framework for the diagnosis of non-alcoholic fatty liver
TRANSCRIPT
The (travel) times they are a changing: A ComputationalFramework for the Diagnosis of Non-Alcoholic Fatty Liver
Disease (NAFLD)
by
Alex Benjamin
M.S. in Computation for Design and OptimizationMassachusetts Institute of Technology, 2017
Submitted to the Department of Computation for Design and Optimizationin partial fulfillment of the requirements for the degree of
Master of Science in Computation for Design and Optimization
at the
MASSACHUSETTS INSTITUTE OF TECHNOLOGY
May 2017
2017 Massachusetts Institute of Technology. All rights reserved.
Signature redactedAuthor............
Certified by.......
,V.........................---.
Computation for Design and OptimizationJune 5, 2014
A
Signature redacted-~ )
Accepted by..
...........................Brian W. Anthony
Principal Research Scientist, Department of Mechanical Engineering
(Thijiesis Supervisor
S ig n a t u re re d a c te d ............MASACUSETSINSITTE .. ... .. ... .. ........
MASSACHSET ILNSTITUTEOF TECHNOLOGY
JUN 19 2017
LIBRARIESARCHIVES
NA4ua.bia i jiconstantinouCo-Director, Computation f*15e ign and Optimization
.. . . . . . . . .
The (travel) times they are a changing: A Computational Framework for the Diagnosis of Non-
Alcoholic Fatty Liver Disease (NAFLD)
by
Alex Benjamin
Submitted to the Department of Computation for Design and Optimizationon May 24, 2017 in Partial Fulfillment of the
Requirements for the Degree of Master of Science inComputation for Design and Optimization
ABSTRACT
We propose and validate a non-invasive method to diagnose Non-Alcoholic Fatty Liver Disease(NAFLD). The proposed method is based on two fundamental concepts: 1) the speed of sound ina fatty liver is lower than that in a healthy liver and 2) the quality of an ultrasound image ismaximized when the beamforming speed of sound used in image formation matches the speed inthe medium under examination. The proposed method uses image brightness and sharpness asquantitative image-quality metrics to predict the true sound speed and capture the effects of fatinfiltration, while accounting for the transmission through subcutaneous fat. Validation using non-linear acoustic simulations indicated the proposed method's ability to predict the speed of soundwithin a medium under examination with little sensitivity to the transducer's frequency (errors lessthan 2%). Additionally, ex vivo testing on sheep liver, mice livers, and tissue-mimicking phantomsindicated the method's ability to predict the true speed of sound with errors less than 0.5% (despitethe presence of subcutaneous fat) and its ability to quantify the relationship between fat contentand speed of sound. Additionally, this work starts to create a framework which allows for thedetermination of the spatial distribution of the longitudinal speed of sound, thereby providing apromising method for diagnosing NAFLD over time.
Thesis Supervisor: Brian W. AnthonyTitle: Principal Research Scientist, Department of Mechanical Engineering & Institute forMedical Engineering and Sciences
2
3
ACKNOWLEDGEMENTS
First and foremost, I would like to thank my advisor, Dr. Brian Anthony (aka The Prince of
Darkness); thank you for guiding me, motivating me, and most of all, for giving me the opportunity
to pursue my true passion.
Next, I would like to thank my brother, Rishon, to whom I owe an insurmountable debt; I
wish I could tell you how much you mean to me, but that would be a futile attempt. You are, in
every way possible, an exemplary human being and a true inspiration.
A heartfelt thank you to my parents who have sacrificed so much and have asked for nothing
in return. Thank you for being such amazing and supportive parents and for making me so proud
to be your son. I love you both very much.
A special thanks to Dr. Pavel Grinfeld, a man who has been my teacher, my friend, my
mentor, and my guide for the past 7 years. Pavel, as much as you would like to deny it, you have
played a huge role in helping me achieve what I have.
Next, I would like to thank my closest friends in the Device Realization Lab (Anne, Ina,
Moritz, Athena, Shawn, Rebecca, Hank, David, Sissy, Melinda, and Robin). In so many ways, you
have been my family for the last two years. I will cherish every memory and every moment that I
have spent with all of you. Even more than my family, you have seen me at my best and worst
times and that vulnerability is the price of a long-lasting friendship ... one that I am more than
willing to pay.
A special thank you to Sid. Sid, you have been like a brother to me, and I don't think I need
to say any more than that, because that is a special bond. One that is much stronger than epoxy!
A heartfelt thank you to my friend Aashita for always being there for me. Sometimes when I
truly didn't deserve it. You are and will always be a very special part of my life.
Dan and Elise, despite not being at MIT, you two will always have a special place in my life.
If not in the present, you have certainly shaped my life in the past. I will never forget that.
Shibani, I would love to thank you for making my time at MIT so memorable in so many
ways, but you and I both despise formal "thank yous!" and so, I trust you to know what I mean.
4
Lastly, I would like to thank a very special person in my life. A man who embodies the
qualities of perseverance, courage, determination, resolve, and ambition. The only man who could
look me dead in the eyes and say this:
"You build your future, it isn't handed to you" - Francis J. Underwood
(two taps on a wooden desk...).
