Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
Dissertation
Presented in Partial Fulfillment of the Requirements for the Degree of Doctor of
Philosophy in the Graduate School of The Ohio State University
By
Graduate Program in
2017
Dissertation Committee:
Md Asiful Islam
2017
Abstract
Magnetic resonance imaging (MRI), X-ray computed tomography (X-ray CT) etc. are
high accuracy imaging modalities but lack portability and cost effectiveness. In contrast,
microwave tomography has the potential for real-time portable imaging due to its simpler
hardware and lower costs. However, conventional microwave tomography has a number
of limitations: a) lengthy times to obtain an image and b) accuracy imbalance between the
Further, on-body microwave imaging suffers from several additional challenges. Two of
them are, a) lack of reference measurements needed to calibrate the data, b) uncertainty in
the positions of the on-body antenna sensors. In this dissertation, we address several of
Newton algorithm to accelerate image reconstruction (more than 25 times faster than the
existing methods) and therefore enable real-time monitoring, 2) to balance real and
relies on a combination of preset permittivity values. For the first time, this algorithm
mitigates, almost entirely, the imbalance of the real and imaginary permittivities, even for
algorithm that images the cross-section of a human torso. To do so, we propose artificial
ii
neural networks (ANNs) to establish the unknown relationship between permittivity and
overcome the unavailability of the reference data. Also, Discrete Fourier Transform
without the need for matrix-inversion. This algorithm exploits ‘reciprocity’ and is among
the first demonstrations of the frequency domain analog of the well-known time reversal
(TR) algorithm of time domain microwave imaging. Overall, this 3D imaging method is
shown to be more robust against data error (works for SNR as low as -5dB), making it
suitable for body-worn microwave imaging. Throughout the dissertation, imaging results
with synthetic and experimental data are provided to validate the methods and
algorithms.
iii
Dedicated to my mother, father, wife, daughter, son and sister…
iv
Acknowledgments
First of all, I would like to thank Allah (swt), the almighty, for giving me the ability to
complete this work. Without His will and help, it would not be possible ever.
for his support and guidance and being patient with me throughout my PhD life. His
I would like to gratefully thank Professor Asimina Kiourti for helping and guiding me
in different phases of my work and playing the role of co-advisor. Special thanks will go
to Professor Niru K. Nahar for always being there to help me out while taking important
decisions. Her role can never be forgotten and I wish her success in each step of her life.
Also, I would like to thank my committee member Professor Fernando Teixeira for
Finally, I would mention and thank my friends for their support during my difficult
times. I especially would like to acknowledge Shah Chowdhury, Cedric Lee, Ushe
Rashed Zuboraj, Syed Saqueb, Anas Abumunshar, Carlos Viteri, Kai Ren and Shubhendu
Bhardwaj for their valuable friendships. I wish them the very best in their lives.
v
Vita
Publications
Journal Papers:
M. A. Islam, A. Kiourti, and J.L. Volakis, “A Modified Gauss-Newton Method for High-
Speed Microwave Imaging with Near-Field Probes,” Microwave and Optical Technology
Letters, Feb. 2017.
M. A. Islam, A. Kiourti, and J.L. Volakis, “A Novel Method of Deep Tissue Biomedical
Imaging Using a Wearable Sensor,” IEEE Sensors Journal, 2015.
Conference Presentations:
M. A. Islam, A. Kiourti, and J.L. Volakis, “A Novel Body-Worn RF Sensor for Deep
vi
Tissue Imaging,” 2015 IEEE Int. Symp. Antennas and Propagation (APS/URSI),
Vancouver, Canada, 19-25 Jul. 2015.
M. A. Islam, A. Kiourti, and J.L. Volakis, “Conformal Sensor Accuracy for Deep Tissue
Biomedical Imaging,” 2015 IEEE International Symposium on Antennas and Propagation
& USNC/URSI National Radio Science Meeting, Boulder, CO, Jan 4-7, 2016.
M. A. Islam, A. Kiourti, and J.L. Volakis, “A Novel Body Worn RF Sensor for Deep
Tissue Imaging,” 2015 IEEE MTT-S 2015 International Microwave Workshop Series on
RF and Wireless Technologies for Biomedical and Healthcare Applications (IMWS-Bio),
Taipei, Taiwan, Sep 15-19, 2015.
Fields of Study
vii
Table of Contents
Abstract ............................................................................................................................... ii
Acknowledgements ............................................................................................................. v
Vita ..................................................................................................................................... vi
List of Tables...................................................................................................................... xi
4.1 Method.................................................................................................................... 42
Chapter 5: Imaging with On-body Antennas Employing Artificial Neural Networks ..... 58
ix
5.3 Imaging Results ...................................................................................................... 67
References ......................................................................................................................... 95
x
List of Tables
xi
List of Figures
Figure Page
Figure 1.2 Envisioned MWT for on-body applications as compared to MRI and X-ray
CT........................................................................................................................................ 6
Figure 3.1 Geometrical configuration of the imaging problem (top view) ...................... 25
Figure 3.2 (a) Simulation set-up, and (b) experimental set-up used to validate the
Figure 3.3 Power distribution inside the imaging domain. The red region denotes the
Figure 3.4 Reconstructed images with synthetic data: (a) noiseless, (b) SNR = 50dB, (c)
SNR = 40 dB, (d) SNR = 30 dB. The dashed circular lines indicate the actual position of
xii
Figure 3.5 Three different imaging scenarios for the measurement setup. PTFE rods
located: a) one near boundary, b) one in the center and, c) two near the opposite
boundaries ......................................................................................................................... 33
Figure 3.6 Reconstructed images using: (a)-(c) simulated noiseless data, (d)-(f) measured
data .................................................................................................................................... 37
Figure 4.2 Reconstructed images for one anomaly: (a) rel. permittivity (direct), (b)
conductivity (direct), (c) rel. permittivity (indirect), (d) conductivity (indirect). The
dashed circular lines indicate the actual position of the anomaly ..................................... 47
Figure 4.3 Reconstructed images for three anomalies: (a) rel. permittivity (direct), (b)
conductivity (direct), (c) rel. permittivity (indirect), (d) conductivity (indirect). The
dashed circular lines indicate the actual position of the anomaly. .................................... 48
Figure 4.4 Reconstructed images for two anomalies: (a) rel. permittivity (direct), (b)
conductivity (direct), (c) rel. permittivity (indirect), (d) conductivity (indirect). The
dashed circular lines indicate the actual position of the anomaly ..................................... 49
xiii
Figure 4.5 Reconstructed images for three anomalies: (a) rel. permittivity (direct), (b)
conductivity (direct), (c) rel. permittivity (indirect), (d) conductivity (indirect). The
dashed circular lines indicate the actual position of the anomaly ..................................... 50
Figure 4.6 Geometrical configuration of the imaging domain a) 3D forward model and b)
Figure 4.7 a) Cross-sectional view of the flow models: case-I (left) and case-II (right),
b)reconstructed image of the entire flow model c) reconstructed image of the phase air
and d) reconstructed image of the phase water. The black circles correspond to the actual
Figure 5.5 Imaging of tumors for four different cases in the lung background. The dotted
circle shows the actual position and shape of the tumors. Dimensions are in meters ...... 69
Figure 5.6 Imaging of tumors for four different cases in the lung background. The dotted
circle shows the actual position and shape of the tumors. Dimensions are in meters ...... 70
xiv
Figure 5.7 a) 3D human body model, b) 2D cross section of the model ......................... 71
Figure 5.8 Image reconstruction on an elliptical grid for a human torso with non-exact
Figure 6.1 a) Electric line source near a circular PEC cylinder (top view), b) Line source
radiation and measurement points on the surface of the cylindrical imaging domain ...... 76
Figure 6.3 Pattern of a dipole: a) 3D pattern, b) a 2D cut in the elevation plane from the
3D pattern .......................................................................................................................... 81
Figure 6.6 2D cuts from the 3D image, a) XY plane, Z=3cm, b) XY plane, Z= 0cm, c)
Figure 6.7 Images with different SNR values, left column: proposed method, right
Figure 6.8 (a)-(b) Tumor growing in size, (c) reconstructed image of tumor growth ...... 89
xv
Chapter 1: Introduction
region of interest. One of the antennas is excited at one time and the others receive the
field radiated by the transmit antenna, see Fig. 1.1. An estimate for the cross-sectional
Object to be
imaged
Antennas
1
19
three-dimensional images instead of cross-sectional images. Mathematically, the
approaches to solve the reconstruction problem have been proposed. Among them two
strong scatterers within the domain of interest (DOI), if any. The transmit antennas are
excited by a wideband signal, and data are collected and processed in time-domain to
obtain the scatterer locations. One example application is medical diagnostics, e.g.,
the actual permittivity of the DOI and can be divided into two categories. The first
approach exploits the first order Born or Rytov approximation [5]-[6]. This method is
[10]. Typically, the scatterers are expected to be of low contrast with respect to the
background (viz. where the Born or Rytov approximation is valid). Among other
techniques, [10] proposed the ‘Distorted Born Iterative Method (DBIM)’ for cases
where the first order Born or Rytov approximation does not hold. Still, DBIM does not
work well for cases where the permittivity of the scatterer is substantially different
2
techniques combines imaging with an optimization algorithm [12]-[13]. This is based
on the minimization of the squared norm of the difference between the measured
boundary electric fields (E-fields) and the calculated boundary E-fields. Of importance
is that this category of techniques mostly does not have dielectric contrast limitations.
Recently, an MWT algorithm has been proposed [14] which can potentially image
the location of a strong scatterer (like tumor or blood clot) in a complex imaging
domain like the human head. Though it can be applied in an out-of-hospital scenario,
it has several drawbacks, such as, it only recovers the strongest scatterer (not any other
scatterers). Also, fixed antenna positions have been utilized with no indication for on-
body applications.
medicine. Among the wide variety of applications, in this thesis the interest is mostly
in medicine. The applicability of MWT for medical purposes can be motivated by the
large variation in the tissue permittivities, see Table I. Due to the large variation, high
contrast between different tissues could be expected. However, in order to utilize this
fact, absolute MWT images should be reconstructed. This would mean that the
nonlinear MWT reconstruction problem should be solved which would usually mean
3
In the past, endeavor has been made to extract the average permittivity of
underlying deep tissue by placing a set of probes on the body [15]. However, there are
several
permittivity value which is hard to relate to any healthy/non-healthy case, ii) shape
uncertainty of the body is not taken care of, iii) how the measurements were calibrated
is not obvious.
A later paper from the same group proposed a method to extract pixel-by-pixel
image of deep tissue [16] (overcoming the limitation (i) above). However, it still
4
1.1 The Aims and Organization of the Dissertation
In general, solving the MWT problem is a time consuming process due to several
Typically, MWT yields two images: real and imaginary parts of the complex
worse in quality as compared to the real part. This may cause ambiguity in decision
making. So, there is definitely a need for an imaging algorithm that deals with this
and Electrical Impedance Tomography (EIT) are the most popular surgery-free
technologies used to image tissues deep into the human body [17]-[20]. However,
these methods do not provide continuous real-time tissue characterization, as they are
not portable systems. Also, these technologies are expensive, implying major access
disparities. For continuous and low-cost monitoring of vital human body signals,
several portable body-worn sensors have been reported [21]-[23]. Such sensors may
5
Motivation
Existing Imaging Technology Future Imaging Technology
strong need to develop low-cost, portable devices for deep tissue imaging for use in
hardware reduction and portability of the existing imaging methods and what MWT
behind MWT has been discussed. The derivation of the classic inverse scattering
formulation starting from the ‘volume equivalence principle’ is shown. Also, it has
been shown that the inverse scattering formulation reduces to the same formulation to
6
that using Gauss-Newton method when first order ‘Born Approximation’ is applied.
