/or damage to persons or property as a matter of products liability, negligence
or otherwise, or from any use or operation of any methods, products, instructions, or
ideas contained in the material herein.
Printed in India
1 Introduction to Lung CT AI 1
2 Three-Dimensional (3D) Digital Images of the Lung Using X-ray Computed
Tomography 12
Index 143
xi
I dedicate this book to all who suffer from lung disease
and to those who strive to research, diagnose, and treat
lung disease.
PREFACE
ix
Conf idence
is ClinicalKey
Evidence-based answers,
continually updated
I want to express my deep thanks and gratitude to all my patients, mentors, students, and colleagues
who have all taught me so much over the past 45 years. I want to thank all the people at Elsevier,
and especially my editor, Rebecca Corradetti and Shereen Jameel, who helped me so much in
improving the quality of this book.
vii
Any screen.
Any time.
Anywhere.
Activate the eBook version
of this title at no additional charge.
Elsevier eBooks+ gives you the power to browse, search, and customize
your content, make notes and highlights, and have content read aloud.
1
2 DEVELOPIING THE DIGITAL LUNG
Start AI
Environment
Program
Observations
Decide and
Objective 3 Objective 1
Take Action
Objective 2
Fig. 1.1 The four elements of an intelligent agent (AI agent): environment, observations, actions,
and objective(s).
Character Typed
Store Character in a
Text File on the Hard
Disk
Fig. 1.2 How the four elements of an AI agent work to enable the objective(s) of a simple word
processing program.
1—Introduction to Lung CT AI 3
reactive machine; (2) limited-memory; (3) theory of mind; and (4) self-aware2 (Box 1.1).
Since the late 1960s and early 1970s, reactive machine type AI has driven the develop-
ment of multiple AI technologies to visually and, subsequently, quantitatively assess the
presence and extent of lung diseases using x-ray CT scanning. More recently, limited-
memory levels of AI have been added to the list of technologies driving progressive
improvements in the visual and quantitative CT assessment of lung disease. Reactive
machines are the most basic type of AI system. This means that they cannot form mem-
ories or use past experiences to influence presently made decisions; they can only react
to currently existing situations—hence the term “reactive.” An existing form of a reac-
tive machine is Deep Blue, a chess-playing supercomputer created by IBM in the mid-
1980s.2 Reactive machine AI programs will only react in the present in the way they are
programmed. Examples of reactive machine AI programs in CT lung AI would include
analytic CT image reconstruction algorithms, analytic CT image lung segmentation
programs, computing the lung CT image voxel histogram, and relevant voxel histogram
statistics that have been previously shown to correlate with normal and diseased lung
tissue. These reactive machine lung CT AI metrics of lung disease can be used to detect
and assess the extent of normal and abnormal lung structure and function caused by
underlying lung disease, such as emphysema, asthma, or pulmonary fibrosis. The advan-
tage of reactive machine learning AI is that it is very clear what the software program
is doing. However, it is not as powerful as limited-memory AI. Limited-memory AI is
comprised of machine learning models that derive knowledge from previously learned
information, stored data, or events. Unlike reactive machines, limited-memory learns
from the past by observing actions, or data fed to them, in order to build experiential
knowledge.2 Limited-memory machine learning AI has been used in identifying specific
patterns of lung disease, such as honeycombing in patients with ILD. Limited-memory
machine learning AI has also been applied to CT image reconstruction, reducing noise
in reconstructed CT images. It has also been implemented in segmentation software
to extract the lungs from the rest of the thoracic anatomy on CT images. Limited-
memory AI machine learning has been used to identify unique patient CT phenotypes
in patients with COPD3 and asthma.4 Limited-memory AI machine learning has also
been used to classify lung nodules into benign and malignant categories.5
Lung CT AI involves a number of simpler AI agents that are linked together to
produce a final lung CT AI objective, which is to detect and assess normal and dis-
eased lung structure and function and make these results widely available to patients
and healthcare providers. The lung CT AI agents that are included in this book, in the
order in which they are usually performed, are: (1) generate high-quality CT images of
the thorax; (2) display the CT images of the thorax; (3) separate or segment the lung CT
images from the rest of the thoracic anatomy; (4) extract quantitative features from the
4 DEVELOPIING THE DIGITAL LUNG
lung CT images that represent meaningful metrics of normal and diseased lung struc-
ture and function; (5) analyze these features to predict the presence or absence of lung
disease, and to assess the extent and impact of any detected lung disease; and (6) make
these results available in near real time for patients and their health care providers (Fig.
1.3). The routine lung CT AI outputs for individual patient care can then be systemati-
cally collected and examined across healthcare systems and countries; these results can
Start
CT Scanning
Protocol
Fig. 1.3 The sequence of lung CT AI agents in the order in which they are performed: generate
high-quality CT images of the thorax, display the CT images of the thorax, separate or segment
the lung CT images from the rest of the thoracic anatomy, extract quantitative CT (QCT) features
from the lung CT images that represent meaningful metrics of normal and diseased lung structure
and function, analyze these features to predict the presence or absence of lung disease and to
assess the extent and impact of any detected lung disease, and make these results available in
near realtime for patients and their health care providers.
1—Introduction to Lung CT AI 5
be used in near real time to assess world lung health and disease, and to inform how
best to allocate scarce resources needed to decrease the prevalence and severity of lung
disease. My biggest reason for writing this book is to make everyone aware that the
lung CT AI objectives outlined above are now achievable. I believe lung CT AI will
work best in a healthcare environment where everyone has access to quality healthcare,
regardless of ability to pay or preexisting conditions.
This book will follow the sequence of AI agents that have been developed over the last
50 years to generate a 3D digital representation of the lung using x-ray CT,6,7 and finally
develop an AI agent that is useful in detecting and predicting the malignant poten-
tial of pulmonary nodules;5 the presence and severity of diffuse lung diseases, such as
chronic obstructive pulmonary disease (COPD),8 idiopathic pulmonary fibrosis (IPF),9
and, most recently, COVID-19 viral pneumonia.10 The progressive improvements in CT
scanner technology have advanced the 3D visualization of normal lung structure and
function and made it possible to develop new AI methods to accurately detect and assess
normal and diseased states of the lung. The progressive increases in our knowledge of
lung CT AI will help guide our narrative from describing the first digital images of the
lung obtained from x-ray computed tomography, to the very recent exciting application
of quantitative CT AI methods for the detection and assessment of the severity of lung
nodules and acute and chronic diffuse lung disease. The high spatial and contrast resolu-
tion of modern chest CT images of the lung has also enabled the creation of sophisti-
cated software to build silicon computer models of normal and diseased lungs, and has
increased our fundamental understanding of lung physiology and pathophysiology.11,12
x-ray tube cooling was needed. The scan time and the spatial resolution of the CT scan-
ner are important metrics that drive what can be done with AI in analyzing images of
the lung. Chapter 2 also follows several key technologies that were developed to improve
the spatial resolution and decrease scan times of x-ray CT scanners; scan time and spa-
tial resolution have greatly improved over the last 40 years.
Chapter 3 discusses the latest generation of x-ray CT scanner technologies and lung
CT scanning protocols available in 2020. The latest generation of CT scanners can
scan the entire thorax in less than 10 seconds with an isotropic resolution of 0.5 mm.
Chapter 3 details the important CT scanning variables that need to be carefully selected,
such as x-ray dose, scan time, z-axis resolution, and image reconstruction method, to
produce the best lung CT images.