5
6
CONTENTS
Abstract.......................................................................................................................................... 2
Acknowledgements ....................................................................................................................... 4
Contents ......................................................................................................................................... 7
Figures............................................................................................................................................ 9
Tables ........................................................................................................................................... 11
1 Introduction.............................................................................................................................. 13
1.1 Overview of Non-Alcoholic Fatty Liver Disease ..................................................... 13
1.2 Motivation for Using Longitudinal Speed of Sound to Diagnose Steatosis .............. 14
1.3 Longitudinal Speed of Sound as an Indicator of Additional M aladies ...................... 15
2 M ethods for Q uantifying Longitudinal Speed of Sound ................................................... 17
2.1 Existing Methodologies for Quantifying the Longitudinal Speed of Sound In Vivo.... 17
2.2 Proposed Methodology for Quantifying the Longitudinal Speed of Sound .............. 18
2.3 Explanation of Image Quality M etrics...................................................................... 19
2.4 Accounting for the Transmission Through Subcutaneous Fat.................................. 22
3 Validation Using Non-Linear Acoustic Simulations.......................................................... 24
3.1 Sim ulation Setup........................................................................................................... 24
3.2 Simulation Results ..................................................................................................... 25
4 Experim ental Validation ..................................................................................................... 30
4.1 Experimental Setup................................................................................................... 30
4.2 Experim ental Results ................................................................................................. 33
5 Iterative Speed of Sound Estimation................................................................................... 42
5.1 Algorithm Description .............................................................................................. 42
5.2 Experimental Validation ............................................................................................ 44
6 Towards a Framework for Longitudinal Speed of Sound Maps...................................... 47
7
6.1 M otivation for Speed of Sound M aps............................................................................... 47
6.2 S3: A Spatial and Spectral Sharpness M easure [V u et al.]............................................. 48
6.3 D etailed D escription of the A lgorithm ............................................................................ 48
6.4 Prelim inary Results (Ex Vivo).......................................................................................... 50
7 Conclusions and Future W ork............................................................................................. 54
R eferences.................................................................................................................................... 57
8
FIGURES
Figure 1: Bi-layer setup to demonstrate the method for accounting for inhomogenous tissue
stru ctu re s ....................................................................................................................................... 2 3
Figure 2: Predicted speed of sound in the simulation medium using the brightness metric (A) and
combined sharpness metric (B) The transducer frequency was held fixed at 4 MHz while the true
speed was varied as follows: 1460 m/s, 1500 m/s and 1540 m/s............................................... 26
Figure 3: Predicted speed of sound in the simulation medium using the brightness metric (A) and
combined sharpness metric (B). The true speed within the medium was held fixed at 1500 m/s
while the transducer frequency was varied as follows: 2 MHz, 3 MH...................................... 28
Figure 4: Diagrammatic representation of the pulse-echo setup for determination of ex vivo
ground-truth speed of sound values. (b) Image of the experiment being performed to determine
ground truth speed of sound in an ex-vivo sheep liver using the............................................... 31
Figure 5: Measurement setup on the Visualsonics Vevo 2100 ultrasound system (40 MHz probe);
the probe was in direct contact with the ex-vivo mice livers and held by an imaging station that
allows for the hands-free imaging of the sample ..................................................................... 32
Figure 6: Pathology results of the sheep liver show a lack of large fat vacuoles within the
parenchyma as seen with Oil Red 0 (a) and H&E stain (b). The ground truth measurements
indicated that the speed of sound within the liver was 1540 m/s (+/- 0.5%), and estimations using
the brightness metric, FFT metric and PCA metric were found to be 1540 m/s each (c)...... 34
Figure 7: Speed of sound prediction in a lean sheep liver (1500 m/s); the b-mode image in the
lower left corner depicts the olive oil layer (true speed of sound = 1400 m/s (+/- 0.5%)) sitting atop
the liver parenchyma (true speed of sound = 1540 m/s (+/- 0.5%)). ......................................... 35
Figure 8: Predicted speed of sound in the lowest layer i.e. phantom B (1500 m/s), skewed by the
transmission through phantom A (true speed = 1440 m/s). The roughly 2.59% error was corrected
by accounting for this transmission and resulted in a renewed estimate of 1551 m/s (0.71% error)
for p h an to m B ............................................................................................................................... 3 7
Figure 9: Predicted speed of sound in the control specimen (1500 m/s) along with pathology
results that demonstrate a lack of fat vacuoles (top); predicted speed of sound in the fatty specimen
(1460 m/s) consistent with the presence of large fat vacuoles in the liver parenchyma........... 39
9
Figure 10: Correlation (R^2=0.75) between predicted speed of sound using sharpness metrics and
absolute fat content. Batch numbers are as follows: Batch 1 (< 5% Fat), Batch 2 (5 - 10 % Fat),
Batch 3 (10 - 15 % Fat), Batch 4 (15 - 20 % Fat) and Batch 5 (> 20% Fat) ........................... 41
Figure 11: Error (logarithmic scale) associated with the iterative bi-section method as a function
of the number of iterations. As seen in the figure, the error decreases linearly in the logarithmic
scale, therby indicative of its underlying exponential decay. .................................................. 45
Figure 12: Variation of initial guesses within the bi-section method. As seen in the figure, speed
I starts at 1400 m/s and speed 2 at 1620 m/s. After 12 iterations, both values converge to 1540 m/s
(ground truth indicated by dashed green line) to within tolerance (0.1 m/s)............................. 46
Figure 13: B-mode image of the CIRS Liver Phantom acquired using the GE Logiq E9 at a
beam form ing speed of 1540 m /s............................................................................................... 51
Figure 14: Speed of sound map for the CIRS Liver Phantom obtained by optimizing the S3
sharpness map over a range of beamforming speeds................................................................. 52
Figure 15: Speed of sound map (blurred with a Gaussian filter (a = 2)) for the CIRS Liver
Phantom obtained by optimizing the S3 sharpness map over a range of beamforming speeds. .. 53
10
TABLES
Table 1: Summary of algorithm validation for fixed center frequency (4 MHz) and varying true
speeds of sound within the medium (1460 m/s, 1500 m/s and 1540 m/s)................................ 27
Table 2: Summary of algorithm validation for a fixed speed of sound (1500 m/s) and varying
transducer frequencies (2 M Hz, 3 M Hz, 4 M Hz)..................................................................... 29
Table 3: Predicted speed of sound in mice livers using FFT metric ....................................... 40
Table 4: Predicted speed of sound in mice livers using PCA metric ....................................... 40
11
12
CHAPTER
1INTRODUCTION
1.1 Overview of Non-Alcoholic Fatty Liver Disease
Non-alcoholic fatty liver disease (NAFLD) is a chronic disease defined as the presence
of hepatic steatosis (> 5% fat by volume) in the absence of any known cause and minimal
alcohol use [1]. It has an estimated worldwide prevalence of 45% [1, 2]. In the United States,
NAFLD afflicts roughly 30% of the general population, and by 2020, NAFLD is projected to
be the leading indicator and cause of liver transplantations in the world [3, 4]. Moreover,
NAFLD has been recognized as an independent cardiovascular disease risk factor with a 70%
overall mortality increase, due to a roughly 300% increase in cardiovascular disease mortality
[5]; it has also been associated with severe metabolic impairments such as insulin resistance,
hypertension, obesity, and Type 2 diabetes mellitus [6, 7].
The spectrum of NAFLD is divided into two categories: (1) simple steatosis (NAFL),
which makes up 70-75% of cases, is defined by excess liver fat without inflammation and (2)
steatohepatitis (NASH), which makes up 25-30% of cases, is defined by excess liver fat with
inflammation and/or fibrosis [8, 9]. Previously, it was believed that NAFL is non-progressive,
and does not result in liver fibrosis, whereas NASH results in chronic liver injury and
progressive fibrosis [8]. However, recent longitudinal paired biopsy studies have shown that
even patients with NAFL may progress to NASH [9]. In light of these results, quantitative
13
characterization of both NAFL and NASH, and their distinction, is highly relevant for effective
diagnosis and treatment.