Then some other methods in the literature are discussed. In chapter 3, the MWT
employed and basic formulation of this method is discussed. The reason why this
method yields slow convergence towards a reconstructed image is pointed out. Then a
solution based on the Taylor series approximation has been proposed. Simulation and
experimental results have been presented to validate the method. It has been shown
that the method yields almost 25 times faster convergence than a previous method.
Chapter 4 proposes a new technique to improve imaginary permittivity image and sets
up a balance between real and imaginary permittivity images. Unlike the ambiguous
situation with the regular image reconstruction, now, both the images carry the same
information causing more certainty to the user. Next, chapter 5 addresses the
pinpointing the challenges, an Artificial Neural Network (ANN) based approach has
Discrete Fourier Transform (DFT) has been employed to reduce the number of
electromagnetic waves has been proposed. The algorithm is derived starting from the
7
formulation of line source scattering by a perfect electric conductor (PEC) cylinder.
The original 2D algorithm is extended to 3D imaging taking into account the far field
pattern and orientation of the antenna. Of importance is that the algorithm is capable
of handling much lower signal-to-noise (SNR) ration (as low as −5dB) as compared to
conventional MWT algorithms. Also, it has been demonstrated that the algorithm can
8
Chapter 2: On the Theory of Microwave Tomography
reconstructed image. In this chapter, we will discuss the widely used formulations to
solve MWT problem: inverse scattering and optimization based methods. Also, we
will set up a connection between these two methods. Additionally, we will discuss the
Here, we will show, how MWT can be done from an inverse scattering point of
view. Let us assume that in the free-space environment (𝜀0, 𝜇0), the source (𝐽̅𝑖 )
̅̅̅0 , ̅̅
generates fields (𝐸 𝐻̅̅0 ). This source and field must satisfy Maxwell’s equations:
̅ × 𝐸̅0 = −𝑗𝜔𝜇0 𝐻
∇ ̅0 (2.1)
∇ × 𝐻0 = 𝐽𝑖 + 𝑗𝜔𝜀0 𝐸̅0
̅ ̅ ̅ (2.2)
When the same source radiates in the medium (having a material obstacle in the free-
̅ × 𝐸̅ = −𝑗𝜔𝜇𝐻
∇ ̅ (2.3)
̅×𝐻
∇ ̅ = 𝐽̅𝑖 + 𝑗𝜔𝜀𝐸̅ (2.4)
9
Taking (2.3)−(2.1) and (2.4)−(2.2):
̅ × 𝐸̅𝑠 = −𝑗𝜔(𝜇𝐻
∇ ̅ − 𝜇0 𝐻
̅0 ) (2.5)
̅ ̅ ̅ ̅
∇ × 𝐻𝑠 = 𝑗𝜔(𝜀𝐸 − 𝜀0 𝐸0 ) (2.6)
̅ × 𝐸̅𝑠 = −𝑗𝜔(𝜇 − 𝜇0 )𝐻
∇ ̅ − 𝑗𝜔𝜇0 𝐻 ̅𝑠 (2.7)
̅×𝐻
∇ ̅𝑠 = 𝑗𝜔(𝜀−𝜀0 )𝐸̅ + 𝑗𝜔𝜀0 𝐸̅𝑠 (2.8)
̅ and 𝑀
We define volume equivalent electric and magnetic current densities 𝐽𝑒𝑞 ̅𝑒𝑞
̅ = 𝑗𝜔(𝜀−𝜀0 )𝐸̅
𝐽𝑒𝑞 (2.9)
̅𝑒𝑞 = 𝑗𝜔(𝜇−𝜇0 )𝐻
𝑀 ̅ (2.10)
̅ × 𝐸̅𝑠 = −𝑀
∇ ̅𝑒𝑞 − 𝑗𝜔𝜇0 𝐻̅𝑠 (2.11)
̅×𝐻
∇ ̅𝑠 = 𝐽𝑒𝑞
̅ + 𝑗𝜔𝜀0 𝐸̅𝑠 (2.12)
Equations (2.11) and (2.12) state that, scattered electric and magnetic fields by a
̅ and 𝑀
material obstacle can be generated by equivalent volume current densities 𝐽𝑒𝑞 ̅𝑒𝑞
function, 𝐺̅ (𝒓̅, 𝒓̅′ ) of the background medium (free-space in this case) is known, we
10
̅ (𝒓̅′) ∙ 𝐺̅ (𝒓̅, 𝒓̅′)𝑑𝒓̅′
𝐸̅𝑠 (𝒓̅) = 𝜇 ∭ 𝐽𝑒𝑞
or, 𝐸̅𝑚𝑒𝑎𝑠 (𝒓̅) = 𝐸̅0 (𝒓̅) + 𝑗𝜔𝜇 ∭(𝜀(𝒓̅′)−𝜀0 (𝒓̅′))𝐸̅ (𝒓̅′) ∙ 𝐺̅ (𝒓̅, 𝒓̅′)𝑑𝒓̅′ (2.13)
Like any other inverse problem [24], MWT can be cast as an optimization problem.
Typically, the following cost function that has to be minimized, to solve the MWT
problem.
1
𝐶({𝜀}) = 2 argmin‖𝐸̅𝑚𝑒𝑎𝑠 − 𝐹̅ ({𝜀})‖2 (2.14)
{𝜀}
where, ||▪|| is the L2 norm of the argument, 𝐸̅𝑚𝑒𝑎𝑠 is the E-field measurement vector,
𝐹̅ ({𝜀}) is the forward scattering solution for a permittivity distribution, {𝜀} in the
imaging domain. 𝐸̅𝑚𝑒𝑎𝑠 is calculated by exciting one antennas at a time and measuring
the E-field from the other antennas. {𝜀} is found by minimizing eqn. (2.14) iteratively
Gauss-Newton method falls in the broad category of line search method. From
conventional line search method [25], the iterative process to determine {𝜀} from
11
{𝜀𝑘+1 } = {𝜀𝑘 } − 𝛼𝑘 {𝑝𝑘 } (2.15)
Here, 𝛼𝑘 is the scalar step length and {𝑝𝑘 } is the search direction which has to be a
descent direction of the functional, 𝐶({𝜀}) in (2.14). To be noted is that when {𝑝𝑘 } is
Where, [𝐵𝑘 ] is a positive definite matrix. To be noted is that, if we insert (2.17) into
12
In (2.18), {∇𝐶𝑘 } is the gradient calculated from (2.14), as shown in (2.19) below. Also,
[𝐵𝑘 ] is the exact Hessian, [𝐵𝑘] = [∇2𝐶𝑘] according to Newton’s method. However, in
From (2.14), the gradient and Hessian of C can be calculated at any point {𝜺𝑘 } as,
𝑇
({ }) ({ })
{∇𝐶𝑘 } = 𝜕𝐶 𝜀𝑘 ⁄ { } = − (𝜕𝐹 𝜀𝑘 ⁄ { }) (𝐸̅ 𝑚𝑒𝑎𝑠 − 𝐹̅ ({𝜀})) (2.19)
𝜕 𝜀 𝜕 𝜀
𝑇
2
({ }) 𝜕𝐹({𝜺𝑘 }) 𝜕𝐹({𝜺𝑘 }) 𝜕2 𝐹𝑗 ({𝜺𝑘 })
[𝐵𝑘 ] = 𝜕 𝐺 𝜺𝒌 ⁄ = ( ⁄ { }) ( ⁄ { }) − ∑𝑀𝑗=1 ((𝐸𝑗,𝑚𝑒𝑎𝑠 − 𝐹𝑗 ({𝜺𝒌 })) ⁄ )
𝜕{𝜺}
2
𝜕 𝜺 𝜕 𝜺 𝜕{𝜺}2
(2.20)
Now, to approximate the Hessian using Gauss-Newton method, only the first term of
(2.20) is retained,
𝑇
𝜕𝐹({𝜺𝑘 }) 𝜕𝐹({𝜺𝑘 })
[𝐵𝑘 ] ≈ ( ⁄𝜕{𝜺}) ( ⁄𝜕{𝜺}) (2.21)
𝜕𝐹({𝜺𝑘 })
We define ⁄𝜕{𝜺} = [𝐽]𝑀×𝑁 as the Jacobian/sensitivity matrix [3] (M is the
13
Inserting (2.19) and (2.21) into (2.18), our Gauss-Newton iterative process reduces to,
In (2.13), 𝐸̅𝑚𝑒𝑎𝑠 (𝒓̅) is the total field in the imaging domain which is unknown. So, in
order to solve the inverse problem, we have to solve two unknowns in (2.13): 𝜀(𝒓̅′)
and 𝐸̅𝑚𝑒𝑎𝑠 (𝒓̅). One convenient way to approximate 𝐸̅𝑚𝑒𝑎𝑠 (𝒓̅) and still get a reasonable
solution is to let 𝐸̅𝑚𝑒𝑎𝑠 (𝒓̅) ≈ 𝐸̅0 (𝒓̅′) which is the incident field and is known a priori.
If we also know 𝐺̅ (𝒓̅, 𝒓̅′), then the matrix representation of (22) becomes:
𝑇 𝑇
{{𝐸̅𝑚𝑒𝑎𝑠 (𝒓̅)} − {𝐸̅0 (𝒓̅)}} 𝑀×1 = [𝐸̅0 (𝒓̅′) ∙ 𝐺̅ (𝒓̅, 𝒓̅′)]𝑀×𝑁 {{𝜀(𝒓̅′)} − {𝜀0 (𝒓̅′)}} 𝑁×1
(2.23)
Now, we define [2] [𝐸̅0 (𝒓̅′ ) ∙ 𝐺̅ (𝒓̅, 𝒓̅′)] = [𝐽] as the Jacobian/sensitivity matrix and
𝑇 𝑇
{{𝐸̅𝑚𝑒𝑎𝑠 (𝒓̅)} − {𝐸̅0 (𝒓̅)}} 𝑀×1 = [𝐽]𝑀×𝑁 {{𝜀(𝒓̅′)} − {𝜀0 (𝒓̅′)}} 𝑁×1 (2.24)
14
Eqn. (2.24) can be solved for {𝜀(𝒓̅′)} employing the least square (LS) method as
below,
{𝜀} = {𝜀0 } − [𝐽𝑇 𝐽]−1 [𝐽]𝑇 ({𝐸̅𝑚𝑒𝑎𝑠 (𝒓̅)} − {𝐸̅0 (𝒓̅)}) (2.25)
{𝜀𝑘+1 } = {𝜀𝑘 } − [𝐽𝑇 𝐽]−1 [𝐽]𝑇 ({𝐸̅𝑚𝑒𝑎𝑠 (𝒓̅)} − {𝐸̅𝑘 (𝒓̅)}) (2.26)
So, eqn. (2.25) is very similar to what we got in eqn. (2.22). For the convenience of
comparison, eqn. (2.22) is stated here again (𝐹̅ ({𝜀𝑘 }) = 𝐸̅𝑘 (𝒓̅)),
{𝜀𝑘+1 } = {𝜀𝑘 } − 𝛼𝑘 [𝐽𝑇 𝐽]−1 [𝐽]𝑇 ({𝐸̅𝑚𝑒𝑎𝑠 (𝒓̅)} − {𝐸̅𝑘 (𝒓̅)}) (2.27)
where, 𝛼𝑘 is scalar constant. Thus, from (2.26) & (2.27), we can see that Gauss-
Newton method and inverse scattering method both reduce to the same iterative form:
15
2.4 “Born Approximation” and Imaging of High Contrast Object
The methods discussed so far are typically employed when the scattering object is in
low contrast with the background. However, in practical applications, cases may arise
when the scatterer is actually in high contrast. We discuss the aspect of addressing
In eqn. (2.23) above, we assumed the total field is equal to the incident field,
𝐸̅ (𝒓̅′) ≈ 𝐸̅0 (𝒓̅′). This approximation is called 1st order “Born approximation” [11]. As
there are two unknowns in eqn. (2.13): 𝜀(𝒓̅ ′) and 𝐸̅ (𝒓̅ ′ ), it cannot be solved easily for
Also, higher order Born approximation can be carried out in a recursive way:
𝐸̅ 𝑛𝐵 (𝒓̅) = 𝐸̅0 (𝒓̅) + 𝑗𝜔𝜇 ∭(𝜀(𝒓̅′)−𝜀0 (𝒓̅′))𝐸̅ (𝑛−1)𝐵 (𝒓̅′) ∙ 𝐺̅ (𝒓̅, 𝒓̅′)𝑑𝒓̅′ (2.28)
Generally, Born approximation (both 1st and higher order) is valid for weak
16
2.4.1 Imaging High Contrast Dielectrics
In this case, the initial permittivity distribution is iteratively updated until a stopping
criterion is satisfied. In each iteration, 1st order Born approximation is applied. The
𝐸̅ (𝒓̅) = 𝐸̅0 (𝒓̅) + 𝑗𝜔 ∭(𝜀 𝑘 (𝒓̅′)−𝜀0 (𝒓̅′))𝐸̅ 𝑘−1 (𝒓̅′) ∙ 𝐺̅ (𝒓̅, 𝒓̅′)𝑑𝒓̅′ (2.29)
To be noted is that the left hand side 𝐸̅ (𝒓̅) is known from measurements and hence
does not depend on the iteration number k. Also, the same Green’s function (which is
𝑘
𝐸̅ (𝒓̅) = 𝐸̅0 (𝒓̅) + 𝑗𝜔 ∭ ∆𝜀 𝑘 (𝒓̅′ )𝐸̅ 𝑘 (𝒓̅′) ∙ 𝐺̅ 𝑘 (𝒓̅, 𝒓̅′)𝑑𝒓̅′
𝜀 𝑘+1 = 𝜀 𝑘 + ∆𝜀 𝑘 (2.30)
17
The main difference between BIM and DBIM is:
So, DBIM is a potential method to image high contrast dielectrics. However, DBIM
requires the forward solution (update of Green’s function, 𝐺̅ and incident field, ̅̅̅
𝐸0 ) at
inverse scattering, the forward problem needs to be solved both to calculate the
numerically. The numerical model to compute the forward solution has to be as close
as possible to the physical imaging set-up used to carry out the actual measurements.