The second segment of the book, Chapters 4 through 7, describes the lung CT AI
methods that have been developed to detect and quantitatively assess focal lung nod-
ules, pulmonary emphysema, pulmonary fibrosis, and acute COVID-19 viral pneumo-
nia from the CT-generated density maps of the lung.
Chapter 4 discusses the concept of the lung nodule and how CT is used to detect and
quantitatively assess the malignant potential of a lung nodule. Exposure to environmen-
tal factors, especially cigarette smoke, increases the risk of developing lung cancer. It has
been shown recently that the use of screening lung CT scans can decrease mortality in
people exposed to cigarette smoke.
Chapters 5, 6, and 7 introduce increasingly more sophisticated lung CT AI methods
to assess diffuse lung disease, starting with CT image voxel histograms in Chapter 5
and ending with sophisticated limited-memory AI machine learning algorithms in
Chapter 7.
Chapter 5 describes the QCT metrics that are readily obtained from a single total
lung capacity (TLC) CT scan done with an appropriate CT protocol, as outlined in
Chapter 3. Lung CT density metrics were the first lung CT AI metrics to be reported
and focused mainly on the reduced density of the lung produced by pulmonary emphy-
sema. Pulmonary fibrosis produces increased density in the lung and can also be assessed
using density measures from lung CT images. Chapter 5 describes the research studies
that were performed to determine if the QCT metrics described do, in fact, represent
important features of normal and diseased lung tissue by correlating CT image find-
ings to other measures of lung disease (e.g., pulmonary function tests, lung pathology).
Chapter 5 discusses how quantitative CT lung images provide critical new information
regarding COPD that was not obtained from other methods (e.g., clinical history, pul-
monary function testing).16 Determining the value of lung CT AI versus other means
of assessing patients with COPD required the funding of large multicenter NIH grants
that studied thousands of subjects with and without COPD. These studies established
the relative value of lung CT AI metrics with other data characterizing COPD-related
lung disease, including genetic studies, pulmonary physiology testing, and standard-
ized healthcare questionnaires. These studies included COPDGene, MESA Lung, and
SPIROMICS. One QCT lung AI study in the SPIROMCS study identified four unique
clusters of subjects with COPD that had different QCT metrics and different disease
trajectories.3 Chapter 5 also discusses the results of multiple smaller-size research studies
in assessing the value of lung CT AI in groups of patients with pulmonary fibrotic lung
disease associated with idiopathic pulmonary fibrosis (IPF), connective tissue disease
8 DEVELOPIING THE DIGITAL LUNG
Start
Acceptable Acceptable
No No
Performance Performance
Yes Yes
Fig. 1.4 The three steps used to train limited-memory type lung CT AI algorithms to detect and
assess normal and diseased lung tissue. The first stage is feature extraction. This can be done
using supervised or unsupervised approaches. Supervised approaches have an expert imaging
physician label the normal and abnormal tissue features on the lung CT images. Unsupervised
approaches let the AI algorithm determine by itself the normal and abnormal tissue types. The
second stage trains the AI algorithm to recognize and quantitate the amount of the normal and
abnormal tissue features, or textures, within the lung CT images that correspond to important
pathologic features associated with different lung diseases. The number of supervised or unsu-
pervised training CT cases can be increased to improve the performance of the limited-memory
lung CT AI program. The design of the limited-memory lung CT AI program can also be altered
to improve its performance. The third stage tests the performance of the trained AI algorithm on
a test set of chest CT scans from a cohort of human subjects separate from the training cohort.
The performance of the lung CT AI algorithm is based on its ability to identify the different tissue
features on the test set of chest CT scans including features such as emphysema, honeycomb-
ing, and consolidation.
10 DEVELOPIING THE DIGITAL LUNG
of the main goals of this book is to show how the advancement of lung CT AI over
the past 45 years has made it possible to provide lung CT AI technologies in the cur-
rent clinical medical imaging environment. These advances in computing have enabled
the emergence of several small clinical CT lung AI imaging companies that began as
research-only enterprises, to help provide the lung CT AI tools needed for large NIH
studies such as COPDGene, Mesa Lung, and SPIROMICS. The success of lung CT AI
in the research realm has motivated these small lung CT AI companies to develop new
clinical products in lung CT AI.
VIDA is one of these companies that has developed lung CT AI software that can
run independently, or on large enterprise-scale medical imaging AI ecosystems so that
the previously validated quantitative CT metrics of lung disease can be extracted in near
real time, and made available to the imaging and referring physician for the immedi-
ate care of the patient at the point of care. Chapter 9 reviews the current information
technologies (IT) that are in use in modern healthcare hospitals and clinics to support
the acquisition, storage, and distribution of medical imaging studies, including x-ray
CT studies of the thorax and how the imaging IT interacts with the larger healthcare
IT that supports the patient’s electronic medical record (EMR). We then discuss the
lung CT AI technology that is now available within the imaging IT ecosystem to auto-
matically assess lung CT images for quantitative metrics of lung disease. Specifically,
we discuss VIDA Insights v3.0 Density/tMPR module and Texture/Subpleural View
module. Both use reactive machine AI and limited-memory AI methods to analyze each
lung and lobe of a chest CT scan for the following quantitative CT metrics: lung and
lobe volume in liters, LAA-950 metric for emphysema, HAA-700 to -250 metric for COVID-
19 pneumonia and ILD, and the texture patterns that make up the HAA-700 to -250 that
include ground-glass/reticular opacities, consolidation, and honeycombing. The chapter
concludes by discussing the importance of responsible AI in any application of AI, along
with guidelines to achieve responsible lung CT AI. The current technology state of lung
CT AI can accomplish what could only have been dreamed about by investigators when
the “Density Mask” CT technique was first published in 1988 as a CT method of assess-
ing reduced lung density by emphysema.20
References
1. Russell S. Human Compatable: Artificial Intelligence and the Problem of Control. New York, NY:
Viking; 2019:336.
2. Reynoso R. 4 Main Types of Artificial Intelligence. 2019. Available at: https://www.g2.com/
articles/types-of-artificial-intelligence.
3. Haghighi B, Choi S, Choi J, Hoffman EA, Comellas AP, Newell Jr. JD, et al. Imaging-
based clusters in current smokers of the COPD cohort associate with clinical characteristics:
the SubPopulations and Intermediate Outcome Measures in COPD Study (SPIROMICS).
Respir Res. 2018;19(1):178.
4. Choi S, Hoffman EA, Wenzel SE, Castro M, Fain S, Jarjour N, et al. Quantitative computed
tomographic imaging-based clustering differentiates asthmatic subgroups with distinctive
clinical phenotypes. J Allergy Clin Immunol. 2017;140(3):690–700.e8.
5. Uthoff J, Stephens MJ, Newell Jr. JD, Hoffman EA, Larson J, Koehn N, et al. Machine learn-
ing approach for distinguishing malignant and benign lung nodules utilizing standardized
perinodular parenchymal features from CT. Med Phys. 2019;46(7):3207–3216.
6. Ledley RS, Di Chiro G, Luessenhop AJ, Twigg HL. Computerized transaxial x-ray tomog-
raphy of the human body. Science. 1974;186(4160):207–212.