A liver biopsy is the accepted reference standard for the diagnosis of NAFLD and
distinction of NAFL and NASH. However, biopsies are costly, invasive, associated with
morbidity and uncommon mortality, and are subject to significant variability and sampling
error. In a disease that affects roughly 30% of the population, it is impractical for diagnosis
and risk stratification. As a result, clinicians prefer to use alternative tests such as serum
biomarkers (e.g. liver enzymes), and radiological imaging tests [9, 10]. These techniques have
low sensitivity and specificity which is further exacerbated by the fact that as many as 50% of
patients with NAFLD have normal liver chemistries [11]. There is a need for a methodology
that allows for the quick and reliable quantification of liver fat at the point of care. This thesis
proposes and evaluates a method for the estimation of liver fat by the use of the longitudinal
speed of sound in liver tissue.
1.2 Motivation for Using Longitudinal Speed of Sound to
Diagnose Steatosis
Fat has a relatively lower longitudinal speed of sound (~1400 m/s) than healthy liver
tissue (~1540 m/s) [1, 2]. Consequently, the expected speed of sound in a fatty liver will be
lower than that in a healthy liver [12, 14]. Ex vivo results have demonstrated this expected
speed reduction in human liver samples; however, the relative contribution of the amount of
water and fat in the tissue were confounding factors in these studies. O'Brien et al. showed a
speed of sound reduction of about 50 m/s in rats with liver lipid content ranging from a normal
of 2-5% to slightly more than 20% [14]. This evidence suggests that the ability to accurately
14
and consistently estimate the longitudinal speed of sound in vivo might provide a viable
alternative to biopsies in the realm of diagnosing NALFD.
1.3 Longitudinal Speed of Sound as an Indicator of
Additional Maladies
The longitudinal speed of sound in vivo is affected by the pathology of the tissue under
examination [15, 16]. In addition to NAFLD, the longitudinal speed of sound in vivo is an
indicator of a host other maladies as well. Consider for example, the use of quantitative
ultrasound to determine the bone quality (specifically bone density) in premature and full-term
infants. Within the pediatric community, metabolic bone disease of prematurity (MBDP) is a
common and significant problem that often gives rise to fractures, osteomalacia, and
osteoporosis [17]. The current standard for measuring bone density is dual energy X-ray
absorptiometry (DXA). However, in recent decades, quantitative ultrasound has emerged as a
viable alternative. It has been demonstrated that the longitudinal speed of sound and
attenuation coefficient in bone correlates to bone density [17]. In addition to being inexpensive
and quick, quantitative ultrasound has the benefit of not exposing infants to ionizing radiation
and the potential to provide accurate and reliable diagnoses at point of care.
Speed of sound estimation is also of relevance in the field of breast imaging and early
detection of breast cancer. Breast density, as assessed by mammograms, reflects breast tissue
composition [18]. Breast has a relatively low longitudinal speed of sound (~1430 m/s) due to
the fact that it is primarily composed of fat. However, tumors within the breast tissue have
higher sound speeds and in addition to being relevant precursors of breast cancer, tend to cause
15
upshifts in the bulk speed of sound [18]. Consequently, the ability to predict the longitudinal
speed of sound in vivo is of importance in monitoring breast health over time.
In summary, the ability to accurately and consistently measure the longitudinal speed of
sound in vivo may improve the quality of ultrasound images and to leverage the low-cost,
portability, and safety of quantitative ultrasound technology for a variety of diagnostic
applications.
16
CHAPTER
2METHODS FOR QUANTIFYING
LONGITUDINAL SPEED OF SOUND
2.1 Existing Methodologies for Quantifying the
Longitudinal Speed of Sound In Vivo
Estimating the speed of sound in vivo is a non-trivial task, because of the lack of ground-
truth values for travel distance. Most commercial ultrasound scanners use an assumed, fixed
speed of sound (1540 m/s) in reconstructing images. This assumption is erroneous, because
the speed of sound in human tissues is known to vary over a range from less than 1450 to over
1600 m/s [19, 20]. Various methods have been proposed to estimate the longitudinal speed of
sound in vivo. Robinson et al. classified these methods into transmission and pulse-echo [16].
Transmission methods measure sound wave propagation time between the transmitter and
receiver whereas pulse-echo methods estimate sound propagation speed by processing the
associated backscattered data. Pulse-echo methods are divided into three categories: image
aberration methods, intersecting beam transit-time methods and axial techniques [16]. A
thorough review of these methods can be found in [16]. Additionally, Anderson and Trahey
[16] proposed a method to determine the speed of sound in vivo based on the best quadratic fit
of the delay patterns applied to individual transducer elements. Napolitano et al. estimated the
17
speed of sound by analyzing the frequency domain of the B-mode image; Cho et al estimated
the speed of sound by adjusting echo signals from several individual elements while Shin et
al. estimated the sound speed using blind deconvolution [16]. Various other techniques have
been developed in the last decade as well to estimate the true speed of sound; some of the
commonly used parameters include backscatter coefficients, speckle size, echo amplitude, and
attenuation coefficients, [14, 19, 20, 21]. All these methods exploit the fact that when the
assumed speed of sound in the beamformer matches that in the tissue being examined, some
measure of the quality of the B-mode image is maximized. More specifically, in speckle-
dominated B-mode images (e.g. those of the liver parenchyma), speckle brightness is
maximized and speckle size is minimized when the appropriate speed of sound is used by the
beamformer during image formation [23]. The use of speckle to assess quality of beam
formation arises from the work of Smith and Wager [22]. They demonstrated, by the use of
speckle statistics, that speckle size is dependent on the use of an accurate speed of sound during
beam formation and also that the smallest speckle dimensions occur near the focus. This
concept was extended by Nock and Trahey who optimized speckle characteristics by adjusting
beam formation delays until the brightness was the greatest at the desired focus [23].
2.2 Proposed Methodology for Quantifying the
Longitudinal Speed of Sound
In this study, we quantify the bulk speed of sound by analyzing two B-mode image
metrics: the sharpness in a region of interest (ROI) around the expected focal zone and the
brightness of the image around the expected focal zone. The brightness of the image is obtained
using a spatial average of the envelope of the corresponding IQ data for a fixed band around
18
the intended focal zone. The IQ data is the modulated radiofrequency data obtained from the
commercial scanner. The sharpness of the B-mode image is quantified using two different
methods: i) the proportion of Fast-Fourier Transform (FFT) coefficients above a defined
threshold of the zero-frequency (i.e. DC) component of the image spectrum [24] and ii) the
singular values of spatial, intensity gradient of the image [25]. The above metrics were selected
for their ease of real-time implementation, thereby offering a technique that can be performed
on current ultrasound imagers at point of care. These metrics can be calculated by operating on
B-mode images and radiofrequency (RF) data (beamformed, envelope-detected data) with
minimal computational cost, and no calibration (no relative measurements or comparisions to
reference phantoms), thus making them compatible with existing commercial ultrasound
machines [19].