However, even after making all the efforts to eliminate the mismatch between these
two, inevitably, there are still some mismatches left. In the literature, this type of
mismatch/error is referred to as “model error” [4]. Not accounting for “model error” in
model error and measurement noise [29]-[30]. Typically, in the inverse problem of
microwave tomography, model error is the dominant error which has to be taken care
of.
ill-conditioned. So, a small amount numerical error or measurement error will destroy
the solution if direct inversion is employed. To overcome this issue, the idea of
regularization comes into play. Among several methods of regularization [27], one is
‘Tikhonov regularization’ which replaces [𝐽𝑇 𝐽]−1 with [𝐽𝑇 𝐽 + μ𝑅]−1 in (2.26) and
(2.27):
Here, the scalar μ is called the regularization parameter, and R is the regularization
solution.
Here, we will like to compare the image quality of MWT in comparison with some
other modalities such as, X-ray Computed Tomography (X-ray CT). The fundamental
domain [31]. Now, we will discuss the image resolution in brief for both modalities:
X-ray CT: As can be seen from Fig. 2.1, the wavelength used in X-ray CT for medical
X-ray CT images is mostly ~1mm, though 𝜇𝑚 resolution images are also reported by
using advanced imaging algorithms [31]. This huge difference between wavelength
20
Microwave Tomography: On the contrary, MWT images are limited in resolution
mostly due to the wavelength (~cm). Along with the ill-posed nature of the problem,
the diffraction and scattering in microwave frequencies from the imaging domain and
Tomography”. This means X-ray CT has very little amount of scattering and
21
discretization dimensions of the imaging domain in MWT are kept smaller than the
Table 2.1 compares some of the main features between X-ray CT and MW
tomography:
22
Chapter 3: A Fast Microwave Imaging Algorithm
are not suited for real-time applications. Specifically, most existing MWT algorithms
image the domain of interest (DOI) by solving the full non-linear inverse scattering
problem. This is done using several optimization algorithms, such as the Newton [33]
of time to generate the image. Keeping in mind the recent advances of high-speed
microwave devices and low-frequency electronics, the bottleneck for realizing real-
time MWI systems is the speed of the image reconstruction algorithms [35]. The
[36]. For example, one high-speed MWI algorithm intended for monitoring continuous
oil/gas flow [35] requires more than 25 sec to reconstruct the image. As such, existing
Impedance Tomography (EIT) has been used over the past decade [37]. Though it can
be made fast, EIT has several limitations including poor resolution (especially when
the image is recovered fast) and a requirement for the sensors to be in direct contact
with the DOI, a rather challenging requirement. With this in mind, we recently
23
proposed a fast microwave imaging algorithm for deep tissue imaging [16]. The
section with accuracy better than 6%. However, this imaging algorithm requires a
large number of ‘training samples’ (required to train the system the relationship
between the measured S-parameters and permittivity of the DOI) that are directly
proportional to the number of pixels in the DOI. Also, it loses accuracy if the required
a priori information is not available. As such, the algorithm is not suitable for higher
resolution images.
be noted is that due to the Born approximation [11] (see ch-2, section 2.4) which is
this method is particularly suited for reconstructions of low contrast objects with
respect to the background. No matter what method is being used, the forward
scattering solution takes the most significant portion of the total time needed for the
each iteration to minimize the total computational time. This is done by approximating
the forward solution by using the Taylor series expansion upto the first order term. To
24
Antenna
Substrate
Domain-of-
Interest (DOI)
Transmitting
Probe
“source”
around the DOI (see Fig. 1) to obtain the S-parameter measurements. We remark that
MWI. The entire imaging domain along with the finite dipole antennas is modeled
using a full wave finite element method (FEM) based software Ansys HFSS to obtain
25
vector FEM based on edge elements to achieve highly accurate electromagnetic
solutions [38]-[40]. This almost exact modeling approach enables us to employ the S-
corresponding electric fields for the imaging algorithm to work. This is because most
works on MWI to date do not use full wave modeling of the actual system with
antennas [41]-[48]. Finally, reconstructed images for both simulated and measured
data at 1.2 GHz for a DOI diameter of 19cm are obtained using the proposed
Fig. 3.1 shows the geometrical configuration of the proposed sensor. As seen, the
sensor has a set (P in number) of transceiver antennas placed conformally around the
DOI. For the scenario shown in Fig. 1, one antenna is transmitting and the others are
receiving. This is repeated for all the antennas/probes to construct the S-parameter
matrix:
𝑆11 ⋯ 𝑆1𝑃
[𝑆] = [ ⋮ ⋱ ⋮ ] (3.1)
𝑆𝑃1 ⋯ 𝑆𝑃𝑃
26
Here, the diagonal entries 𝑆𝑖𝑖 refer to the reflection coefficients of the P antennas. As
is the case with any reciprocal medium, 𝑆𝑖𝑗 = 𝑆𝑗𝑖 , implying that the upper- and lower-
triangular sectors of [S] have the same values. That is, of the P2 entries in [S], only
𝑃×(𝑃−1)
𝑀= elements are independent. These M independent measurements will be
2
We also define the relative permittivity of the pixels comprising the DOI (see Fig. 1),
In (3.3), 𝜀𝑛 represents the relative permittivity of the n-th pixel of the DOI and N =
total number of pixels. Our goal is to find {𝜺} using the collected measurement vector
{𝑺}𝑚𝑒𝑎𝑠 defined in (3.2). To find the desired {𝜺}, we proceed to minimize the function
1
𝐺({𝜺}) = 2 argmin‖{𝑺}𝑚𝑒𝑎𝑠 − 𝐹({𝜺})‖2 (3.4)
{𝜺}
where F({𝜺}) is the corresponding {𝑆} for a given set of {𝜺}, and ||. || is the L2 norm of
a vector. To solve for {𝜺}, we will employ a modified version of the Gauss-Newton
iterative method. In its original form, the Gauss-Newton iterations have the form,
−1 𝑇
{𝜺𝒌+𝟏 } = {𝜺𝒌 } − [ J kT J k ] [ J k ] ({𝑺}𝑚𝑒𝑎𝑠 − 𝐹({𝜺𝒌 })) (3.5)
27
where [ J k ] is the M×N Jacobian matrix. We note that the elements of the Jacobian
matrix are given by J 𝑖𝑗 = 𝜕𝐹𝑖 ⁄𝜕𝜀𝑗 , that is, [ J k ] is calculated by perturbing each
pixel’s permittivity an amount ∆𝜀𝑗 and measuring the change in 𝐹𝑖 , referred to as ∆𝐹𝑖 .
To avoid the majority of the computational burden in solving (3.5), we employ the
1) The Jacobian matrix is calculated only once, at some initial ‘point’ {𝛆} = {𝛆𝟎 }. This
2) To avoid solving the direct problem, viz. that of calculating F({𝛆𝐤 }) for every
was used. Other regularization techniques may also be applied including the Truncated
28
Singular Value Decomposition (TSVD) method [27]. Following Tikhonov’s
19cm
10cm
Dipole
transceivers
(a)
(b)
Figure 3.2: (a) Simulation set-up, and (b) experimental set-up used to validate the
proposed imaging concept.
29
In (3.8), the scalar μ is called the regularization parameter, and [R] is the
regularization matrix. In the literature, there are several possible choices for [R],
including the identity matrix, positive diagonal matrix, approximations of the first and
second order differential operators, etc. [26]-[27]. In this work, we will set [R] as the
identity matrix. Importantly, we do note that the regularization term μR also serves to
make the matrix positive definite, a necessary condition for convergence [27].
Notably, for a typical imaging problem, the number of unknowns is higher than the
solution of (3.8) is the regularized minimum norm solution of the system [27].
placed conformally on the periphery of the DOI, depicted in Fig. 3.2. The antennas
and antenna size. As the antennas are very close to the DOI at this frequency, almost
all of the DOI can be considered to be in the near-field. We note that in previous
works, antennas with higher directivity were used [49]. However, for near-field
applications, higher directivity antennas are not required [44]-[46]. Instead, we prefer
30
to employ simple dipole antennas and focus on reducing the so called ‘model error’ by
modeling them exactly through full- wave FEM method. Using simple antenna thus
reduces associated computational costs. The employed dipoles are 6cm-long (~λ/4 in
free-space), and are matched to 50Ω. They are fabricated using copper tape on a
these copper strips can also be replaced with conductive textile antennas for added
flexibility [50]. As shown in Fig. 3.2, the DOI considered in this study is a finite
the cylinder’s dielectric cross-section. For the direct problem solution, as mentioned
Figure 3.3: Power distribution inside the imaging domain. The red region denotes the
power from the excited dipole.
31
(a) (b)
(c) (d)
Figure 3.4: Reconstructed images with synthetic data: (a) noiseless, (b) SNR = 50dB,
(c) SNR = 40 dB, (d) SNR = 30 dB. The dashed circular lines indicate the actual
position of the anomaly.
earlier, Ansys HFSS was utilized where more than 80,000 tetrahedra were used to
rectangular grid. Therefore, the inverse problem amounts to solving the dielectric
constants corresponding to 133 variables. We note that the solution time of (3.8) was
32
about 0.25 sec on a desktop computer with processor of Intel core i7 @ 3.40 GHz and
algorithm [3], where the forward problem was solved once in each iteration, required
An example field power distribution within the cross section of the cylinder in Fig. 3.2
is shown in Fig. 3.3. This corresponds to a single dipole excitation while the others
were receiving. Of importance is that the dipole field penetrates the entire cylinder
Figure 3.5: Three different imaging scenarios for the measurement setup. PTFE rods
located: a) one near boundary, b) one in the center and, c) two near the opposite
boundaries.