1—Introduction to Lung CT AI 11
7. Schellinger D, Di Chiro G, Axelbaum SP, Twigg HL, Ledley RS. Early clinical experience
with the ACTA scanner. Radiology. 1975;114(2):257–261.
8. Humphries SM, Notary AM, Centeno JP, Strand MJ, Crapo JD, Silverman EK, et al.
Deep learning enables automatic classification of emphysema pattern at CT. Radiology.
2020;294(2):434–444.
9. Humphries SM, Yagihashi K, Huckleberry J, Rho BH, Schroeder JD, Strand M, et al.
Idiopathic pulmonary fibrosis: data-driven textural analysis of extent of fibrosis at baseline
and 15-month follow-up. Radiology. 2017;285(1):270–278.
10. Li L, Qin L, Xu Z, Yin Y, Wang X, Kong B, et al. Artificial intelligence distinguishes
COVID-19 from community acquired pneumonia on chest CT. Radiology. 2020:200905.
11. Tawhai M, Clark A, Donovan G, Burrowes K. Computational modeling of airway and pul-
monary vascular structure and function: development of a “lung physiome”. Crit Rev Biomed
Eng. 2011;39(4):319–336.
12. Tawhai MH, Bates JH. Multi-scale lung modeling. J Appl Physiol. (1985).
2011;110(5):1466–1472.
13. Heron M. Deaths: leading causes for 2017. National Vital Statistics Reports. 2019;68(6):1–76.
14. Fischer A, Patel NM, Volkmann ER. Interstitial lung disease in systemic sclerosis: focus on
early detection and intervention. Open Access Rheumatol. 2019;11:283–307.
15. VIDA. VIDA 2020. Available at: https://vidalung.ai.
16. Lowe KE, Regan EA, Anzueto A, Austin E, Austin JHM, Beaty TH, et al. COPDGene((R))
2019: redefining the diagnosis of chronic obstructive pulmonary disease. Chronic Obstr Pulm
Dis. 2019;6(5):384–399.
17. Best AC, Lynch AM, Bozic CM, Miller D, Grunwald GK, Lynch DA. Quantitative CT
indexes in idiopathic pulmonary fibrosis: relationship with physiologic impairment. Radiology.
2003;228(2):407–414.
18. Best AC, Meng J, Lynch AM, Bozic CM, Miller D, Grunwald GK, et al. Idiopathic pulmo-
nary fibrosis: physiologic tests, quantitative CT indexes, and CT visual scores as predictors of
mortality. Radiology. 2008;246(3):935–940.
19. Podolanczuk AJ, Oelsner EC, Barr RG, Hoffman EA, Armstrong HF, Austin JH, et al. High
attenuation areas on chest computed tomography in community-dwelling adults: the MESA
study. Eur Respir J. 2016;48(5):1442–1452.
20. Muller NL, Staples CA, Miller RR, Abboud RT. “Density mask”. An objective method to
quantitate emphysema using computed tomography. Chest. 1988;94(4):782–787.
C H A P T E R 2
Three-Dimensional (3D) Digital Images
of the Lung Using X-ray Computed
Tomography
This chapter will discuss the digital lung, x-rays, and key components of the x-ray CT
scanner to help better understand lung CT AI scanning protocols, and to briefly review
the historical progression of advancements in x-ray computed tomography of the lungs
from the 1970s through the development of multidetector spiral CT (MDCT) scanners
in the early 2000s. Each CT technology advancement improved visual and quantitative
assessment of x-ray CT images of the lung. The challenges that needed to be overcome
included decreasing the examination time, increasing the spatial resolution of the CT
images, and decreasing the x-ray dose to the patient. Each of these improvements, dur-
ing this timeframe, was very important toward enabling future AI approaches in the
diagnosis and assessment of lung diseases.
12
2—THREE-DIMENSIONAL (3D) DIGITAL IMAGES OF THE LUNG 13
y = 0.5 mm
x = 0.5 mm
z = 0.5 mm
z-axis
y-axis
A B x-axis
Fig. 2.1 (A) 3D rendering of both lungs obtained using a modern multidetector x-ray computed
tomographic scanner and software to render the millions of voxels that were acquired into a
single 3D rendering of the lungs. [Courtesy of VIDA.] (B) The cube shape of a 0.5-mm isotropic
x-ray CT voxel is shown here and its relationship to a 2D axial chest CT image.
The x-ray CT scanner consists of several major components: x-ray tube and electron-
ics, x-ray detector and electronics, CT gantry, patient table, computer hardware, soft-
ware, and displays (Fig. 2.3). From the very first x-ray CT scanner to the current modern
x-ray CT scanners of the 2020 s, they all have needed these components to produce and
display a CT image.2 The advances in CT technology over the past five decades were
made possible by advances in each of these key x-ray CT scanner components. It should
be emphasized that without the AI software program running the x-ray CT scanner,
there would be no practical way to reconstruct the CT images. The advent of relatively
small and capable minicomputers in the 1960s made it possible to construct commer-
cially viable x-ray CT scanners. The basic design of CT scans from the 1970s to the
present share several common design features. Since the advent of human CT scans, it
has been desirable to have the patient lay still in a supine position on the CT “scanning”
table. The table moves the patient along their long axis, head-to-toe or z-axis, so the
entire body can be scanned, if desired. The need to obtain multiple 1D projections of the
supine patient laying in one position on the scanning table led to a CT scanner design
where the x-ray tube rotates around the supine stationary subject with its corresponding
x-ray detector on the opposite side of the patient (Fig. 2.4). The x-ray tube produces a
beam of x-rays and these x-ray beams can have several geometric shapes including pen-
cil beam, broad parallel beam, fan beam, or cone beam, depending on the CT scanner
design. The most common current configuration of the x-ray beam is a rotating 50- to
60-degree cone beam (Fig. 2.5). The shape of the x-ray beam determines the shape of
the x-ray detector array. The rotation axis of the x-ray tube and x-ray detector assembly
is around the z-axis of the subject, and the highly collimated x-ray beam travels through
the patient and onto the x-ray detectors. Today, the xy and z-axis width of individual
detectors is less than 1 mm depending on the x-ray CT scanner model.
14 DEVELOPIING THE DIGITAL LUNG
Patient
x
r
a
y
b
e
a
m
Z Axis
Fig. 2.2 Graphical representation of a tightly collimated x-ray beam in the z-axis that is transmit-
ted through the thorax of a subject. The tightly collimated x-ray beam reduces the ionizing radia-
tion exposure to adjacent tissues and increases the contrast resolution of the scan over what is
possible with projection radiography, also shown here.
There are other medical imaging techniques that can generate 3D digital images of
the lung; including magnetic resonance computed tomography (MRI), single-photon
emission computed tomography (SPECT), positron emission computed tomography
(PET), and hybrid imaging that combines two of these techniques (e.g., PET/CT,
SPECT/CT, PET/MRI). X-ray transmission computed tomography (CT) is the most
widely available computed tomographic method and is uniquely capable of rapid 3D
imaging of the lungs, with both high image contrast and high spatial resolution, linear
density scale over typical lung density values (−1000 HU to 0 HU), low claustropho-
bia, and low dose radiation (0.15 mGy to 1.5 mGy using modern x-ray CT scanners).3
2—THREE-DIMENSIONAL (3D) DIGITAL IMAGES OF THE LUNG 15
Patient
CT Gantry
CT Table
Patient
Fig. 2.3 Several major components of an x-ray CT scanner, including the gantry containing the
x-ray tube, high energy voltage generator and corresponding x-ray detectors, and the CT scan-
ning table with a patient positioned supine at the isocenter of the gantry. The CT technologist’s
computer and monitors are typically located in an adjoining room.
x-ray tube
x-ray beam
isocenter
patient
max FOV
detector array
Fig. 2.4 The geometric relationships of the x-ray tube, patient, maximum scan field of view, and
the x-ray detector array of third-generation CT scanner. Many 1D projections of the patient’s
anatomy are obtained as the x-ray tube and x-ray detector move around the patient on the CT
scanning table. The patient is positioned at the isocenter of rotation of the CT gantry. The maxi-
mum scan field of view is indicated. The reconstruction or display field of view (DFOV) can be
smaller than the maximum scan field of view.