2.3 Explanation of Image Quality Metrics
The following three image-quality measures were used to estimate the speed of sound in
speckle-dominated B-mode images: 1) Image Brightness (MB), 2) Image Sharpness: Fast
Fourier Transform Coefficients (MF), and 3) Image Sharpness: PCA Analysis (Mp).
Focal Zone Image Brightness (Brightness): For an assumed beamformation speed of sound, the
RF data envelopes are determined using the Hilbert Transform. The samples corresponding to
a band around the expected focal zone (Bf) are extracted and a spatial mean is determined across
all available A-lines (A). This results in an estimate of the average brightness around the focal
zone. The following equations summarizes the metric:
MB = H(F)(1)A B )
19
where H is the Hilbert Transform. The rationale behind analyzing the echogenicity around the
intended focal zone is as follows: in commercial scanners, beam steering is achieved by the
application of appropriate time delays across the active aperture [20]. In order to determine the
timing delays, most scanners make an assumption about the speed of sound within the medium
e.g. 1540 m/s. If this assumption is correct, the wave fronts from the individual elements will
constructively interfere at the intended focal zone and maximize the corresponding
echogenicity. More often than not, however, this assumption is faulty and the echogenicity at
the intended focal zone is diminished due to inadvertent destructive interference.
Image Sharpness: FFT Coefficients (FFT): Errors in the assumed beamforining speed of sound
cause lateral blurring within the resulting B-mode image [20, 22]. In order to quantify the
resulting blurring around the focal zone, the image sharpness metric proposed in [24] is utilized:
Th
MF = Mx gratr2haMFN
DCwhere Th is the number of FFT coefficients whose magnitudes is greater than where DC is
1000
the magnitude of the zero-frequency component, and M and N are the dimensions of the image.
The metric fundamentally extracts the effect of speed of sound mismatches and is easily applied
to a B-mode image with no calibration and minimal processing. The region of interest (ROI)
for this method is centered around the intended focal zone, but can vary with minimal effects
on the accuracy. The intuition behind this metric is that the largest FFT co-efficient of an image
is its zero-frequency component (i.e. DC component). The proposed metric MF extracts the
number of coefficients that are larger than a defined fraction of the DC component i.e. the
number of high-frequency components within the image. Since a mismatche between the
assumed speed of sound and the true speed of sound within the medium causes lateral image
blurring, sharper images (one with a larger number of high frequency components) are more
20
likely to have been acquired at the true sound speed within the medium (i.e. beamforming speed
= true speed within the medium)
Image Sharpness: PCA Analysis (PCA): The metric proposed by [25] is utilized to quantify
the magnitude of the intensity gradients within the B-mode image and thereby obtain a measure
of the associated blurring. The metric is summarized as follows [25]:
P= S (3)
where s, is the largest singular value corresponding to the matrix, G given by:
G = [Ix ly] (4)
where Ix and ly are vectors corresponding to the directional gradients of intensity within the
image, and E is a small, arbitrary number (usually 5e-4)[ 25]. The largest singular vector
corresponds to the direction along which the intensity gradient is the largest and the largest
singular value captures the magnitude of this variation. It is noted that this metric requires only
a region of interest (ROI) within the B-mode image and no additional calibration. As was
mentioned above, if the beamforming speed does not match the true speed of sound within the
medium, the resulting ultrasound image experiences lateral blurring. The proposed metric (Mp)
fundamentally extracts the sharpness of a given image by analyzing the magnitude of its spatial
intensity gradients. As such, sharper images would have larger intensity gradients and would
correspond to beamforming speeds that are close to the true speed within the medium.
Combination of Image Sharpness Metrics: The two distinct image sharpness metrics detailed
above were combined to yield a single sharpness metric:
MS = (MF)(lr)(Mp)r (5)
where r was chosen to be 0.25. The value was chosen optimally over the course of validating
the proposed method.
21
2.4 Accounting for the Transmission Through
Subcutaneous Fat
A common problem that arises during the estimation of the longitudinal speed of sound
in vivo is the depth inhomogeneity of the underlying tissue i.e. the presence of layers.
Ultrasound typically propagates through a layer of subcutaneous fat before entering the organ
of interest. This propagation can skew the speed of sound prediction within the organ of interest
[21].
A correction metric, similar to the one presented by Weijers et al 2012, shows that the
image can be segmented into the two layers (organ layer, and skin/subcutaneous fat layer) and
then linearly proportioned to account for the transmission through the occluding layer of
subcutaneous fat. Weijers et al. assumes that even though the speed of sound within the organ
of interest is unknown, the speed through the occluding layer (subcutaneous fat) is well-known.
We propose a method that is similar to the one proposed by [21], but extracts the speed of sound
in the successive layers by shifting the intended focal zones. For example, if the configuration
is a bilayer setup (Layer 1 over Layer 2), as seen in Figure 1, the proposed algorithm is first
used to extract the speed of sound in Layer I (occluding layer) and then in Layer 2 (layer of
interest); the image segmentation method described above is then used to correct for the
transmission through Layer 1 (linear proportioning). This is captured by the following equation:
Ctotal = (% Layer 1 )(cl) + (% Layer 2)(c 2 ) (6)
22
-c4.'0.
C(U4-a0I-
Layer 2 (c2 )
CO
CD'a
0D
Figure 1: Bi-layer setup to demonstrate the method for accounting for inhomogenous tissue structures
where the percentages of each layer are defined as follows:
% Layer 1 =
% Layer 2 =
Layer 1 Depth
Total Depth
Layer 2 Depth
Total Depth
(7)
(8)
23
CHAPTER
3VALIDATION USING NON-LINEAR
ACOUSTIC SIMULATIONS
3.1 Simulation Setup
A MATLAB-based, Fast Object-oriented C++ Ultrasound Simulation (FOCUS) was
utilized to validate the proposed technique. FOCUS utilizes the fast near-field method to solve
the non-linear, wave equation in three-dimensional space [26]. A tissue-mimicking domain was
constructed with the following dimensions: 5 cm x 1 cm x 6 cm (x, y and z). A scatterer density
of 1.7 x 103 scatterers/cm 3 was used and point scatterers were randomly distributed within the
bulk of the phantom. Scatterers were designed to have randomly distributed reflection
coefficients. Simulations were conducted with a 256-element array with a radius of curvature
equivalent to the focal depth. The element width was set to the effective wavelength of the
transducer, the element height was set to 20 mm, the kerf was set to 50 microns and the number
of active elements was set to 128. The sampling frequency was set to 200 MHz to eliminate
aliasing effects and the axial and azimuthal focal depths were held constant at 3 cm (sound wave
propagation was along the z direction).