33
cross- section. A necessary condition for the accuracy of the inverse solution is to
necessary to design a sensor that ensures enough field penetration into the DOI. Also,
the imaging algorithm has to account for inter-antenna coupling. To do so, the sensor
antenna and the whole system must be designed and solved using full wave
parameter data. The employed simulation set-up is shown in Fig. 3.2(a). As already
Therefore, even small errors can lead to erroneous images [27]. Additionally,
numerical round-off errors may result in fictitious solution singularities [26]. So, to
test the robustness of the approach, we proceeded to add random noise to the synthetic
S-parameter data. The scalar regularization parameter, 𝜇, in (3.8) played a critical role
in reconstructing the image. Specifically, the higher the value of 𝜇, the more robust the
algorithm is against noise. At the same time, the solution becomes overly smooth, i.e.,
34
As in practice, the measured data will inevitably have some errors/noise. Hence, it
is worth investigating its effect on the reconstructed image. To do so, we add different
2
Here, 𝑃𝑠𝑖𝑔𝑛𝑎𝑙 and 𝑃𝑛𝑜𝑖𝑠𝑒 are the signal and noise power, respectively, 𝐸𝑖,𝑠𝑖𝑔𝑛𝑎𝑙/𝑛𝑜𝑖𝑠𝑒 is
the square of the magnitude of the i-th element of signal/noise E-field vector and N is
the number of elements in the signal/noise vector. Example reconstructed images for
different SNR values are given in Fig. 3.4. For these cases, the 19cm-diameter DOI
15. The dashed circular lines in Fig. 3.4 indicate the actual position of the anomaly. As
seen, very good image reconstruction is achieved. As would be expected, the higher
The measurement set-up used for validation is depicted in Fig. 3.2(b), and the
specific imaging scenarios under consideration are given in Fig. 3.5. The 19cm-
diameter DOI consisted of air background where one or two dielectric cylindrical rods
35
were inserted, each of diameter 3.15 cm. The rods were constructed of PTFE (Poly
For the experiment, each of the 8 transceiver antennas was sequentially excited, and
S-parameters were collected using a Vector Network Analyzer (VNA). The calibration
used to extract S-parameters with the DOI being of uniform background, 𝑆𝑠𝑖𝑚,𝑏𝑘𝑔𝑛𝑑 .
Subsequently, S-parameter measurements were carried out and two sets of data were
collected: 1) when the DOI is filled with the background permittivity, 𝑆𝑚𝑒𝑎𝑠,𝑏𝑘𝑔𝑛𝑑 ,
and 2) for the actual DOI permittivity distribution with the anomaly to be imaged
Section II, all these data sets are vectors. On the assumption that changes in S-
parameters in presence and absence of the anomaly, remains the same between
𝑆
𝑆𝑐𝑎𝑙 = (𝑆𝑚𝑒𝑎𝑠,𝑎𝑐𝑡𝑢𝑎𝑙 ) × 𝑆𝑠𝑖𝑚,𝑏𝑘𝑔𝑛𝑑 (3.10)
𝑚𝑒𝑎𝑠,𝑏𝑘𝑔𝑛𝑑
In logarithmic form,
36
(a) (d)
(b) (e)
(c) (f)
Figure 3.6: Reconstructed images using: (a)-(c) simulated noiseless data, (d)-(f)
measured data.
This calibrated data, 𝑆𝑐𝑎𝑙 |dB, are subsequently used in the inversion algorithm (viz.
37
Fig. 3.6 shows the reconstructed images for all three scenarios under consideration
(see Fig. 3.5). Images extracted using the calibrated measured data (right column) are
compared with those from simulated noiseless data (left column). Clearly, the images
based on experimental data are in very good agreement with those based on
simulations. Some small imaging artifacts are only observed in the experimental
results. We noticed that there is a tendency for the artifacts to appear near the
boundary of the DOI. This is because higher field density regions were present near
38
Chapter 4: A Novel Method to Mitigate Real-
Imaginary Image Imbalance
MWI has the capability of reconstructing both relative permittivity and conductivity of
an imaging domain, which, in principle, can convey distinct information. For example,
in industrial flow imaging, the relative permittivity distribution is related to the flow
pattern and conductivity distribution may convey information about the impurity in the
domain [51]. In biomedical applications, the relative permittivity is more related to the
bound water content [43]. Bound water may be a potential indicator of pathological
conditions like cancer, edema etc [43]. On the other hand, conductivity distribution of
a biological domain has formed the basis of noninvasive temperature sensing for
Evident from our experience [53] and an extensive literature search is that, the
[54]. There are very few attempts made to address this issue. An attempt to mitigate
this issue was reported in [55] using a pre-scaling approach on the real and imaginary
parts of the permittivity. However, the robustness of this approach is not clear. To this
needed to develop a microwave imaging algorithm that has the ability to produce
imaging has been done previously by different imaging modalities [56]-[57] to obtain
map of material concentrations in the domain. Unlike those works, we do not target to
obtain the material concentration of the imaging domain. Rather, in this work, the
real and imaginary permittivity images. For this, an algorithm based on Gauss-Newton
method that minimizes a cost function with respect to the fraction parameters, has
been formulated. Both the relative permittivity and conductivity images are calculated
from the fraction parameters once they are reconstructed. Of importance is that, the
Also, as the real and imaginary parts are coupled in the inverse algorithm, this
approach renders better image recovery for both. For the forward problem solution,
unlike other two and three dimensional microwave tomographic approaches [42]-[48],
in this work we use a complete full-wave analysis considering all the transceiver
antennas incorporated in the model. This would help in reducing the so-called model-
error. Reconstructed images from experimental data are compared with those from a
efficacy of the method. Overall, the proposed method in this letter has clearly shown a
40
Domain of
Imaging (DOI)
Cylindrical
imaging
chamber
Dipole
transceivers
(a)
Antenna
Substrate
Domain-of-
Interest (DOI)
Transmitting
Probe
“source”
(b)
(c)
microwave tomography.
4.1 Method
Here, the domain to be imaged is brought inside a cylindrical chamber (shown in Fig.
4.1a,c) with a set of 8 dipole antennas placed on the surface of the cylinder. One
antenna is excited at a time and the others are used as receivers. This process is
repeated until all the antennas are excited one-by-one. Scattering parameters are
collected from all the receiver antennas and stacked in one vector, {𝑺}𝑚𝑒𝑎𝑠 to form the
The forward problem is solved using the finite element method (FEM) based
software package Ansys HFSS. More than 80,000 tetrahedra were used to mesh the
whole 3D geometry including the antennas. To be noted is that, unlike most other 3D
microwave imaging approach, we use a full wave forward solver taking into account
all the finite antennas as part of the forward model. Any possible surface wave effect
and inter-antenna couplings are thus taken care of during the forward solution. This in
turns, yields more accurate solution of the forward problem which helps reducing any
is shown in Fig. 1b. To be noted is, this kind of dual mesh approach (one for the
forward problem and another for the inverse problem) has already been used in [3].
We define the complex relative permittivity of the domain of interest (DOI) shown in
Fig. 4.1,
Here, Npix refers to the total number of pixels subdividing the DOI. Next, we choose
the frequency 𝜔. The relative permittivity at the n-th pixel can be written as,
Here, 𝑟1,𝑛 , 𝑟2,𝑛,….. , 𝑟𝑁𝑑𝑖𝑒𝑙𝑒𝑐,𝑛,𝑛 are the fractions of 𝑒1 , 𝑒2 , … , 𝑒𝑁𝑑𝑖𝑒𝑙𝑒𝑐 respectively for any
the permittivity of each pixel, 𝜀𝑛 , one can employ more sophisticated expressions [58]
43
4.1.2 Reconstruction of Fraction Parameters
In conventional microwave imaging methods, the goal is usually to find {𝜺} using the
measured data, {𝑺}𝑚𝑒𝑎𝑠 . In such scenario, to find the desired {𝛆}, one proceeds to
where F({𝜺}) is the forward problem solution for a given set of {𝜺}, and || ∙ || is the L2
norm of a vector. The second term in (4.4) is the regularization functional of which [R]
controlling the amount of regularization imposed (user specified) and {𝜺𝟎 } is any
Now, instead of solving for the desired {𝜺}, in this work, we propose to solve the
imaging problem for the fraction parameters. Thus, for our problem, the optimization
1 ̃𝟎 } )‖2 )
𝐶({𝒓̃}) = argmin 2 (‖{𝑆}𝑚𝑒𝑎𝑠 − 𝐹({𝒓̃})‖2 + 𝜇 ‖𝑅({𝒓̃} − {𝒓 (4.5)
{𝒓}
where,
44
In the above, {𝒓̃𝒌 } represents the stacked vector of all the fraction parameters
{𝒓𝒌 }1 , {𝒓𝒌 }2 … . , {𝒓𝒌 }𝑁 for all the pixels in the DOI where the superscript ‘k’
𝑑𝑖𝑒𝑙𝑒𝑐
Now, the optimization problem in (4.5) can be solved via the conventional Gauss-
Newton method. As reported in the authors’ previous work [53], a fast implementation
̃𝒌 }) = 𝐹({𝒓
𝐹({𝒓 ̃𝒌 } − {𝒓
̃𝟎 }) + [𝐽𝑟̃ ]({𝒓 ̃𝟎 }) (4.7)
Here, [𝐽𝑟̃ ] is the Jacobian matrix. To be noted is that, unlike in typical microwave
reconstructed [42]-[48], here we calculate the Jacobian matrix with respect to the
45
Here, ‘m’ refers to the m-th measurement point among the total M number of
measurements, ‘n’ refers to the n-th pixel and ‘t’ refers to the t-th known permittivity
(see Eqn. (4.2)). In (4.9) 𝑒𝑡 is a constant and 𝐽𝜀 is calculated using adjoint method [3].
.
.
.
{𝒓𝑵𝒅𝒊𝒆𝒍𝒆𝒄 } = 1 − ({𝒓𝟏 } + {𝒓𝒌+𝟏
𝒌+𝟏 𝒌+𝟏 𝒌+𝟏
𝟐 } … + {𝒓𝑵𝒅𝒊𝒆𝒍𝒆𝒄 −𝟏 })
(4.10)
where [𝐺𝑡𝑘 ] and [𝐵𝑡𝑘 ] (t=1, 2, …,𝑁𝑑𝑖𝑒𝑙𝑒𝑐 ) are the gradient and Hessian matrices from
(4.5) with respect to 𝑟𝑡 at the k-th iteration. Also, the last equation of (4.10) directly
𝑇
[𝐺𝑡𝑘 ] = [𝐽𝑟𝑡 ] ({𝑺}𝑚𝑒𝑎𝑠 − 𝐹({𝒓𝒌𝒕 })) + 𝜇[𝑅 𝑇 𝑅]({𝒓𝒌𝒕 } − {𝒓𝟎𝒕 }) (4.11)
Upon obtaining the fraction parameters {𝒓𝒕 } (t=1, 2, …,𝑁𝑑𝑖𝑒𝑙𝑒𝑐 ) from (4.10), we can
46
(a) (c)
(b) (d)
Figure 4.2: Reconstructed images for one anomaly: (a) rel. permittivity (direct), (b)
conductivity (direct), (c) rel. permittivity (indirect), (d) conductivity (indirect). The
dashed circular lines indicate the actual position of the anomaly.