16 DEVELOPIING THE DIGITAL LUNG
x-ray tube
2D detector array
Fig. 2.5 Narrow 3D cone x-ray cone beam and its corresponding curved 2D x-ray detector
array. This is the most common current geometric configuration of the x-ray beam and detector
array used in clinical x-ray CT scanners.
The arrival of MDCT scanners in the 2000s, with 64 or more z-axis detector channels,
has fueled multiple high-quality research studies that have increased our understanding
of lung CT AI, its application in detecting, and assessing lung cancer and diffuse lung
disease. Many research studies that have been done using x-ray CT in assessing COPD,
ILD, and lung cancer, and these research studies provide the scientific foundations for
this book (see Chapters 4 through 8).
X-RAYS
There are two forms of ionizing electromagnetic radiation, x-rays, and gamma rays.
They are both high-energy photons capable of breaking covalent chemical bonds and
because they can break covalent chemical bonds, are referred to as ionizing electromag-
netic radiation.4 Liquid water is the most abundant molecule in the human body and
the ionization energy to remove an electron from water is 11.2 electron volts (eV).4 A
photon with an energy equal to or greater than 11.2 eV is considered ionizing radiation.4
Covalent organic chemical bonds link together atoms in carbon-based living organisms
and breaking these bonds can have serious consequences for the organism. The 3D dis-
tribution of energy concentration in units of joules per kilogram of the x-ray photons
passing through the body is referred to as ionizing radiation absorbed dose. X-ray and
gamma-ray photons are defined by how they are produced. The x-ray is produced out-
side the nucleus of an atom. In contrast, gamma rays are produced within the nucleus of
the atom through radioactive decay of an excited nucleus. The energy of the gamma-ray
is equal to the difference in the initial excited energy state of the nucleus and a lower
state, or a ground state, of the nucleus.
2—THREE-DIMENSIONAL (3D) DIGITAL IMAGES OF THE LUNG 17
The differential attenuation of the x-ray beam photons in biologic material provides
the mechanism to generate diagnostic medical images of the human body. The differ-
ential attenuation of the photons involves elastic (Rayleigh) scattering of x-ray photons,
photoelectric absorption of x-ray photons, and Compton scattering of x-ray photons, as
the principal mechanisms of differential attenuation of the x-ray beam (using peak x-ray
tube voltages (kVp) between 25 kV and 150 kV).5 The dominant attenuation mechanism
of x-rays in lung tissue and other soft tissue using a kVp between 70 kV and 150 kV
(typical of non-contrast x-ray CT of the thorax) is Compton scattering.5 The probability
of a Compton scattering event per unit mass of incident x-rays on low atomic number
materials (typical of normal lung tissue) is nearly independent of the atomic number (Z)
of the scattering lung tissue.5 The probability of a Compton scattering event per unit
volume of incident x-rays on low atomic number materials is proportional to the density
of the material.5 The reason for this is, except for the hydrogen atom, the number of elec-
trons per gram of tissue for low atomic number materials, such as carbon and oxygen, is
relatively constant, so the number of electrons per gram of tissue is driven mainly by tis-
sue density; the result is that the attenuation of x-rays in the lung tissue is proportional
to the density of the lung tissue.5 A good friend and colleague of mine, Jim Hogg, MD,
has likened the x-ray CT scanner to a lung densitometer, and the above discussion sup-
ports this view. Photoelectric absorption of x-rays does come into play when intravenous
iodinated contrast material is given to the patient just before the scan is performed. The
presence of a high-atomic-number material like iodine (Z = 53, with a K shell orbital
electron binding energy of 33 kV) will mean that the absorption of incident x-rays by
iodine atoms will depend primarily on the photoelectric effect when incident x-rays
interact with an iodine atom.5
resistance to heat and wear, as well as a high atomic number (Z = 74), so it is an ideal
anode target material. A Tungsten (90%)-Rhenium (Re, Z = 75) (10%) alloy can be
used to increase the resistance to anode surface damage.6
The impact of the electrons on a tungsten metal anode produces the x-ray photons
using two different methods: characteristic x-ray photon method and Bremsstrahlung
x-ray photon method.6 The characteristic x-rays are generated when an electron with
sufficient kinetic energy strikes and ejects an inner electron (K electron orbit) in
material with a high atomic number (Z), such as tungsten (W, Z = 74). Subsequently,
an outer electron with the same high Z material transitions to the vacancy in the
inner electron orbital and, in this process, an x-ray photon is generated with energy
equal to the difference in the binding energy of the outer electron and the inner elec-
tron that was ejected. X-ray photons greater than 100 eV are termed characteristic
photons whose energies are determined by the metal anode target. The most ener-
getic characteristic x-rays from tungsten are due to transitions from the L (57.98 keV,
59.32 keV ), or M shell (67.24 keV ) to the K shell of tungsten.6 The Bremsstrahlung,
or breaking method of x-ray production, involves the deceleration of electrons as they
penetrate a high Z anode target material, such as tungsten, and the deceleration of
the electrons generates x-rays with photon energies between the working function of
the metal and the keV used in the x-ray tube. The Bremsstrahlung x-rays are usually
more abundant than characteristic x-rays in a CT scanner x-ray tube. The efficiency
of generating Bremsstrahlung x-rays increases with the atomic number of the anode
(Z = 74 for tungsten) and with the kilovoltage that is applied between the cathode
and the anode.6
Modern x-ray tubes designed to be used with x-ray CT scanners have high power
ratings, 5–7 megajoules, to be able to produce a large number of x-ray photons in a very
short period of time.1 The size of the x-ray tube focal spot on the tungsten anode is an
important factor in determining the spatial resolution of the CT scanner. A smaller focal
size can improve spatial resolution. However, a smaller focal spot size decreases the max-
imum number of x-ray photons that can be produced in a given period of time. Selection
of the optimum focal spot size is an important CT scanning protocol variable for lung
CT AI. The focal spot size of current x-ray CT scanners ranges from 1 to 2 mm.1
The x-ray tube peak kilovoltage (kVp) and tube current (mA) are two key variables
in a lung CT scanning protocol. The kVp determines the x-ray beam peak energy,
x-ray beam energy spectrum, and the efficiency of producing x-rays. The increase in
efficiency of photon production will increase the ionizing radiation dose to the patient.
The higher the kVp, the greater efficiency of producing x-rays for a given tube current.