In order to simulate the effect of uncertainty in the speed of sound, the true speed of sound
within the computational domain was held fixed at 1460 m/s, 1500 m/s, and 1540 m/s (each
simulated independently), while the assumed speed of sound for the beamforming process was
24
varied from 1400 m/s to 1600 m/s in increments of 10 m/s, and the transducer frequency was
held constant at 4 MHz.
In order to demonstrate the invariance with respect to the transducer frequency, the
transducer's center frequency was varied as follows: 2 MHz, 3 MHz, 4 MHz (each simulated
independently) while the speed of sound in the medium was held fixed at 1500 m/s; once again,
the assumed speed of sound for the beamforming process was swept from 1400 m/s to 1600 m/s
in increments of 10 m/s. The outputs of the simulations were the resulting beamformed RF data
and corresponding B-mode images.
3.2 Simulation Results
Figure 2 shows the normalized image-quality metrics, brightness metric (MB) (Fig. IA)
and combined sharpness metric (Ms) (Fig. IB), as a function of the beamforming speed of sound
for a transducer frequency of 4 MHz and true speeds of 1460 m/s, 1500 m/s and 1540 m/s within
the medium.
25
A
1350 1370 1390 1410 1430 1450 1470 1490 1510 1530 1550 1570 1590 1610 16
Speed of Sound (m/s)
-6-1460 m/s -0-1500 rn/s -*-1540 m/s
1350 1370 1390 1410 1430 1450 1470 1490 1510 1530 1550 1570 1590 1610 1630
Speed of Sound (m/s)Figure 2: Predicted speed of sound in the simulation medium using the brightness metric (A) and combined
sharpness metric (B) The transducer frequency was held fixed at 4 MHz while the true speed was varied as
follows: 1460 m/s, 1500 m/s and 1540 m/s.
26
.4
1
0.9
0.8
0.7
0.6
0.5
0.4
1
0.95
4Cu
B
A
-________________
0.9
0.85
0.8
0.75
-A- 1460 mn/s -<D-1500 m/s -*- 1540 m/s
30
For both metrics, the predicted sound speed corresponds to a maximum normalized value of
1.0. The dashed lines represented the true speeds of sound within the medium: 1460 m/s (black),
1500 m/s (blue), 1540 m/s (red). Table 1 summarizes the results:
Table 1: Summary of algorithm validation for fixed center frequency (4 MHz) and varying true speeds of
sound within the medium (1460 m/s, 1500 m/s and 1540 m/s)
True Speed of Sound (Frequency)
1460 m/s (4 MHz)
1500 m/s (4 MHz)
1540 m/s (4 MHz)
1460 m/s (4 MHz)
1500 m/s (4 MHz)
1540 m/s (4 MHz)
Predicted Speed of Sound
1480 m/s
1500 m/s
1530 m/s
1470 m/s
1500 m/s
1470 m/s
Percent Error
1.37%
0.00%
0.65%
0.69%
0.00%
4.54%
As can be seen in the table above, each of the metrics independently predicts the true speed of
sound within the medium, thereby demonstrating the algorithm's invariance with respect to the
true speed within the medium.
Figure 3 shows the normalized image-quality metrics, brightness (MB) (Fig. 2A) and
sharpness (Ms) (Fig. 2B), as a function of the beamforming speed of sound for transducer
frequencies of 2 MHz, 3 MHz and 4 MHz, and a true speed of 1500 m/s within the medium.
For both metrics, the predicted sound speed corresponds to a maximum normalized value of
1.0. The dashed green line represents the true speeds of sound within the medium i.e. 1500 m/s.
27
Metric
MB
MB
MB
Ms
Ms
Ms
1
0.9 -
0.8 -
0.7
0.6
0.5 -
0.4135
-A-2 MH
1450 1500
Speed of Sound (m/s)
z -0-3 MHz
1450 1500
Speed of Sound (m/s)
1550 1600 1650
-0-4 MHz
1550 1600 1650
Figure 3: Predicted speed of sound in the simulation medium using the brightness metric (A) and combined
sharpness metric (B). The true speed within the medium was held fixed at 1500 m/s while the transducer
frequency was varied as follows: 2 MHz, 3 MHz
28
A
S
1400
B1
3 0.9
0.8
0.71350 1400
-C>-3 MHz -*-4 MHz-1&2 MHz
0
Table 2 below summarizes the results:
Table 2: Summary of algorithm validation for a fixed speed of sound (1500 m/s) and varying transducer
frequencies (2 MHz, 3 MHz, 4 MHz)
Metric True Speed of Sound (Frequency) Predicted Speed of Sound Percent Error
MB
MB
1500 m/s (2 MHz)
1500 m/s (3 MHz)
1500 m/s (4 MHz)
1500 m/s (2 MHz)
1500 m/s (3 MHz)
1500 m/s (4 MHz)
Ms
Ms
1500 m/s
1500 m/s
1500 m/s
1510 m/s
1510 m/s
1470 m/s
0%
0%
0%
0.67%
0.67%
2.00%
As can be seen in Table 2, each of the metrics independently predicts the true speed of sound
within the medium with minimal error (less than 2%) without regard to the transducer's frequency,
thereby demonstrating its invariance with respect to the transducer's center frequency.
29
CHAPTER
4EXPERIMENTAL VALIDATION
4.1 Experimental Setup
The proposed method was tested in the following experimental configurations: i) sheep
liver (N=1) ii) sheep liver immersed in olive oil to simulate transmission through subcutaneous
fat iii) bilayer phantom (tissue mimicking phantoms of different speeds of sound in a stacked
configuration) iv) fatty and control mice liver (N = 2) v) mice livers with varying grades of fat
infiltration (N = 14).
Both layers in the bilayer phantom were calibrated phantoms manufactured by the
Computerized Imaging Reference Systems (CIRS). The sheep liver was obtained from a local
butcher shop in Cambridge, Massachusetts under Massachusetts Institute of Technology
Committee on Animal Care approval (E17-07-0320) and the olive oil was obtained at a local
retail store. The mouse liver samples were obtained from the Laboratory for Lipid Medicine &
Technology at Massachusetts General Hospital Charlestown, MA under Massachusetts General
Hospital or Harvard Committee of Animal Care approval number 2010N000038.