As shown in Fig. 4.1, the DOI considered, is a glass cylinder of diameter=20cm and
47
(a) (c)
(b) (d)
Figure 4.3: Reconstructed images for three anomalies: (a) rel. permittivity (direct), (b)
conductivity (direct), (c) rel. permittivity (indirect), (d) conductivity (indirect). The
dashed circular lines indicate the actual position of the anomaly.
the cylinder. All dipoles are identical, 6 cm long and matched to 50Ω. For imaging
non-biological DOI with low dielectrics, the operating frequency was chosen as 1.75
GHz that yields a suitable size of the antennas. The permittivity interior to this
measurements were carried out with a network analyzer, Agilent N5230A PNA-L
48
series. The dielectric permittivity of the materials were measured using Agilent
(a) (c)
(b) (d)
Figure 4.4: Reconstructed images for two anomalies: (a) rel. permittivity (direct), (b)
conductivity (direct), (c) rel. permittivity (indirect), (d) conductivity (indirect). The
dashed circular lines indicate the actual position of the anomaly.
49
In Figs. 4.2 and 4.3, we show image reconstruction for one and three vegetable oil
anomalies. The anomalies are prepared by adding salt with vegetable oil to achieve,
𝜀𝑟 = 2.5 − 𝑗1.41 (conductivity=0.1 S/m) and placed inside the DOI in the background
of air (𝜀𝑟 = 1). In the left column of both the Figs. 4.2 and 4.3, real and imaginary
(a) (c)
(b) (d)
Figure 4.5: Reconstructed images for three anomalies: (a) rel. permittivity (direct), (b)
conductivity (direct), (c) rel. permittivity (indirect), (d) conductivity (indirect). The
dashed circular lines indicate the actual position of the anomaly.
50
minimize a cost function w.r.t. the complex permittivity are shown (direct). While in
the right column of the Figs. 4.2 and 4.3, real and permittivity images through
from Figs. 4.2 and 4.3 that, by employing the proposed fraction imaging method: a)
the difference in the image quality between real and imaginary permittivity images is
largely mitigated, b) the quality of both real and imaginary part reconstruction is
improved. The latter is likely due to the fact that the real and imaginary permittivity
reconstructions are coupled together in the inverse algorithm described in section II.
In this case, the same imaging chamber has been utilized with the same set of dipole
transceivers. Only the DOI inside the chamber is replaced with biological phantoms:
background phantom and anomaly phantom. The background phantom was prepared
following the procedure in [59] through adjusting the ingredients to achieve 𝜀𝑟𝑏𝑎𝑐𝑘 =
39.88 − 𝑗5.8 (conductivity=0.39 S/m) at the operating frequency 1.25 GHz. The
𝑎𝑛𝑜𝑚𝑎𝑙𝑦
anomaly phantom was prepared by mixing water and glycerin to achieve 𝜀𝑟 =
60.2 − 𝑗8.77 (conductivity=0.61 S/m) at the same frequency. Real and imaginary
permittivity images using the proposed fraction parameter reconstruction are shown in
Figs. 4.4 and 4.5 along with those using the conventional Gauss-Newton method.
51
Again, using the proposed method, the real-imaginary image imbalance is largely
mitigated and the overall quality of both the images is improved for the biological case
as well.
Microwave imaging methods have been employed in the past to image dielectrics in
electrical capacitance tomography (ECT) [58], [60]-[68] is the most widely used
expected to have several advantages over ECT but it is yet to be extensively explored
calculations have been shown, this technique has several limitations including: 1)
measurements at more than one frequencies are required, 2) at least one of the phases
multiphase flow overcoming the limitations of ECT as mentioned above. The main
52
benefits of this method are: 1) measurements at only one frequency are sufficient for
tomography (ECT) [58]. Though ECT is the most widely used modality to image two-
phase flow scenarios, it struggles to obtain meaningful image when more than two-
phases are encountered in the imaging domain [58]. Also, ECT images collapse as the
conductivity of the medium starts rising [68] and when the permittivity of the bulk
medium is very high, i.e. water. Another severe limitation of ECT measurements is, it
has much lower sensitivity on the central region of the imaging domain than on the
periphery of the domain. This later issue results in an overemphasized image for an
anomaly which is present near the periphery than that of one near the center [68]. To
technique has been recently proposed [58]. Here, by doing the measurements for two
frequencies, an additional dimension in the data has been added to resolve the
53
domain of
Imaging (DOI)
cylindrical
imaging
chamber
dipole
transceivers
(a)
antenna
substrate
domain-of-
interest (DOI)
transmitting
probe
“source”
(b)
54
air
water
oil
a
Figure 4.7: a) Cross-sectional view of the flow models: case-I (left) and case-II
(right), b)reconstructed image of the entire flow model c) reconstructed image of the
phase air and d) reconstructed image of the phase water. The black circles correspond
to the actual position of the phases.
55
obtain the fractions of all individual components of an N phase flow scenario if
measurements in N-1 frequencies are available. Also, in order for the method to work,
one of the phases of the medium must have some conductivity [58]. But, on the other
hand, ECT has been shown to have failed when the bulk medium is conductive [68].
In this work, we show that, with our fraction imaging method employing
only single frequency measurements, for a multiphase flow scenario. We show our
results for a three-phase flow model, however, our method should theoretically work
for any number of phases. Exploiting the capabilities of microwave tomography, it has
been already shown in the previous section that this method works even when the
medium is highly lossy. Here, we show that unlike the ECT- based method [58], our
method does not have the limitation of one phase to be slightly lossy, thus it works for
The imaging set-up is shown in Fig. 4.6 where a set of 8 dipole antennas are
show our proof-of-concept fraction imaging on a 2D grid similar to the one shown in
Figure 4.6(b). We remark that this method can naturally be extended to 3D without
difficulty.
56
Inside the cylindrical pipe, we have the multiphase flow occurring. We consider two
flow-models which are shown in Fig. 4.7(a), where air and water are the so-called
dispersed flow and oil is the continuous flow. We assumed here that the dispersed
phases have a columnar flow, however, the method is expected to work for other flow
Newton method [53] is shown in Fig. 4.7(b) where both air and water can be observed
in the oil background. The images of the individual phases are shown in Fig. 4.7(c)
and 4.7(d). We can observe that the individual phases: air and water are clearly
separated by this proposed technique, most importantly, without the need for
totally lossless (unlike [58] where any one of the phases has to be slightly lossy) to
4.4 Summary
In this work, a new method is proposed to improve the accuracy in reconstructing the
imaginary part of the permittivity in microwave tomography. In doing so, the image
quality of the real permittivity is also shown to be improved. As the proposed method
is not problem specific, the algorithm can be adapted to imaging applications like
challenges: 1) there is hardly any option to calibrate the measurements, 2) the shape of
the body to be imaged in not exactly known. On top of the regular challenges with
MWT, these additional ones make on-body MWT a formidable task. In this chapter,
we propose an Artificial Neural Network (ANN) based approach to address these two
issues.
Artificial neural networks (ANNs) are computing systems inspired by the biological
neural networks that constitute animal brains [69]. Such systems learn (progressively
unidirectional signal with an activating strength that varies with the strength of the
58
neurons) are strong enough, the receiving (postsynaptic) neuron activates and
Typically, neurons are organized in layers. Signals travel from the first (input), to
the last (output), possibly after traversing the layers multiple times. In addition to
receiving and sending signals, units may have state, generally represented by real
numbers, typically between 0 and 1. A threshold or limiting function may govern each
connection and neuron, such that the signal must exceed the limit before propagating.
Also, at each node, an activation function can be defined which dictates the output of
that node given an input or set of inputs. Typical activation functions are binary,
logistic function, sigmoid function etc. [69]. An example ANN is shown in Fig. 5.1.
As mentioned in [69], an ANN can potentially predict any complex function with
sufficient number of hidden layers. We will exploit this capability in this work.
As in the case of supervised learning [69], an ANN has to be trained first to be used
later to predict any unknowns. In the ‘training’ phase, the ANN has to be presented
with2 sets of data: inputs and targets as shown in Fig. 5.2(a). Employing any suitable
optimization algorithm, the ANN will adjust its weights based on the data set. Typical
training algorithms include backpropagation training [69]. Here, in the case of MWT,
59
1
I1 1 2 1 O1
I2 v2 v3 v2 O2
. . .
. . .
. . .
IN N
v v
N1-1 v
M OM
Inputs Targets
v1
N
Hidden layer
the inputs would be the S-parameter measurements collected around the body-to-be-
imaged and the corresponding targets are the pixel-by-pixel permittivity image. These
2 sets of data has to be presented for many different scenarios that may encounter in
the on-body MWT set-up to make sure the ANN is trained well. After training the
ANN, it is presented with an input to determine the unknowns as shown in Fig. 5.2(b).
One thing to mention about ANN is, while training the network, one should be careful
against ‘overfitting’ the network. ‘Overfitting’ refers to the case when an ANN learns
60
the training data very well but becomes unable to predict results with unseen data [69].
This frequently happens if the ANN employed is too complex (has too many
parameters) than
Inputs
(E-field
measurements) ANN
Training
algorithm weights
Targets (w’s)
(permittivity
image)
(a)
Network with
Inputs weights learned
Targets
(E-field (permittivity
from Training image)
measurements)
(b)
61
the actual unknown function to be predicted. In this work, ‘overfitting’ is avoided by
choosing an ANN which is just complex enough (with 3 hidden layers, 35 neurons in
each layer) to yield similar errors both with training data and testing data.
utilizing the data itself. A three-step approach for this purpose is proposed below:
i) Mean subtraction:
The mean of the data for each frequency is subtracted from the data itself:
(5.1)
This subtraction helps to get rid of any additive error/noise in the data.
62
This is done to project the data onto a different subspace [70]. It turns out the data is
better distinguished in this subspace across different images, hence, enabling better
image reconstruction.
Let,
[𝑇]𝑖𝑛𝑖𝑡
𝐾×𝑀𝑡𝑜𝑡𝑎𝑙 = the data matrix with ‘K’ number of sample scenarios to train the ANN.
[𝑇]𝑖𝑛𝑖𝑡 𝑇
𝐾×𝑀𝑡𝑜𝑡𝑎𝑙 = [𝑈]𝐾×𝑅 [𝛴]𝑅×𝑅 [𝑉]𝑅×𝑀𝑡𝑜𝑡𝑎𝑙
(5.2)
𝑘
{𝑆𝑑𝑟 }1×𝑅 = {𝑆}𝑡𝑜𝑡
𝑚𝑒𝑎𝑠 1×𝑀 [𝑉]𝑀𝑡𝑜𝑡𝑎𝑙 ×𝑅
𝑡𝑜𝑡𝑎𝑙
(5.3)
63
on this subspace will reduce the data to an R dimensional data by retaining only the
meaningful dimensions.
iii) Normalization:
This is done to bring the data in the same numerical range for different scenarios that
𝑘
may be encountered. The normalization of {𝑆𝑑𝑟 } is done with respect to the maxima of
𝑘
{𝑆𝑑𝑟 }1×𝑅 . So, the numerical value of each element in the data lies between 0 and 1.
Dimensions
Typically, ANNs work well when the number of unknowns are not very high [69]. It is
observed that, employing ANN directly on the 2D image it becomes difficult to obtain
Transform (DFT) on the image and retain some significant DFT coefficients to train
the ANN. To be noted is that, the idea of employing DFT for imaging purposes has
been used before in [71]. Once the ANN is trained, it is employed in a case where the
image is unknown. Again, the significant DFT coefficients of the unknown image is
obtained from the ANN. Finally, inverse DFT is applied to recover the actual
64
(a) (b)
Anomaly
Antennas
65
unknown image. In this work, we pixelized the original image in 100×100=10000
pixels. In the DFT domain, instead of retaining 10000 coefficients, we choose to retain
only upto M=N=7 (=49 complex coefficients) and discard the rest. This particular
number is chosen just to make sure the image has similar quality as compared to an
method). We remark that, retaining more DFT coefficients adds little value to the
Fig. 5.3 shows an image represented by 49 (upto M=N=7) DFT coefficients. It can be
seen that though the shape of the anomaly is not constructed well, it gives the correct
position of the anomaly without major image artifacts. This is similar to typical
microwave tomography where shape of the anomaly is usually not reconstructed well.