The higher the kVp (between 70 kV and 150 kV), the lower the tissue contrast. It is
important in quantitative CT work to agree on the same kVp, since different kVp set-
tings will change the HU values assigned to lung voxels for a given mA. The higher
the kVp, the more efficient the x-ray photon production, so for a given mA, a higher
kVp will result in higher radiation absorbed dose to the patient. For a given kVp, the
tube current will determine the amount of x-ray photons produced. The higher the tube
current, the more x-ray photons are produced for a given kVp. A higher tube current
results in less image noise, but also higher radiation absorbed dose to the patient for a
given kVp.
2—THREE-DIMENSIONAL (3D) DIGITAL IMAGES OF THE LUNG 19
The x-ray tube current in mA determines the number of electrons that are acceler-
ated between the x-ray tube cathode and anode per unit time. The tube current in mA
multiplied by the length of time the x-ray tube is turned on, exposure time in seconds
is called the tube current-time product or mAs. The mAs is proportional to the total
number of x-ray photons the x-ray tube produces. The higher the mAs, the greater the
number of x-ray photons produced and the higher ionizing radiation absorbed dose to
the patient. The selection of a mAs value has a direct effect on the image noise and the
radiation dose to the patient. Lung CT scanning protocols use mAs values that ensure
adequate image quality with the lowest amount of radiation dose to the patient or sub-
ject being scanned.
Modern x-ray CT scanners use a bow tie-shaped x-ray beam filter to shape the beam
so that all the detector elements in the detector array see a more uniform number of
photons per unit area striking them.1 The use of additional x-ray beam filtration can also
modify the spectrum and intensity of the x-ray beam before it is transmitted through
the patient. The addition of 5–10 mm thick aluminum filters in x-ray CT imaging will
decrease the amount of lower energy x-rays that cannot penetrate the thickness of the
thorax and contribute to the formation of the image, and also increase the fraction of
higher energy photons that can reach the x-ray detector if they are not scattered or
absorbed.1
It is now possible to control the kVp and mAs on modern CT scanners in realtime,
and these values can be adjusted to provide a consistent signal-to-noise ratio in the CT
image and a lower and more uniform radiation dose to the imaged tissues. These are
referred to as tube kilovoltage modulation and tube current modulation (Fig. 2.6). The
use of mA modulation in lung CT AI work has been successfully implemented recently.
The use of kVp modulation is an issue in lung CT AI work when CT image voxel den-
sity values need to be consistent across time and patients. The varying kVp will change
the x-ray beam energy spectrum, hence the CT image contrast and voxel density values
will vary in the lung in a way that is challenging to correct at the present time.
In summary, the important factors regarding the operation of the x-ray tube that
need to be factored into a CT scanning protocol include focal spot size, tube current-
time product (mAs), peak kilovoltage (kVp), additional x-ray beam filtration, tube cur-
rent modulation, and kilovoltage modulation.
CT X-ray Beam Shape and Energy Spectrum
The x-ray photon beam has several properties that impact scan time, image contrast,
and ionizing radiation dose. Beam intensity and cross-sectional size of the photon beam
determine how much tissue can be imaged per unit time. For a given beam intensity,
the larger the cross-sectional size of the beam, the shorter time it takes to scan the
lungs. The early CT scanners had very narrow pencil or rod-shaped photon beam cross-
sections. This limitation was due to the reconstruction algorithms that were used, as well
as limitations in the x-ray tube power and available x-ray detectors. The scan times were
very long. The size of the x-ray beam cross-section has steadily increased over time as
corresponding increases in x-ray tube output, x-ray detector array size, and computer
power have been able to deal with large amounts of data per unit time coming from the
x-ray detectors.
20 DEVELOPIING THE DIGITAL LUNG
Fig. 2.6 Modern x-ray dual source multidetector CT scanner, Siemens SOMATOM FORCE CT
scanner, showing the patient table in relationship to the large aperture of the CT gantry. Courtesy
Siemens Healthineers.
The x-ray beam photon energy spectrum that irradiates the patient is a major factor
in the image contrast of the final CT image. This energy spectrum needs to be carefully
selected so that adequate image signal to noise ratio and contrast to noise ratio are pres-
ent in the lung CT images at the lowest possible radiation dose. The energy spectrum
needs to be the same in order to compare one patient to another and to assess the same
patient at multiple time points. This is an important concept in quantitative CT of the
lung because the values of the linear absorption coefficients, which determine the value
assigned to each voxel in the image, are a function of the x-ray beam energy spectrum.
To be able to compare voxel values over time from the same lung, or between lungs
2—THREE-DIMENSIONAL (3D) DIGITAL IMAGES OF THE LUNG 21
of different individuals, the x-ray energy spectrum needs to be the same. The photon
energy spectrum for x-ray CT scanners is determined by the peak kilovoltage applied
across the x-ray tube cathode and anode, and the type and thickness of metal filters
placed in the path of the x-rays emitted from the x-ray tube anode. Higher x-ray tube
peak kilovoltage (kVp) will shift the peak energy and the entire energy spectrum toward
higher energies, and also increase the number of x-ray photons generated for a given
mA setting. Lower peak kilovoltage will shift the peak energy and the energy spectrum
to lower energies and decrease the number of photons generated for a given mA setting.
The addition of metal filters between the x-ray tube anode and the patient can further
shape the energy spectrum without increasing the peak energy by filtering out lower
energies that are not able to penetrate the patient and reach the x-ray detectors. The peak
kilovoltage and metal filters need to be carefully selected to optimize image quality and
minimize the radiation dose to the patient or subject.
X-ray CT Detectors
The x-ray CT photon detectors that have been used since the first commercial CT scan-
ners made in the 1970s to the current generation of CT scanners use energy integrating
detectors (EID). The total energy of one or more photons that a single EID detec-
tor element measures during the measurement time are integrated to provide a total
energy signal. This signal is often produced by many different photons of different ener-
gies. There are new commercial CT scanners in development that use photon-counting
detectors (PCD). The photon-counting CT scanners detect each photon and its energy.
The PCD CT scanners compared to the current EID CT scanners have the potential
to further reduce radiation dose, increase image contrast and spatial resolution, correct
beam hardening artifacts, improve CT intravenous contrast media enhanced imaging,
and create new quantitative metrics for lung CT AI.8
The x-ray CT detectors are located opposite the x-ray tube. The x-ray CT detectors
detect photons that are not scattered or absorbed by the tissues. These photons are trans-
mitted through the tissue, without being scattered or absorbed, and then are detected
by the x-ray detectors. The different tissues scatter/absorb the x-ray photons differently,
so the number of and energy of x-ray photons that impact the x-ray detectors is related
to the thickness and specific composition of the tissues imaged. The EID type detector
elements are typically made from high-density ceramics containing rare earth materials,
for instance, gadolinium oxysulfide (Gd2O2S).1 The ceramic material is at the front of
the individual detector element and absorbs the transmitted x-ray photons coming from
the patient. Visible light is emitted by the ceramic material. This light is detected by an
array of electronic solid-state photodiodes that register the location and total energy of
the absorbed photons during a very short measurement time interval.1 The electronic
signal from the photodiodes is then digitized by an analog to digital converter and these
digital signals are sent to the CT scanner’s computer.
The physical size and shape of the detector array are determined by the shape of the
x-ray beam. The current x-ray CT scanners use a narrow-angle cone beam, so the x-ray
detector is a two-dimensional curved surface. The detector curvature is optimized to
minimize differences in path length from the x-ray source to a given detector element.