The fatty mice were FAT-2 transgenic mice, which ranged in levels of fat infiltration into
their liver based on a specialized high fat diet [34]. They have significantly higher total fat and
triglyceride (TG) fraction than the healthy control wild-type mice. The healthy mice were fed a
regular diet and have low omega-6 essential fatty acids. The FAT-2 mice are able to create
30
omega-6 polyunsaturated fatty acids (PUFA) endogenously, and therefore have high omega-6
PUFA. The high PUFA is a major contributor to fatty liver [34].
The imaging of all specimens, except the 14 mouse livers, was performed using a GE
Logiq E9 System with a 9 L-D probe. RF and B-mode images were collected from the machine
with differing assumed speeds of sound in the beamforming process ranging from 1400 to 1620
m/s (1400, 1420, 1480, 1500, 1540, 1580, 1620 m/s). The center frequency was held constant
at 9 MHz with a single focal zone within the depth of the tissue under examination.
(a) (b)Ultrasound Probe
U Sample
Reflector
Figure 4: Diagrammatic representation of the pulse-echo setup for determination of ex
vivo ground-truth speed of sound values. (b) Image of the experiment being performed
to determine ground truth speed of sound in an ex-vivo sheep liver using the
Ground-truth speed of sound measurements were obtained using the experimental setup
in Figure 4. A mechanical fixture firmly held the ultrasound probe at a known distance from a
surface. The specimen was placed between a fixed surface and the ultrasound probe. The
distance measurement was obtained using a linear magnetic encoder accurate to 0.001 inches
and a lead screw accurate to 0.050 inches. The travel time was calculated from the envelope of
31
the reflected signal from the interface between the tissue and an air gap between the sample and
the acrylic plate refelctor. The true speed of sound was then calculated by taking into account
the round-trip travel distance. If the error propagation is assumed to be correlated (additive)
with a unit time of I second, the total error in the ground truth would be plus or minus 0.051
m/s, which is negligible on this time scale.
2-axis Stage
ltrasou Probe
Liver Sample
Figure 5: Measurement setup on the Visualsonics Vevo 2100 ultrasound system (40 MHz
probe); the probe was in direct contact with the ex-vivo mice livers and held by an imaging
station that allows for the hands-free imaging of the sample.
A similar setup was utilized for the 14 mouse livers as seen in Figure 5; the key difference being
that a VisualSonics Vevo 2100 system was used. A 40 MHz probe (256 elements) was utilized
(40 MHz on transmit and 32 MHz on receive). The probe was in direct contact with the excised
livers and held by a 2-axis translational stage allowing hands-free imaging of the sample. The
data was collected with standard settings for that system. RF and B-mode data was collected
from the machine with varying speeds of sound ranging from 1400 to 1600 m/s in increments
of 10 m/s. The livers were approximately 4 mm in thickness.
32
Post data acquisition, the samples were frozen in dry ice and transported to the Pathology
Department at the Massachusetts General Hospital for fat quantification. A small sample of the
liver was extracted and the fat was quantified with a histology method similar to human biopsy
sample fat quantification. The histology grading was performed with a standard Oil Red 0 stain
and H&E stain. This protocol was used for the sheep liver samples and the mice liver samples.
The remainder of each samples was then frozen and transported to MGH Charlestown Campus
for biochemical analysis.
A biochemical analysis was performed on the livers as described by Kaliannan et al [33].
This method quantifies the total hepatic fat and triglyceride (TG) quantification purposes rather
than just a small section like the histology protocol. The tissue was mixed with a
choloroform:methanol solution (2:1 volume mixture). The sample was then vortexed and
incubated for 5 minutes. Additional chloroform:methanol solution was then added and
vigorously homogenized for 3 minutes. The sample was centrifuged at 2,000 x g for 10 minutes
and the supernatant layer containing lipids was collected. The additional chloroform was
allowed to evaporate and mass of the lipids was recorded. For TG quantification, lipids were
mixed methanol for resuspension. An aliquot was taken and vortexed with 1 M magnesium
chloride and then centrifuged for 5 minutes. A colorimetric, guinoneimine dye assay was then
performed to quantify the TG based upon lipoprotein lipase-mediated release of glycerol
(Sigma: Serum Triglyceride Determination Kit) [33].
4.2 Experimental Results
Sheep Liver: The true speed of sound in the sheep liver was 1540 m/s (+/- 0.051 m/s) as
obtained by the experimental setup in Figure 4. Experiments performed on the sheep liver
estimated the speed of sound to be: (1) 1540 m/s using brightness (2) 1540 m/s using the FFT
33
metric, and (3) 1540 m/s using the PCA metric as can be seen in Figure 6c; the predicted values
correspond to a normalized metric of 1.0. When compared, the predictions of each of the three
metrics were found to be in close agreement with the ground-truth measurement with the
caveat that the ground-truth measurement had an associated uncertainty of 0.051 m/s. The
sheep liver histology grading was < 5% fat as seen in the Oil Red 0 stain (Figure 6a) and H&E
j
-&-"Brightness"
1400 1450
4.
?~ *--A
-0-FF1
1500
-<>-PCA
1550 1600 1650
Speed of Sound (m/s)
Figure 6: Pathology results of the sheep liver show a lack of large fat vacuoles within the parenchyma as seen
with Oil Red 0 (a) and H&E stain (b). The ground truth measurements indicated that the speed of sound
within the liver was 1540 m/s (+/- 0.5%), and estimations using the brightness metric, FFT metric and PCA
metric were found to be 1540 m/s each (c).
stain (Figure 6b).
34
1
0.954-,
C
eu
4-,0)
- (c)-
(C.
V4
0.9 I
0.851350
Sheep Liver with a Simulated Fat Layer: The true speed of sound in the simulated fat layer
(olive oil) was 1400 m/s (+/- 0.5%) as measured using the experimental setup described in
detail above (Figure 3) and the true speed of sound in the sheep liver was measured to be 1540
m/s (+/-0.5%). The B-mode image of the experimental setup (Figure 7) shows the "fat" layer
on top, and the sheep liver submerged in oil. The speed of sound in the sheep liver, as predicted
by the brightness, FFT and PCA metrics independently, was found to be 1500 m/s (2.59% error)
as can be seen in Figure 7. The skewed prediction was likely due to the transmission through a
layer of olive oil with a relatively lower speed of sound. Upon accounting for this transmission,
the corrected sound speed was found to be 1548 m/s (0.52% error), thereby suggesting the
feasibility of applying the proposed algorithm in an in vivo setting.