DFT:
𝑁−1 𝑀−1
1 2𝜋𝑥𝑘 2𝜋𝑦𝑙
∑ ∑ 𝐼(𝑥, 𝑦)𝑒 −𝑖( + )
𝑎(𝑘, 𝑙) = 𝑅𝑒 [ 𝑀 𝑁 ]
√𝑀𝑁 𝑥=0 𝑦=0
𝑁−1 𝑀−1
1 2𝜋𝑥𝑘 2𝜋𝑦𝑙
∑ ∑ 𝐼(𝑥, 𝑦)𝑒 −𝑖( + )
𝑏(𝑘, 𝑙) = 𝐼𝑚 [ 𝑀 𝑁 ]
√𝑀𝑁 𝑥=0 𝑦=0
(5.4)
66
Inverse DFT:
2𝜋𝑥𝑘 2𝜋𝑦𝑙
1 𝑖( + )
I(𝑥, 𝑦) = ∑𝑁−1 𝑀−1
𝑘=0 ∑𝑙=0 (𝑎(𝑘, 𝑙) − 𝑖𝑏(𝑘, 𝑙)) 𝑒
𝑀 𝑁 (5.5)
√𝑀𝑁
shown in Fig. 5.4. A set of 12 dipole antennas around a cylindrical imaging domain
can be seen. The dipoles are 5.5 cm in length and resonate from 0.6 GHz to 1 GHz.
This relatively wideband performance is possible because the permittivity of the outer
layer (mimicking skin) is 46.5±10% and loss tangent 0.39±10% [72]-[73] which is
directly touching the antennas. Also, the lung permittivity and loss tangent are set to
accumulation in the lung, so we set its permittivity to 80 (-10%) and loss tangent
0.1±10%.
In Fig. 5.5, some imaging results have been shown for four different scenarios. In
this case, the ANN training is carried out for 400 different known cases. These 400
cases are generated by varying the position, radius and permittivity of the anomaly, the
permittivity of background, permittivity of outer layers (skin) and major and minor
axes of the elliptical shape of the body. Matlab Neural Network Toolbox [74] has been
67
employed to train the ANN where backpropagation algorithm was used. The ANN has
3 hidden layers each having 35 nodes (artificial neurons) in it. Sigmoid function
1
[f(x) = 1+𝑒 −𝑥 ] is used as activation function [69]. It can be seen from Fig. 5.5 that, the
position of the anomaly is correctly recovered in the reconstructed image in all the
four cases.
The images of Fig. 5.5 are for the scenario when we have the exact information of
the shape of the body. Now, we consider a case when the shape and hence, the
position of the sensors are not exactly known. Fig. 5.6 shows some images when we
allow the shape of the body to become ellipsoid instead of circular and also the two
axes of the ellipsoid are kept as variables. Specifically, we allow the two axes to vary
±5% around a nominal value equal to 15cm. To be noted is that, the nominal shape is
still circular and images are plotted on this circular shape as shown in Fig. 5.6. From
this, it can be seen that, the images successfully recovers the positions of the
anomalies, but with some artifacts. This is expected as the ANN now is trained with a
set of data which has been generated considering two additional variables.
Here, we show a realistic case in Fig 5.7 where the imaging is carried out after
placing 8 dipole antennas around a numerical human body model obtained from [75].
The permittivity and loss tangent of the domain are varied around nominal values
68
(a) (b)
(c) (d)
Figure 5.5: Imaging of tumors for four different cases in the lung background. The
dotted circle shows the actual position and shape of the tumors. Dimensions are in
meters.
59±10% and 0.39±10% for heart and 35±10% and 0.4±10% for lung (the bulk
medium), respectively. The anomaly permittivity and loss tangent are set to 80 (-10%)
69
(a) (b)
(c) (d)
Figure 5.6: Imaging of tumors for four different cases in the lung background.
The dotted circle shows the actual position and shape of the tumors. Dimensions
are in meters.
70
antennas anomaly
heart
(a)
tumor
heart
Shape is
uncertain
(b)
71
Figure 5.8: Image reconstruction on an elliptical grid for a human torso with non-
exact shape information.
and 0.1±10%, respectively. This human body model is a simplistic one which ignores
the presence of skeleton and other biological details. We remark here that, these
variations and others like thickness of outer layer etc. are already mitigated through
the 3 step ‘self-calibration’ process employed on the data. That is why it is anticipated
that this simplistic model of human body should mimic a realistic scenario reasonably
well. The image is reconstructed on an elliptical grid as shown in Fig. 5.8, which
72
closely resembles the actual body shape (which is not perfectly elliptical). The ANN
has been trained by generating data using the human body model. The shape variation
variation is assumed along the z axis. Also, the DFT coefficients for the training are
calculated from the elliptical grid, not the human body model. Again, 400 different
cases are generated to train the ANN. Once the ANN is trained, it is used to generate
the image on an elliptical grid as shown in Fig. 5.8, by employing S-parameter data as
input to the ANN obtained from a human body model simulation. As can be seen in
73
Chapter 6: Novel 3D Imaging Method Based on
Reciprocity
A major limitation of the above mentioned MWT algorithms is that they involve
model is necessary to reduce the so-called ‘model error’ which implies high
computational cost. Above all, the time required for successful image reconstruction is
monitoring is required. There are few attempts in the literature to overcome these
issues [14], [53], [76]. However, they suffer from several issues themselves. The
algorithm in [53] proposed by the author shows high-speed imaging, but still requires
matrix free imaging, however, complicated cases, i.e. imaging domain with multiple
big scatterers are not presented. The method in [14] can only image the strongest
scatterer in the imaging domain. Above all, none of the methods work well for highly
noisy/erroneous data.
advantages over the conventional techniques in the literature. One of the major merits
is that it does not require solving the inverse problem, i.e. no matrix inversion is
from a perfect electric conductor (PEC) cylinder and reciprocity between scattering
source and measured E-fields. At first, the electric fields around an imaging domain
are measured (either synthetic or actual measurements) after exciting one antenna
employing the idea of reciprocity, the measured electric fields are considered as
secondary sources which radiates back to the location(s) of significant scatterers in the
imaging domain, where they emanated from initially. All these secondary radiations
are then summed-up for each pixel in the imaging domain to obtain an image of the
(SNR) goes as low as −5 dB. To our knowledge, microwave imaging with such a low
SNR has not been proposed before. We will show that this feature can potentially be
used in settings with portable and conformal antennas, where noise and sensor position
Time reversal (TR) is a well-known technique that has been widely used for time
75
̅ Measurement
y Point
̅′
Line Source
a ′
x
PEC Cylinder
(a)
PEC
Cylinder
Measurement
Point Line Source
(b)
Figure 6.1. a) Electric line source near a circular PEC cylinder (top view), b) Line
source radiation and measurement points on the surface of the cylindrical imaging
domain.
76
knowledge, this is the first demonstration of such technique that exploits the idea of
For the imaging problem, we exploit the formulation of line source scattering from a
circular PEC cylinder [9]. Let a circular PEC cylinder of radius ‘a’ and infinite length
be located near a line source (invariant along the z-axis), as shown in Fig. 6.1(a). We
assume a co-ordinate system such that the PEC cylinder’s center coincides with the
center of the co- ordinates, the line source is located at (𝜌′ , ′) and any measurement
point is located at (𝜌, ) where 𝜌 > a. Now, let the electric line source is of the
constant current ‘I’ and we can express the incident E-field radiated by the line source
(2) ′ 𝑗𝑛(𝜑−𝜑 ) ′
𝛽𝑠 2 𝐼 ∑𝑛=+∝
𝑛=−∝ 𝐽𝑛 (𝛽𝑠 𝜌)𝐻𝑛 (𝛽𝑠 𝜌 )𝑒 , 𝜌 ≤ 𝜌′
𝐸𝑧𝑖 = − 4𝑤 𝜀 { (2) ′ (6.2)
∑𝑛=+∝ ′ 𝑗𝑛(𝜑−𝜑 )
𝑠
𝑛=−∝ 𝐽𝑛 (𝛽𝑠 𝜌 )𝐻𝑛 (𝛽𝑠 𝜌)𝑒 , 𝜌 ≥ 𝜌′
77
In (6.1) and (6.2), 𝑤𝑠 is the operating frequency, 𝛽𝑠 is the propagation constant at the
(2)
frequency, 𝑤𝑠 , 𝐻𝑛 is the Hankel function of the second kind and order n, 𝜀 is the
Now, as shown in Fig. 6.1(b), a line source and the measurement points are
assumed to be located on a circular trajectory. Let in Fig. 6.1(b) too, the center of the
PEC scatterer coincides with the center of the coordinates. With the incident field 𝐸𝑧𝑖 ,
the scattered filed at (𝜌′ , ′ ), emanated from the circular PEC cylinder can be
expressed as below,
𝛽𝑠 2 𝐼 (2)
𝐸𝑧𝑠 (𝜌, 𝜌′ , , ′
)=− ∑𝑛=+∞
𝑛=−∞ 𝑐𝑛 𝐻𝑛 (𝛽𝑠 𝜌), 𝜌 > a (6.3)
4𝑤𝑠 𝜀
total field along the circular PEC boundary is zero (see [8] for detailed derivation).
(with 𝜌 > a)
where n is the order of the Hankel and Bessel function. Now, by introducing the
𝐸𝑧𝑠 (𝜌, 𝜌′ , , ′
) = 𝐼 𝐺(𝜌, 𝜌′ , , ′
) (6.5)
78
Coordinate
transformation
(0,0) (−xpix, − ypix)
(xpix,ypix) (0,0)
𝛽 2 (2) 𝐽𝑛 (𝛽𝑠 𝑎) − ′) ]
𝐺(𝜌, 𝜌′ , , ′)
= 4𝑤𝑠 𝜀 ∑𝑛=+∞ ′
𝑛=−∞ [𝐻𝑛 (𝛽𝑠 𝜌 ) (2) 𝐻(𝑛2) (𝛽𝑠 𝜌)𝑒𝑗𝑛( (6.6)
𝑠 𝐻𝑛 (𝛽𝑠 𝑎)
In Fig. 6.1(b), exciting each point as a line source at one time and measuring from
all other points on the circular trajectory, we have a measurement column vector,
𝑇
𝑬𝒎𝒆𝒂𝒔 = (𝐸12 , 𝐸13 , … , 𝐸𝑝𝑚 , … , 𝐸𝑃𝑀 ) where ‘p’ refers to the p-th line source and ‘m’
Then employing the idea of reciprocity, we hypothesize that these secondary sources
will radiate back to the location(s) of significant scatterers in the imaging domain. One
79
of the merits of this hypothesis is that, though the formulation is for PEC cylinders, we
have observed that the algorithm works for any other significant scatterers (other than
PEC) in the imaging domain. Now, all the secondary radiations from the secondary
line sources are summed-up at each pixel located at (𝑥𝑝𝑖𝑥 , 𝑦𝑝𝑖𝑥 ) in the imaging domain
𝛽 2
𝐼𝑆𝐹 (𝑥𝑝𝑖𝑥 , 𝑦𝑝𝑖𝑥 ) = 4𝑤𝑠 𝜀 ∑𝑃𝑝=1 ∑𝑀 ′
𝑚=𝑝+1‖𝐸𝑝𝑚 (𝜌 , 𝜌, ,
′)
𝐺(𝜌, 𝜌′ , , ′ )‖
(6.7)
𝑠
Here, P is the total number of line sources, M is the total number of measurement
points. To be noted is that the right hand side of (6.7) is calculated only after the
coordinate transformation to ensure the center of the pixel coincides with the center of
the coordinate system (see Fig. 6.2). This is necessary as per the requirement of the
formulation (6.4).