The size of the individual x-ray detector elements determines the maximum spatial reso-
lution possible for the CT scanner. The rotation speed of the CT x-ray tube detector
22 DEVELOPIING THE DIGITAL LUNG
pair determines how fast the thorax can be scanned; all other factors held constant. The
current generation of commercial CT scanners have a 360-degree rotation time as short
as 0.25 seconds.
CT Gantry
The CT gantry has a large aperture at its center for the patient and the patient table to
transit (Fig. 2.6). The modern CT gantry consists of two concentric cylinders, slip ring
design, where the x-ray tube, x-ray detectors, and electronics, including the high voltage
generator, rotate around the patient continuously. The control signals and scan data are
transmitted from the rotating cylinder to the outer stationary cylinder (Fig. 2.7). The
x-ray tube and corresponding detector array maintain an accurate fixed alignment with
each other as they rotate around the patient. The x-ray photons that pass through the
object are detected by the corresponding detector array located opposite the x-ray tube
that is aligned with the detector array. Most CT scanners have a single x-ray tube and
corresponding x-ray detector, but Siemens has made dual source CT scanners for over
15 years, which have two x-ray tubes and two corresponding x-ray detectors 90 degrees
apart (Fig. 2.8). The current third-generation dual-source CT scanners have effective
scan times that are half the scan times of similar single-source CT scanners. The dual-
source CT scanner enables dual-energy CT imaging as well.
Fig. 2.7 Diagram of the concentric cylinder, or slip ring arrangement, of a spiral CT scanner. The
inner ring rotates continuously around the patient while the outer ring is stationary.
2—THREE-DIMENSIONAL (3D) DIGITAL IMAGES OF THE LUNG 23
x-ray tube 1
2D detector array 2
x-ray tube 2
2D detector array 1
Fig. 2.8 Multidetector dual source CT (MDCT) scanners have two x-ray tubes and correspond-
ing x-ray detector arrays that both rotate continuously around the patient. There is a 90-degree
offset between the two x-ray tubes. When both x-ray systems are collecting data, the data col-
lection rate is doubled for each 180 degrees of rotation. This enables very short scan times that
are critical in cardiac CT imaging and combined cardiac and lung CT imaging. The maximum
scan field of view of x-ray tube 1 is typically greater than the maximum scan field of view of x-ray
tube 2 due to the constraints on space. The maximum field of view for dual acquisitions is limited
to the maximum field of view of x-ray tube 2. Dual acquisitions can be used to increase data
collection rates by a factor of two or to obtain acquisitions at two different x-ray tube kVps to
generate dual energy CT scans.
protocol. Pitch is defined as the table feed distance per 360-degree rotation of the x-ray
tube and x-ray detector array divided by the z-axis width of the x-ray detector array.
A table speed of 80 mm/second, 360-degree gantry rotation time of 0.5 seconds, and
a z-axis detector array width of 40 mm would result in a spiral CT scanning pitch of
1.0. A pitch of 1.0 in spiral CT scanning mode generates CT images that are similar to
contiguous axial scanning mode (see Scanning Modes). The radiation dose is inversely
proportional to the pitch. Pitch values higher than 1.5 enable faster scanning, lower
radiation dose, and lower image quality. Pitch values less than 1.5 results in slower scan-
ning, higher radiation dose, and higher image quality (Fig. 2.9). The pitch value usually
chosen for lung CT AI are 1.0 or very close to it.
Scanning Modes
The two CT scanning modes available for lung CT are spiral mode and axial mode.
Spiral mode is the main mode that that is currently used in lung CT AI protocols.
Spiral mode involves moving the CT table that the patient rests on through the CT
gantry aperture as the x-ray tube and x-ray detector array rotate continuously around
the patient, see Fig. 2.9. Axial CT scanning mode keeps the patient table station-
ary as the x-ray tube and x-ray detector array rotate around the patient and, after a
360-degree rotation, the CT table moves the z-axis length of the x-ray detectors, and
another set of axial mode CT images are obtained. Axial CT scanning mode was the
first CT scanning mode that began in 1972 with the first EMI head CT scanner.
Spiral CT scanning began in the late 1980s and early 1990s. Both spiral and axial CT
scanning modes are used today depending on the CT scanner being used and the body
part being scanned.
Collection of a Scanned Object’s Projection Data
The primary objective of the CT AI agent that controls the actual scanning is to col-
lect 1D projections from the x-ray detector array. The precise and accurate produc-
tion of these 1D projections is very important in both visual and quantitative lung
CT AI.
Image Reconstruction
The 1D projections that the CT scanner generates are considered forward projections of
the object being scanned. These projections are log-transformed since the x-ray attenu-
ation process is an exponential one, and the log transformation makes the reconstruc-
tion process a linear problem rather than an exponential problem.1 The linear problem
is easier to solve. The computer can reconstruct the object by back projection of the
log-transformed projections to reconstruct the object of interest (Fig. 2.10). This back
projection process is an inverse problem that attempts to reconstruct a 2D object from a
set of 1D projections that are obtained at different angles of the x-ray tube and detector
array. Simple back projection will cause blurring of the original object, (Fig. 2.10). The
blurring can be corrected by using a deconvolution kernel.1 This deconvolution kernel
0°
270° 90°
180°
Fig. 2.10 Process of unfiltered back projection and why unfiltered back projection needs a cor-
rection factor or filter to create an accurate 2D image of an object from a set of 1D projections.
The is an example of four 1D projections of a 2D circular object that are back projected and form
a 2D star pattern, not the original 2D circular object.
26 DEVELOPIING THE DIGITAL LUNG
is designed to undo the blurring caused by simple back projection. The deconvolution
kernel is combined with each projection in a mathematical process called convolution.
The original projections are each corrected by this convolution process. This process of
correcting the blurring in the projections, and then back projection of the corrected
or deblurred projections, is referred to as filtered back projection (FBP) (Fig. 2.10). It
should be noted that inverse Radon transforms, as well as Fourier transforms, can be
used in both unfiltered and FBP image reconstruction to recover a 2D image from a
set of 1D projections of the object.1 There are several available deconvolution kernels
for x-ray CT scanners, and they have a big impact on the value of the CT numbers
assigned to each image voxel. These CT numbers are the cornerstone of all AI applica-
tions involving x-ray CT of the lungs. Careful selection of the reconstruction method
and reconstruction kernel in the CT scanning protocol is very important to achieve
optimal results in lung CT AI.
CT system optics include information about the geometry of the CT system. This can
include the distance between the x-ray tube and the gantry isocenter, distance between
the x-ray tube and the detectors, x-ray tube focal spot size and shape, size and shape of
the individual detector elements, and the size and shape of the x-ray beam and detector
array.12 The more detailed the MBIR model of the CT imaging process, the greater the
demands will be on computational resources. There is a tradeoff in the robustness of the
MBIR model of the CT imaging process and the reconstruction times needed in a busy
clinical practice.