-&-"Brightness" -O-FFT -<-PCA
1.00 -
0.95
0.90 -Error
- due to
layer of
0.80 - oE
0.80
0.701350 1400 1450 1500 1550 1600 1650
Speed of Sound (m/s)
Figure 7: Speed of sound prediction in a lean sheep liver (1500 m/s); the b-mode image in the lower left
corner depicts the olive oil layer (true speed of sound = 1400 m/s (+/- 0.5%)) sitting atop the liver parenchyma
(true speed of sound = 1540 m/s (+/- 0.5%)).
35
Bilayer Phantom: The ability of the proposed method to determine the speed of sound in
sequential layers of tissue was tested using two phantoms with different sound speeds in stacked
configuration. Phantom A (top layer) and Phantom B (bottom layer) were both certified by
CIRS to have true speeds of 1440 m/s and 1540 m/s respectively. The predicted speed of sound
in Phantom B using the FFT metric was 1500 m/s and the PCA metric was 1500 m/s. A mean
of both metrics resulted in a predicted speed of 1500 m/s. When compared to the ground-truth
value provided by CIRS, the prediction yielded a relative error of 2.59% as seen in Figure 8.
The skewed prediction was largely due to the transmission through Phantom A with a relatively
lower speed of sound. Upon accounting for this transmission, the corrected sound speed was
found to be 1551 m/s (0.71% error), thereby indicating the robustness and accuracy of the
proposed algorithm.
36
-G-FFT -0-PCA1.01
1.00
0.99
0.98
S0.97
"0 Error -
LE due to0.96 A
0.951350 1400 1450 1500 1550 1600 1650
Speed of Sound (m/s)
Figure 8: Predicted speed of sound in the lowest layer i.e. phantom B (1500 m/s), skewed by the transmission
through phantom A (true speed = 1440 m/s). The roughly 2.59% error was corrected by accounting for this
transmission and resulted in a renewed estimate of 1551 m/s (0.71% error) for phantom B.
Fatty and Control Mice Liver: Having tested the efficacy of the algorithm in predicting the true
speed of sound in tissue despite the presence of subcutaneous fat, the next step involved
demonstrating the effect of fat infiltration on the predicted speed of sound in liver. The metrics
were utilized to predict the speed of sound in a healthy mouse liver (ex vivo) and in a fatty
mouse liver - that was subjected to a high-fat diet for a prolonged duration of time as described
in the methods. Figure 9 shows the predicted speed of sound in the fatty and normal livers
along with pathological estimation which indicates the extent of fat infiltration in each of the
samples. As can be seen in the figure, the control sample was devoid of large fat vacuoles while
the fatty sample demonstrated a larger concentration of these vacuoles. The proposed method
37
readily captured this fact with all three metrics independently predicting a speed of 1500 m/s
for the control specimen and 1460 m/s for the fatty specimen. The absolute fat content in the
livers were found to be ~5% and ~ 15% (pathology based fat-extraction) for the control and
fatty specimens respectively.
38
-- FFT -- PCA
1
0.95
0.9
0.85
0.8
0.75
0.7
--- "Brightness"
0 1500Speed of Sound (m/s)
-O-FFT
1450 1500
Speed of Sound (m/s)
Figure 9: Predicted speed of sound in the control specimen (1500 m/s) along with pathology results that
demonstrate a lack of fat vacuoles (top); predicted speed of sound in the fatty specimen (1460 m/s) consistent
with the presence of large fat vacuoles in the liver parenchyma.
39
U,
C
.6'U
4j~
4-.a)
145
- - I
1350 1400 1550 1600 1650
-+- PCA
1
0.95
0.9
0.85
0.8
0.75
U,
C
.6'U
0)
0.7 '135
Steatosis
1400 1550 1600 1650
I I I I I
,6--Brightness
;0
Batch Study of Mice Livers: To further test the validity and applicability of the method, a set
of 14 mice liver were gathered; the fat content of each of the livers was estimated by the
Massachusetts General Hospital Pathology Laboratory, thereby providing a rought estimate of
the fat content within the livers. All three metrics were utilized to predict the speed of sound in
each of the ex vivo livers. Figure 10 demonstrates the Pearson's correlation between the absolute
fat content in the livers and the predicted speed of sound using the Mp (R 2 = 0.75) and MF
metrics (R 2 = 0.75). As can be seen in the figure, the speed of sound is inversely correlated
with the corresponding fat content; the batches are defined as follows: Batch 1 (< 5% Fat),
Batch 2 (5 - 10 % Fat), Batch 3 (10 - 15 % Fat), Batch 4 (15 - 20 % Fat) and Batch 5 (> 20%
Fat). Tables 3 and 4 summarize the predicted speeds of sounds obtained using the MF and Mp
metrics.
Table 3: Predicted speed of sound in mice livers using FFT metric
Batch # 1 2 3 4 5 Metric
Mean 1487 1483 1490 1465 1445 MF
Std. Dev 1.9% 1.4% 2.8% 1.4% 0.49% MF
Sample Size 4 4 2 2 2 MF
Table 4: Predicted speed of sound in mice livers using PCA metric
Batch # 1 2 3 4 5 Metric
Mean 1490 1480 1490 1465 1445 MP
Std. Dev 2.3% 1.2% 2.8% 1.4% 0.49% MP
Sample Size 4 4 2 2 2 MP
40
-1
While the standard deviations for the predicted speeds of sound may seem large (max 2.8%),
the observation can be explained as follows: firstly, the sample sizes were relatively small;
secondly, within the individual batches, the amount of fat in each of the associated livers varied,
which in turn, caused variations in the predicted speed of sound. Unfortunately, the brightness
metric (MB) did not yield results consistent with the FFT (MF) and PCA (Mp) metrics. This
can largely be attributed to the fact that the mice livers were compressed (albeit by
undocumented amounts) during the testing process and compression has been known to
artificially boost the echogenicity of scatterers and interfaces within tissue. This suggests the
use of image sharpness metrics over brightness metrics to predict the speed of sound in-vivo.