Wideband imaging techniques are sometimes more efficient than single frequency
methods. The higher frequency components can potentially give more resolution and
lower frequency components yields more stability in the reconstruction process [39].
Keeping that in mind, (6.7) is directly extended as below, to include ‘S’ number of
80
𝛼
(a) (b)
Figure 6.3. Pattern of a dipole: a) 3D pattern, b) a 2D cut in the elevation plane from
the 3D pattern.
𝑆 𝑃 𝑀
𝛽𝑠 2
𝐼𝑀𝐹 (𝑥𝑝𝑖𝑥 , 𝑦𝑝𝑖𝑥 ) = ∑ ∑ ∑ ‖𝐸𝑝𝑚 (𝜌′ , 𝜌, , ′)
𝐺(𝜌, 𝜌′ , , ′ )‖
4𝑤𝑠 𝜀
𝑠=1 𝑝=1 𝑚=𝑝+1
(6.8)
Finally, (6.8) is employed for each individual pixel of the domain to obtain its image.
For the purpose of easy interpretation, we normalize and scale the final image between
0 and 1.
We now extend our 2D algorithm to 3D imaging. One way of doing this would be
using the 3D Green’s function directly in the imaging formulation. However, for
simplicity, we will employ some multiplicative factors with the 2D Green’s function
81
to achieve the 3D imaging. To do that, we exploit the understanding of how the dipole
antennas radiate in a 3D medium. Specifically, we take into account two things: 1) the
pattern of the antenna, 2) the orientation of the antenna. Firstly, we explain the pattern
of a dipole and how we incorporate this information in the imaging algorithm. Figure
6.3 shows the radiation pattern of a typical dipole. Fig. 6.3(a) is the power magnitude
plotted in 3D and Fig. 6.3(b) shows a 2D cut in the elevation plane ( =constant). As
we can see from here, dipole has a donut shaped pattern in 3D space. In the elevation
plane as shown in Fig. 6.3(b), the pattern can be modeled as some power of co-
𝑝
sinusoidal function, [𝑐𝑜𝑠(𝛼) ] , where ‘p’ is a real positive number that can be
adjusted to match the pattern and 𝛼 is the angle from the azimuth plane as shown in
Fig. 6.3(b). So, while imaging a pixel in a plane not located in the azimuth plane of the
𝑝
dipole, the pattern, [𝑐𝑜𝑠(𝛼) ] must be incorporated. Another thing that has to be taken
in the formulation (6.8). To be noted is that, when we image a pixel which is out of the
azimuth plane of a dipole, we have to consider and correct for the orientation of the
dipole. As can be seen in Fig. 6.4, this can be done by employing the factor, 𝑐𝑜𝑠(𝛼).
82
𝑆 𝑃 𝑀
𝛽𝑠 2
𝐼3𝐷 (𝑥𝑝𝑖𝑥 , 𝑦𝑝𝑖𝑥 ) = ∑ ∑ ∑ ‖𝐸𝑝𝑚 (𝜌′ , 𝜌, , ′)
𝐺(𝜌, 𝜌′ , , ′ )𝐹
1,𝑇 𝐹2,𝑇 𝐹1,𝑅 𝐹2,𝑅 ‖
4𝑤𝑠 𝜀
𝑠=1 𝑝=1 𝑚=𝑝+1
(6.9)
Here,
F2,T/R= factor correcting antenna orientation for Transmit/Receive antenna (as the
Following the above discussion, we can set F1,T/R and F1,T/R as, F1,T/R =
𝑝
[cos(𝛼) ] 𝑇/𝑅 and F2,T/R = [cos(𝛼)] 𝑇/𝑅 . And, (6.9) can be written as,
Dipole Dipole
pos. #1 pos. #2
𝛼 plane B 𝛼
plane A
Tilted TM polarized
dipole dipole
Hence, eqn. (6.10) is the final formula for 3D imaging. To be noted is that, just like
the 2D imaging using (6.8), coordinate transformation for the calculation pertaining to
The three dimensional (3D) imaging set-up is shown in Fig. 6.5(a). Total 36 number of
dipole antennas are employed in 3 layers (12 dipoles in each layer) surrounding the
imaging domain. The dipoles resonate around 1 GHz and the imaging domain has
background permittivity 40 and conductivity 0.4 S/m with the anomaly permittivity of
80
and conductivity 0.1 S/m. 13 frequency samples are employed in the imaging
algorithm in the range 0.7-1.3 GHz. As usual, one dipole is excited at one time and the
others are receiving signals. The S-parameters measurements are carried out both
Fig. 6.5(b) shows the reconstructed image in 3D grid. To better understand the
image, different 2D cuts from the 3D image are shown in Fig. 6.6 (a)-(d). It is
84
observed that the 3D imaging algorithm successfully reconstructs the image in all
cases.
is that it can withstand more noise in the measurements. In Fig. 6.7, images using the
proposed method is compared with those using Gauss-Newton method. It can be seen
that the proposed method yields good quality image even in when the SNR is -3dB.
The proposed method can be applied in the case of on-body imaging. As shown in Fig.
4.7(a) earlier, the imaging is carried out after placing 8 dipole antennas, resonating
samples are employed in the imaging algorithm in the range 0.7-1.3 GHz. The
permittivity and loss tangent of the domain are set to values 59 and 0.39 for heart and
35 and 0.4 for lung (the bulk medium), respectively. The anomaly permittivity and
85
anomaly
antennas
(a)
(b)
86
(a) (b)
(c) (d)
Figure 6.6: 2D cuts from the 3D image, a) XY plane, Z=3cm, b) XY plane, Z= 0cm, c)
YZ plane, X=2cm, d) YZ plane, X= -2.5cm. Note that the anomalies are circular in XY
plane but rectangular in YZ plane.
87
Proposed Method Gauss-Newton Method
(a) (b)
SNR = 30 dB
(c) (d)
SNR = -3 dB
Figure 6.7: Images with different SNR values, left column: proposed method, right
column: Gauss-Newton method.
loss tangent are set to 80 and 0.1, respectively. Two sets of measurements (one is the
incident field and the other is the total field) will be carried out in two different time
instances. The goal is to image the growth of tumor, if any. To be noted is that this
human body model is a simplistic one which ignores the presence of skeleton and
88
other biological details. We remark, the effect of any other dielectric other than the
field (with stage-1 anomaly) from the total field (with stage-2 anomaly). That is why it
is anticipated that this simplistic human body model should mimic a realistic scenario
reasonably well.
tumor is
growing in size
(a) (b)
(c)
Figure 6.8: (a)-(b) Tumor is growing in size, (c) reconstructed image of the tumor
growth.
89
Fig. 6.8 shows the growth of tumor and also the reconstructed image on an elliptical
grid that closely matches the human body shape. Of importance is that, though the
shape information of the human body is not exactly known, the image reconstruction
90
Chapter 7: Conclusion
Microwave Imaging has been a research topic for over thirty years. Despite its
potentials, researchers have not been able to make it fully functional in a practical
setting. Efforts has been employed for the application of breast cancer imaging in the
past with limited success. The issues with MWT encompasses a multitude of factors,
such as, ill-posedness of the problem, low spatial resolution, high nonlinearity. In this
context, the focus of this dissertation was on developing fast, reliable and portable
modalities, like MRI, X-ray CT etc. in out-of-hospital setting. The goal was not to
have a very good image (as MWT resolution is limited anyways) rather to have some
information about the imaging domain through fast algorithms. In this process, several
method. A 2D image now can be obtained in less than 1 sec. This is very
imaging oil-gas flow in industrial pipes etc. are require real-time or high-speed
MWT algorithms.
91
b) Improving imaginary permittivity image: A method, the first to our
biological medium. The reason is, in this case, imaginary permittivity image
for pixel-by-pixel MWT has been proposed for the first time. Of importance is
that the method can handle shape uncertainties and is able self-calibrate the
data itself. The image dimensions (number of unknowns) has been reduced by
has been developed starting from the formulation of line source scattering by a
radiated secondary waves have been shown to converge to the same points of
92
where the scattered fields were originated from. This matrix inversion-free
algorithm is more tolerant against noisy data and can render successful image
demonstrated that the algorithm can successfully reconstruct the images in the
As can be seen in Fig. 5.5 and 5.8, the reconstructed images are blurry and has
coefficients. However, this would result in much larger ANN and eventually need
more extensive computation. With the ongoing advent of computational power, this
should not be difficult. Also, exploring advanced ANNs [83] can be another option to
Throughout the dissertation, the focus was to develop novel imaging algorithms and
keep the antennas simple. However, employing directive conformal antennas may
problem and worth investigating. For example, the imaging algorithm in chapter 6
93
develops the formulation using TM polarized dipole antennas. It would be really
fields from any antenna can be decomposed into TM and TE components to carry out
measurement hardware and measurement time. To overcome this issue, an ANN based
94
References
[1] E. C. Fear, S. C. Hagness, and M. A. Stuchly, “Confocal Microwave Imaging for
Breast Cancer Detection: Localization of Tumors in Three Dimensions,” IEEE
Trans. Biomed. Eng., vol. 49, no. 8, pp. 812–821, Aug. 2002.
[2] X. Li, S. K. Davis, S. C. Hagness, D. W. Weide, and B. D. V. Veen, “Microwave
imaging via space-time beamforming: Experimental investigation of tumor
detection in multilayer breast phantoms,” IEEE Trans. Microw. Theory Techn.,
vol. 52, no. 8, pp. 1856–1865, Aug. 2004.
[3] Q. Fang, P. M. Meaney, S. D. Geimer, A. V. Streltsov, and K. D. Paulsen,
“Microwave image reconstruction from 3-D fields coupled to 2-D parameter
estimation,” IEEE Trans. Med. Imag., vol. 23, no. 4, pp. 475–484, 2004.
[4] M. Ostadrahimi, A. Zakaria, J. LoVetri, and L. Shafai, “A near-field dual polarized
(TE-TM) microwave imaging system,” IEEE Trans. Microw. Theory Techn., vol.
61, no. 3, pp. 1376–1384, Mar. 2013.
[5] T. Rubæk,O. Kim, and P.Meincke, “Computational validation of a 3-D microwave
imaging system for breast-cancer screening,” IEEE Trans. Antennas Propag., vol.
57, no. 7, pp. 2105–2115, Jul. 2009.
[6] W. C. Chew, Waves and fields in inhomogeneous media, IEEE Press, Piscataway
NJ, 1995.
[7] C. Gilmore, P. Mojabi, A. Zakaria, M. Ostadrahimi, C. Kaye, S. Noghanian, L.
Shafai, S. Pistorius, and J. LoVetri, “A wideband microwave tomography system
with a novel frequency selection procedure,” IEEE Trans. Biomed. Eng., vol. 57,
no. 4, pp. 894–904, Apr. 2010.
[8] P. M. Meaney, M. W. Fanning, D. Li, S. P. Poplack, and K. D. Paulsen, “A clinical
prototype for active microwave imaging of the breast,” IEEE Trans. Microw.
Theory Techn., vol. 48, no. 11, pp. 1841–1853, Nov. 2000.
[9] C. Balanis, Advanced electromagnetic engineering, 2nd Edition, Wiley, Jan. 2012.
[10] J. L. Volakis and Kubilay Sertel, Integral equation methods for electromagnetics,
Sci Tech Publishing, Raleigh, NC, 2012.
95
[11] W. C. Chew and Y. M. Wang, “Reconstruction of two-dimensional permittivity
distribution using the distorted born iterative method,” IEEE Trans. Med. Imag.,
vol. 9, no. 2, pp. 218–225, Jun. 1990.
[12] M. Pastorino, Microwave Imaging. New York, NY, USA: Wiley, 2010, vol. 208.
[13] Q. Marashdeh et al., 'A nonlinear image reconstruction technique for ECT using
combined neural network approach,' Meas. Sci. Tech., vol. 17, no. 8, pp. 2097-
2103, 2006.