The reduction in image noise for a given radiation dose is the compelling advantage
of IR versus FBP methods of image reconstruction. The main disadvantage of IR is
that it is a non-linear image reconstruction method, so the use of image quality metrics,
such as SNR, CNR, MTF, and NPS, assume a linear and stationary imaging system. It
is important in lung CT AI to assess image quality objectively. Improved methods to
accurately assess MBIR CT images are needed.12
Scan Field of View (SFOV), Display Field of View (DFOV), and
Reconstruction Matrix Size
Several other important parameters need to be optimized for lung CT AI work. These
include scan field of view (SFOV), display field of view (DFOV), and the reconstruc-
tion matrix that is superimposed on the DFOV. The SFOV is the x–y-plane area that is
scanned and is greater than or equal to the DFOV (Fig. 2.11). The DFOV is the actual
x–y-plane area that is used to reconstruct a CT image (Fig. 2.11). The DFOV can be
less than or equal to the SFOV. You might wonder, why bother with the DFOV? The
additional parameter that needs to be considered is the size of the reconstruction voxel
matrix that is superimposed on the DFOV. The current reconstruction matrix size is 512
× 512, but 1024 × 1024 is available for some CT scanner models. The size of the recon-
struction matrix has a big impact on the speed at which images can be reconstructed and
the size of these images. The larger the matrix, the greater the computing and storage
resources necessary to reconstruct, display, and store the images. However, the larger
the matrix size, the smaller the voxel size; hence, the greater the in plane or x–y-plane
resolution. For a given x-ray dose, the smaller the voxel size, the greater the image noise.
For example, if you doubled the matrix size, you would need to increase the x-ray dose
by a factor of four to achieve the same signal-to-noise ratio for the measured Hounsfield
number assigned to each of the smaller voxels in the reconstructed image. The x–y-plane
dimensions of the reconstructed voxel for a 512 × 512 reconstruction matrix are driven
by the following relationship:
X Y DFOV/512
For example, if the DFOV is 350 mm, the x–y dimensions of the reconstructed
voxel would be 0.68 mm. If the DFOV is 500 mm, the x–y dimensions of the recon-
structed voxel would be 0.98. This is a 50% increase in the x–y dimensions of the
reconstructed voxel. For these reasons, the DFOV in lung CT AI work is optimally
chosen to include only the lungs to keep the DFOV as small as possible, typically
32 cm. It is also important to use the same DFOV and matrix size when multiple
scans are obtained over time on the patient. It should be noted that all scanners can
reconstruct the acquired 1D projections using various DFOV sizes without exposing
the patient to additional radiation. For example, there can be one reconstruction for
lung CT AI work, 32 cm DFOV, and a larger DFOV to be sure to assess all of the
chest wall structures, 45 cm DFOV.
Hounsfield Units and the CT Voxel
After the CT image reconstruction process is completed, we have a 3D representation
of the thorax containing millions of voxels. Each voxel is assigned a value that represents
the average x-ray photon linear attenuation coefficient of the tissue contained within the
voxel. These are small numbers with the units cm−1 and symbol μ. The linear attenuation
coefficient of liquid water, μwater, using 50 keV photons is 0.214 cm−1.5 The linear attenu-
ation of dry air, μair, using 50 keV photons is 0.000290.5
Since the beginning of CT scanning in the early 1970s, there has been a desire
to express these linear attenuation coefficients in a manner that is simple and easy to
understand. This resulted in the Hounsfield unit (HU) scale, named after Sir Godfrey
Hounsfield who invented the first human x-ray CT scanner in 1972.1,2
The Hounsfield Unit is defined using the linear attenuation coefficients of air, water,
and the tissue in question, as follows:
HU 1000 (( tissue water )/( water air ))
Where HU is the tissue density in Hounsfield Units. The Hounsfield scale range
assigns a value to air of −1000 HU and a value to water of 0 HU. The scale typically runs
from −1000 HU to 3095 HU for a 12-bit CT scanner.1 The range of densities in the
2—THREE-DIMENSIONAL (3D) DIGITAL IMAGES OF THE LUNG 29
normal lung range from −1000 HU to 40 HU. The normal lung is comprised of 80% air
and 20% tissue, so it is not surprising that the normal lung density range is −843 HU in
adult males and −824 HU in adult females.13
It should be noted that tissue density measured in HU numbers can be converted to a
more familiar measure of tissue density, mass per unit volume in units of grams per liter,
by adding 1000 to the HU value. This assumes the density of air is 0 g/L and the density
of water is 1000 g/L. For example, the HU number for water is 0, and if we add 1000 we
get the density of water, which is 1000 g/L or 1 g/cm3.
Visually Display of Lung Images
The next step is deciding how best to visually interact with the 3D digital images of the
thorax that contains the lung. The most common method is to view the entire anatomy
contained in the DFOV. This includes structures with air density, such as air in the
trachea with a density of −1000 HU, to cortical bone with a density of 1800 HU. Here
we assume a high-quality medical display capable of displaying 2048 grey levels with a
luminance range of 0.5 to 2000 cd/m2.14 If the range of CT numbers is assumed to be
4096, or 12-bits of information, then how does the human eye evaluate the anatomy
when the human eye at best can distinguish 900 shades of grey on a state of the art
medical display?14 The answer is that we look at a subset of the grey levels and, as a
result, we need to look at images that are optimized to look at lung tissue, soft tissues,
or bony structures. Here we are interested in looking at the human lung, so we are
interested in looking at CT density values between −1000 HU to 40 HU. The range, or
window width (WW), of HU values is typically set at 1500 HU instead of 1040 HU
for viewing x-ray CT images of the lung. The wider WW decreases image contrast,
which, in the case of the x-ray CT images of the lung, yields a more acceptable image
contrast. A WW of 1040 is selected to assess very subtle differences in tissue density.
The mid-point of this 1040-1500 scale or window level (WL) is usually set between
−500 and −600 HU. Fig. 2.12 shows an axial x-ray CT image of the thorax with differ-
ent WW and WL settings. There are other enhanced visualizations available to assess
A B
Fig. 2.12 2D axial image of the thorax with WW = 1500 and WL = −500, which is optimized for
looking at the lungs (A), and WW = 350 and WL = 50, which is optimized to look at the soft tissue
structures in the mediastinum and chest wall (B).
30 DEVELOPIING THE DIGITAL LUNG
A B
Fig. 2.13 Example of multiplanar lung CT images of the thorax in the sagittal (A) and coronal (B)
planes using WW = 1500 and WL = −500. This is the same normal CT study as Fig. 2.12. The
intersection of the horizontal and vertical lines in each image (arrows) are identifying the same
CT voxel, which is near the bifurcation of the trachea into the right and left main-stem bronchi.
the human lung, including multiplanar reformatted images in the sagittal and coronal
planes (Fig. 2.13). There are also maximum intensity projections (MIPS) and minimum
intensity projections (MinIPS) available in axial, coronal, and sagittal planes. The MIPS
help in identifying lung nodules separate from lung blood vessels. The MinIPS are help-
ful for identifying air-containing structures, such as emphysema, and normal or dilated
airway lumens. Two new lung CT image formats that have recently been introduced by
the lung CT AI company VIDA Diagnostics Inc. in Coralville, Iowa. These new image
formats are topographic multiplanar images that are optimized to view the airway tree,
airway tMPR (Fig. 2.14), and the subpleural space of the lungs, subpleural images.
We will use representative examples of lung CT images in this book that visually
show the lung disease features that are being assessed by corresponding lung CT AI
methods.
Quantitative CT Metrics
There are a number of quantitative x-ray CT metrics that have been developed in lung
CT AI studies of normal and diseased lung structure and function, and these will be
discussed in Chapters 4 through 7. These include measurements of lung volume, lung
tissue intensity, lung tissue texture, air trapping, and lung tissue biomechanics.