[14] A. Zamani, A. M. Abbosh, and A. T. Mobashsher, “Fast frequency-based
multistatic microwave imaging algorithm with application to brain injury
detection,” IEEE Trans. Microw. Theory Techn., vol. 64, no. 2, pp. 653-662, 2016.
[15] S. Salman, Z. Wang, E. Colebeck, A. Kiourti, E. Topsakal, and J.L. Volakis,
“Pulmonary edema monitoring sensor with integrated body-area network for
remote medical sensing,” IEEE Trans. Antennas Propag., vol. 62, no. 5, pp. 2787-
2794, Feb. 2014.
[16] M. A. Islam, A. Kiourti, and J. L. Volakis, “A novel method of deep tissue
biomedical imaging using a wearable sensor,” IEEE Sensors Journal, vol. 16,
no.1, pp. 265-270, 2015.
[17] L. Curiel, R. Chopra and K. Hynynen, “Progress in Multimodality Imaging: Truly
Simultaneous Ultrasound and Magnetic Resonance Imaging,” IEEE Trans. Med.
Imag., vol. 26, no. 12, pp. 1740-1746, Dec. 2007.
[18] J.P. Green, “Bayesian reconstructions from emission tomography data using a
modified EM algorithm,” IEEE Trans. Med. Imag., vol. 9, no. 1, pp. 84-93, Mar.
1990..
[19] G. Pratx, C.M. Carpenter, C. Sun, and L. Xing, “X-Ray luminescence computed
tomography via selective excitation: a feasibility study,” IEEE Trans. Med. Imag.,
vol. 29, no. 12, pp. 1992-1999, Dec. 2010.
[20] P.M. Edic, J. Saulnier, J.C. Newell, and D. Isaacson, “A real-time electrical
impedance tomography,” IEEE Trans. Biomed. Eng., vol. 42, no. 9, pp. 849-859,
Nov. 1995.
[21] R. Paradiso, G. Loriga, and N. Taccini, “A wearable health care system based on
knitted integrated sensors,” IEEE Trans. Inf. Technol. Biomed., vol. 9, no. 3, Sept.
2005.
96
[22] A. Bonfiglio and D. De Rossi, Eds., Wearable Monitoring Systems, Springer,
New York, NY, USA, 2011.
[23] T. Yilmaz and Y. Hao, “Compact resonators for permittivity reconstruction of
biological tissues”, in Proc. URSI General Assembly and Scientific Symposium,
pp. 1-4, Aug. 2011.
[24] A. Tarantola, Inverse problem theory and methods for model parameter
estimation, SIAM series on Other Titles in Applied Mathematics, 2005.
[25] J. Nocedal and S. Wright, Numerical Optimization, T. Mikosch, Ed., 2nd ed.
Springer: LLC, 2006.
[26] P. C. Hansen, Rank-Deficient and Discrete Ill-Posed Problems: Numerical
Aspects of Linear Inversion, SIAM series on Mathematical Modeling and
Computation, 1998.
[27] P. C. Hansen, Discrete Inverse Problems: Numerical Aspects of Linear Inversion,
SIAM series on Fundamentals of Algorithms, 2010.
[28] C. R. Vogel, Computational methods for inverse problems, SIAM series on
Frontiers in Applied Mathematics, 2002.
[29] M. Ostadrahimi, P. Mojabi, C. Gilmore, A. Zakaria, S. Noghanian, S. Pistorius,
and J. LoVetri, “Analysis of incident field modeling and incident/scattered field
calibration techniques in microwave tomography,” IEEE Ant. & Wireless Prop.
Lett., vol. 10, pp. 900-903, 2011.
[30] M. Ostadrahimi, P. Mojabi, S. Noghanian, L. Shafai, S. Pistorius, and J. LoVetri,
“A novel microwave tomography system based on the scattering probe technique,”
IEEE Trans. On Instrumentation & Meas., vol. 61, no. 2, pp. 379-390, 2012.
[31] J. Hsieh, “Computed Tomography: Principles, Design, Artifacts and Recent
Advances,” SPIE Press, 2003.
[32] https://en.wikipedia.org/wiki/X-ray
[33] J. Zaeytijd, A Franchois, C. Eyraud, and J. Geffrin, “Full-Wave three-
dimensional microwave imaging with a regularized gauss-newton method- theory
and experiment,” IEEE Trans. Antennas Propag., vol. 55, no. 11, pp. 3279–3292,
Nov. 2007.
97
[34] A. Desmal and H Bagci, “Nonlinear microwave imaging using Levenberg-
Marquardt method with iterative shrinkage thresholding,” Radio Science Meeting
(Joint with AP-S Symposium), USNC-URSI, Memphis, TN, USA, 2014.
[35] Z Wu, H McCann, L E Davis, J Hu, A Fontes, and C G Xie, “Microwave-
tomographic system for oil and gas multiphase flow imaging,” Meas. Sci. and
Tech.., vol. 20, no. 10, May 2009.
[36] Z Wu, “Developing a microwave tomographic system for multiphase flow
imaging: advances and challenges,” Trans. of the Institute of Meas. and Control,
pp. 1-9, Sep., 2014.
[37] P.M. Edic, J. Saulnier, J.C. Newell, and D. Isaacson, “A real-time electrical
impedance tomography,” IEEE Trans. Biomed. Eng., vol. 42, no. 9, pp. 849-859,
Nov. 1995.
[38] P. Monk, Finite element methods for maxwell's equations, Oxford, U.K.: Oxford
Univ. Press, 2003.
[39] J. L. Volakis, A. Chatterjee, and L. C. Kempel, Finite element method for
electromagnetics, IEEE Press, Piscataway NJ, 1998.
[40] J. F. Lee et al., "Numerical methods in antenna modeling," in: F. B. Gross (ed.),
Frontiers in Antennas: Next Generation Design and Engineering, pp. 443-502,
McGraw-Hill, New York, 2011.
[41] A. T. Nugroho and Z. Wu, “Inexact Newton backtracking method for solving
microwave tomography inverse problem,” Proceed. of the IEEE Intl.Conf. on
Imaging Sys. and Tech. (IST), 2015.
[42] M. Benedetti, M. Donelli, A. Martini, M. Pastorino, A. Rosani, and A. Massa,
“An innovative microwave-imaging technique for nondestructive evaluation:
Applications to civil structures monitoring and biological bodies inspection,” IEEE
Trans. Instrum. Meas., vol. 55, no. 6, pp. 1878–1884, Dec. 2006.
[43] N. K. Nikolova, “Microwave imaging for breast cancer,” IEEE Microw. Mag.,
vol. 12, no. 7, pp. 78–94, Dec. 2011.
[44] M. Ostadrahimi, A. Zakaria, J. LoVetri, and L. Shafai, “A near-field dual
polarized (TE-TM) microwave imaging system,” IEEE Trans. Microw. Theory
Techn., vol. 61, no. 3, pp. 1376–1384, Mar. 2013.
98
[45] T. Rubæk,O. Kim, and P.Meincke, “Computational validation of a 3-D
microwave imaging system for breast-cancer screening,” IEEE Trans. Antennas
Propag., vol. 57, no. 7, pp. 2105–2115, Jul. 2009.
[46] C. Gilmore, P. Mojabi, A. Zakaria, M. Ostadrahimi, C. Kaye, S. Noghanian, L.
Shafai, S. Pistorius, and J. LoVetri, “A wideband microwave tomography system
with a novel frequency selection procedure,” IEEE Trans. Biomed. Eng., vol. 57,
no. 4, pp. 894–904, Apr. 2010.
[47] P. M. Meaney, M. W. Fanning, D. Li, S. P. Poplack, and K. D. Paulsen, “A
clinical prototype for active microwave imaging of the breast,” IEEE Trans.
Microw. Theory Techn., vol. 48, no. 11, pp. 1841–1853, Nov. 2000.
[48] E. C. Fear, S. C. Hagness, and M. A. Stuchly, “Confocal Microwave Imaging for
Breast Cancer Detection: Localization of Tumors in Three Dimensions,” IEEE
Trans. Biomed. Eng., vol. 49, no. 8, pp. 812–821, Aug. 2002.
[49] R. K. Amineh, A. Trehan, andN. K. Nikolova, “TEM horn antenna for ultra-wide
band microwave breast imaging,” Progress In Electromagnetics Research B, vol.
13, pp. 59–74, 2009.
[50] A. Kiourti, C. Lee, and J. L. Volakis, “Fabrication of textile antennas and circuits
with 0.1 mm precision, IEEE Antennas Propag. Lett., vol. 15, pp. 151-153, 2016.
99
[55] P. M. Meaney, N. K. Yagnamurthy, and K. D. Paulsen, “Pre-scaled two-
parameter Gauss-Newton image reconstruction to reduce property recovery
imbalance,” Phys. Med. Biol., vol. 47, pp. 1101-1119, 2002.
[56] E. Malone, M. Jehl, S. Arridge, T. Betcke, and D. Holder, “Stroke type
differentiation using spectrally constrained multifrequency EIT: evaluation of
feasibility in a realistic head model,” Phys. Meas., vol. 35, pp. 1051-1066, 2014.
[57] O. Crabeck et. al., “Imaging air volume fraction in sea ice using non-destructive
X-ray tomography,” The Cryosphere, vol. 10, pp. 1125–1145, 2016.
[58] R. Rasel, C. Zuccarelli, Q. Marashdeh, L. Fan, and F.L. Teixeira, “Towards
multiphase flow decomposition based on electrical capacitance tomography
sensors,” IEEE Sensors Journal, 2017.
[60] Q. Marashdeh and F. L. Teixeira, “Sensitivity matrix calculation for fast 3-D
electrical capacitance tomography (ECT) of flow systems,” IEEE Transactions on
Magnetics, vol. 40, no. 2, pp. 1204-1207, 2004.
[64] J. Weber and J. Mei, "Bubbling fluidized bed characterization using electrical
capacitance volume tomography (ECVT)," Powder Technol., vol. 242, pp. 40-50,
July 2013.
100
[65] Q. Marashdeh, F. L. Teixeira, and L.-S. Fan, "Adaptive electrical capacitance
volume tomography," IEEE Sensors Journal, vol. 14, no. 4, pp. 1253-1259, 2014.
[70] J. D. Ullman et. al., “Mining of Massive Datasets,” published online at:
http://infolab.stanford.edu/~ullman/mmds/book.pdf
101
[77] J. Devaney, “Time reversal imaging of obscured targets from multistatic data,”
IEEE Trans. Antennas Propagat., vol. 53, pp. 1600-1610, 2005.
[78] M. E. Yavuz and F. L. Teixeira, “Full time-domain DORT for ultrawideband
electromagnetic fields in dispersive, random inhomogeneous media,” IEEE Trans.
Antennas Propagat., vol. 54, no. 8, pp. 2305-2315, 2006
[79] D. Liu, S. Vasudevan, J. Krolik, G. Bal, and L. Carin, “Electromagnetic time-
reversal source localization in changing media: Experiment and analysis,” IEEE
Trans. Antennas Propagat., vol. 55, pp. 344-354, February 2007.
[80] M. E. Yavuz and F. L. Teixeira, “Space-frequency ultrawideband time-reversal
imaging,” IEEE Trans. Geosci. Remote Sens., vol. 46, no. 4, pp. 1115-1124, 2008.
[81] M. E. Yavuz and F. L. Teixeira, “Ultrawideband microwave remote sensing and
imaging using time-reversal techniques: A review,” Remote Sensing, vol. 1, no. 3,
pp. 466-495, 2009.
[82] A. Fouda and F. L. Teixeira, “Imaging and tracking of targets in clutter using
differential time-reversal techniques,” Waves in Random and Complex Media, vol.
22, no. 1, pp.66-108, 2012.
[83] https://cs.stanford.edu/~quocle/tutorial2.pdf
102