CT Scanning Protocols
CT scanning protocols of the lung are designed to achieve accurate and precise CT
images of the lung, with the needed spatial and contrast resolution, using the small-
est amount of ionizing x-ray radiation dose to the scanned tissues. These CT scanning
2—THREE-DIMENSIONAL (3D) DIGITAL IMAGES OF THE LUNG 31
A B
Fig. 2.14 Normal coronal image of the thorax (A) and a corresponding airway topographic MPR
image that is optimized to show the trachea, main-stem bronchi, lobar bronchi, and segmental
bronchi in a single image (B).
protocols are very detailed and are critical to the field of lung CT AI. The important
elements of these CT protocols will be discussed in Chapter 3 and will build on the
discussions in this chapter. These CT protocols can be pre-programmed into the current
generation of CT scanners available in the 2020 s. This greatly simplifies the CT scan-
ning process for clinical patients and research subjects.
in deciding how much radiation dose to give a patient or research subject for an x-ray
CT scan of the thorax.16 The ALARA principle means we should give only as much
radiation as necessary to achieve the goal of the x-ray CT scan.16 It is important to fol-
low the ALARA principle when designing lung CT scanning protocols.
CTDIvol is a method that is used to try and quantitate the absorbed radiation dose
to the thorax that a patient receives from a CT scan of the thorax.15 The units of the
CTDIvol are Gys. A typical CTDIvol dose from an x-ray CT scan of the thorax is 17
mGy. A low-dose CT scan is considered one in which the CTDIvol is less than 3 mGy.18
The CTDIvol value is reported by the x-ray CT scanner.
The length of the scan (L) in cm in the z-axis is also important to take into consider-
ation.15 The typical L value for a tall adult male is 30 cm. The dose length product (DLP)
is the product of the CTDIvol multiplied by L and has units of mGy-cm. The DLP is
also reported by the CT scanner; the CTDIvol and DLP are included in the x-ray report.
The effective dose for a chest CT scan is a stochastic method of assessing what the
whole-body radiation dose to a “generic patient” would have to be to impart the same
risks of inducing cancer as the absorbed dose to the chest using a given CTDIvol, L, and
conversion factor (k). k varies by the body part that is being scanned. The chest CT has
a k factor equal to 0.014 mSv/mGy-cm.15 The effective dose can be calculated using the
following formula:
for the smaller effective diameter. Further discussions of SSDE go beyond the scope of
this book, the interested reader is referred to the literature.15,20,21
(ART), as opposed to a filtered back projection method. The need for a water bath made
it impractical to scan the thorax and abdomen of a patient. The EMI Mark 1 x-ray CT
scanner was introduced at the 1972 Annual Meeting of the Radiological Society of
North America and orders were taken.22 EMI was the only vendor in 1972, but by the
1973 Annual Meeting of the RSNA, Ohio Nuclear had announced the development of
a clone of the EMI Mark 1 scanner with scan times reduced from 5 minutes to 2 min-
utes, and image matrix size increased to 120 × 120. EMI was still selling the Mark 1 but
announced the improved Mark 2 scanner at the 1973 RSNA meeting. During this time,
Georgetown University sent Robert Ledley to EMI in the United Kingdom to look into
acquiring an EMI Mark 1 head CT scanner for the university.22 Upon his return, Ledley
convinced Georgetown University that he could develop a head and body CT scanner,
and obtained approval to do this. Ledley was able to develop his new whole-body CT
scanner in just 18 months.22
The first whole-body x-ray CT scanner was the Automatic Computerized Transverse
Axial (ACTA) Scanner. It was developed by Robert Ledley, MS, DDS, at Georgetown
University in Washington, DC, in 1973 (Patent No. 3,922,552).23–25 Robert Ledley was
a biophysicist and dentist at Georgetown University who reported the first CT images
of the lungs in 1975 in the journal Radiology.26 It is helpful to go through the com-
ponents of this scanner to see where x-ray chest CT technology began, so as to better
appreciate the technical advancements that have led to the current highly capable chest
x-ray CT scanners of the 2020 s.
The ACTA scanner consisted of a gantry that housed the x-ray tube and two x-ray
detectors, with the patient positioned on a translatable table located in the center of the
gantry aperture between the x-ray tube and the x-ray detectors. The x-ray tube beam
was collimated to form a pencil beam that exposed 7.5 mm of two separate x-ray detec-
tors, with each detector separated by 3 mm. This would be described as a multidetector
x-ray CT scanner since two adjacent CT images were obtained for each 180-degree
rotation of the x-ray tube and x-ray detector array. The x-ray detectors were composed
of a sodium iodide crystal attached to a photomultiplier tube. The x-rays would produce
visible light when they interacted with the iodide atoms in the sodium iodide crystal
and the photomultiplier tube would amplify the light signal. Then, this signal would be
digitized and stored in the computer. The sampling window was 5 ms to measure the
x-ray intensity at one point of the projection. The x-ray tube and crystal assembly would
translate in a line across the patient making 160 of these 5-ms measurements to obtain
a 1D projection of the object being scanned. The x-ray tube and detector array would
then rotate 1 degree and the process would repeat itself. This process was done until the
x-ray and detector array had rotated 180 degrees around the patient. This produced 180
projections of the patient. The combination of each 160 points collected for each projec-
tion, multiplied by 180 projections, would result in a total of 28,800 measurements for
each reconstructed 2D image of the object being scanned. The 180 projections of the
object were then reconstructed using the Fourier slice theorem.23 The reconstructed 2D
image was made up of 25,600 points forming a 160 × 160 image matrix.23 The spatial
resolution of the CT images in the z-direction was 7.5 mm. The x–y resolution was
1.5 mm. There were two scanning modes available: short pass, 24 cm, and long pass,
48 cm. This would be similar to the SFOV, as described previously. In the long pass, the
x-ray tube and detector array would linearly translate 48 cm across the patient to obtain
2—THREE-DIMENSIONAL (3D) DIGITAL IMAGES OF THE LUNG 35
one complete 1D projection and was used for scans of the thorax, abdomen, and pelvis.
This process would be repeated 180 times at 1-degree increments. This scan would take
6 minutes to obtain two 2D axial images of the thorax. The short pass scan mode was
used to scan the head, neck, and extremities. There was doubt as to how useful the CT
scans of the thorax would be given the 6-minute scan time to obtain two images—total
examination time to scan a 30-cm adult thorax would be at least 90 minutes. The respi-
ratory and cardiac motion was well-tolerated and the CT images presented an average
appearance over the time of the scan. The first images of the thorax showed a lung cancer
nodule in the left lung; the trachea, transverse aorta, and superior vena cava were identi-
fied, along with the muscles and bones in the chest wall.23
It can be seen from the detailed description of the first whole-body x-ray CT scan-
ner that certain design features would be retained and others improved upon; if the goal
was collecting more projection data, in a shorter period, with higher spatial resolution,
while also maintaining the high intrinsic contrast resolution of x-ray CT compared to
x-ray projection radiography. The basic layout of a gantry housing the x-ray tube and
corresponding detector array with the patient positioned between the x-ray beam and
detector is retained, along with having the patient lay quietly on a table that traverses the
central aperture of the CT gantry. Improvements that needed to be made over the EMI
Mark 1 and ACTA CT scanner designs included increased size of the x-ray beam with
corresponding increases in the overall size of the detector array, increased number of
detector elements, increased speed at which 1D projections are obtained, and improve-
ments in electronics and computing power.
[4] Enough.