-+- PCA1500
-^. 1480E
o. 1460
)
1440
- I
- I
0
a
0CK
-0- FFT
Batch 1 (< 5% Fat)Batch 2 (5 -10 % Fat)
Batch 3 (10 -15 % Fat)Batch 4 (15 - 20 % Fat)
Batch 5 (> 20% Fat)
Increasing fat content
1 2 3
Fat Batch #
4 5 6
Figure 10: Correlation (RA2=0.75) between predicted speed of sound using sharpness metrics and absolute fat
content. Batch numbers are as follows: Batch 1 (< 5% Fat), Batch 2 (5 - 10 % Fat), Batch 3 (10 - 15 % Fat),
Batch 4 (15 - 20 % Fat) and Batch 5 (> 20% Fat)
41
CHAPTER
5ITERATIVE SPEED OF SOUND
ESTIMATION
5.1 Algorithm Description
In all the aforementioned experiments and simulations, the assumed beamforming speed
of sound was varied manually and the corresponding image quality metrics (MB and MS) were
computed to determine the true speed of sound in the tissue under examination. In reality,
however, the process can be fully-automated. All metrics (sharpness and brightness) are
computed with minimal cost and rely on the B-mode images only. An iterative bi-section
method can be employed to automate the process of finding the speed of sound that maximizes
the corresponding B-mode image quality. The following is an outline of the algorithm:
42
Pseudo-code for Iterative Speed of Sound Estimation
set speed 1 = 1400 m/s
set speed 2 = 1620 m/s
while convergence criteria is not met
acquire ultrasound images at speed 1 and speed 2
compute sharpness and brightness metrics for both speeds.
if metric (speed 1) > metric (speed 2)
update speed 1= speed 1
update speed2= average of speed 1 and speed 2
elseif metric (speed 1) < metric (speed 2)
update speed 1= average of speed 1 and speed 2
update speed 2 = speed 2
else
update speed
update speed
end if loop.
update convergence criteria
end while loop
1 = speed 1
2 = average of speed 1 and speed 2
The initial guesses of 1400 m/s and 1620 rn/s are driven by the fact that speed of sound in soft-
tissues is known to fall within that range and is unlikely to exceed those bounds. This approach,
43
while simplistic, is extremely robust given the fact that the metrics display a global maxima within
the search-space (beamformation speeds ranging from 1400 m/s to 1620 m/s). The convergence
criteria for the method was chosen to be:
mASOS = Ispeed 1 - speed 21 ; 0.1 - (7)
S
As can be seen above, the proposed framework allows for the determination of the true speed of
sound to a sufficient degree of precision.
5.2 Experimental Validation
The proposed algorithm was validated for the case of the CIRS Liver Phantom. The phantom is
manufactured by CIRS Inc., and has a calibrated speed of sound of 1540 m/s. The following figure
shows the convergence criteria (logarithmic scale) from (7) as a function of the iteration number:
44
-*-Error
2.5
2
1.5
I--
a) 0.5
0
0)
0
-0.5
-1
-1.50 2 4 6 8 10 12 14
Iteration Number
Figure 11: Error (logarithmic scale) associated with the iterative bi-section method as a function of the
number of iterations. As seen in the figure, the error decreases linearly in the logarithmic scale, therby
indicative of its underlying exponential decay.
As can be seen in the figure above, the algorithm converged in 12 steps and yielded an accuracy
of 0.1% with respect to ground truth measurements. Moreover, the error decreases linearly on a
logarithmic scale which implies an exponential decay of the error as a function of the number of
iterations. The following figure shows the initial guesses (speed 1 and speed 2 in the pseudo-code)
varying from iteration to iteration and eventually converging to the ground truth value of 1540
M/s.
45
-&-Speed 1 -<>-Speed 2
1650
1600
Ground Truth = 1540 m/sE 1550 -
0
0_0W) 1500CD
1450
1400 Z,1 2 3 4 5 6 7 8 9 10 11 12 13
Iteration Number
Figure 12: Variation of initial guesses within the bi-section method. As seen in the figure, speed 1 starts at
1400 m/s and speed 2 at 1620 m/s. After 12 iterations, both values converge to 1540 m/s (ground truth
indicated by dashed green line) to within tolerance (0.1 m/s).
As can be seen in Figure 11, the initial guess values for the speed of sound (speed 1 and speed 2
from the pseudo-code) start at 1400 m/s and 1620 m/s respectively. Both values are updated per
the algorithm described above to reach the ground truth of 1540 m/s in under 13 iterations. This
has important implications for the use of this technique in clinical settings. The proposed bisection
method relies on the fact that the image quality metrics can be computed with minimal
computational cost and hence, can be optimized. The proposed method does exactly that. By
treating the ultrasound data as images, the proposed method is able to quickly extract the associated
brightness and sharpness metrics, thereby making possible an iterative search over the allowable
range of beamforming speeds (1400 m/s to 1620 m/s).
46
CHAPTER
6TOWARDS A FRAMEWORK FOR
LONGITUDINAL SPEED OF SOUND MAPS
6.1 Motivation for Speed of Sound Maps
As was mentieond earlier, the spectrum of NAFLD is divided into two categories: (1)
simple steatosis (NAFL), which makes up 70-75% of cases, is defined by excess liver fat
without inflammation and (2) steatohepatitis (NASH), which makes up 25-30% of cases, is
defined by excess liver fat with inflammation and/or fibrosis [8, 9]. The focus of the work thus
far has been the estimation of the bulk speed of sound within the liver parenchyma and its
associated correlation with the percentage of fat within the liver. The speed of sound within the
liver parenchyma is not isotropic and uniform. This is especially true when considering patients
who suffer from NASH; the pathology of NASH is such that in addition to the diffuse
infiltration of fat within the liver (NAFL), the liver also exhibits localized fibrosis and
inflammation. These regions of the liver will exhibit higher longitudinal speeds of sounds.
Additionally, a bulk speed of sound would be insufficient in this case, because a liver that
exhibits fibrosis and/or inflammation might be mistakenly diagnosed as one that is healthy. The
reason being that while fibrosis/inflammation serves to increase the speed of sound within the
liver, diffuse fatty infiltration serves to lower the associated speed of sound. These competing
mechanisms might obscure the diagnostic predicitons. As such, there is a need for a technique
47
that allows for the non-invasive, quick and accurate estimation of the spatial distribution of the
longitudinal speed of sound within the liver.
6.2 S3 : A Spatial and Spectral Sharpness Measure [Vu et
al.]
The proposed solution stems from the work of [27]. The authors propose and validate a
metric called S3 which extracts the spatial distribution of a particular sharpness metric within
an image (described below). The spatial distribution is obtained by dividing the image into
sufficiently small, overlapping blocks, obtaining a measure of sharpness within each block, and
then stitching the results together [27]. This method is readily applied to the problem at hand,
because an error in the assumed beamforming speed exhibits itself as laterial blurring of the
associated ultrasound image. As such, the S3 metric can be used to create spatial sharpness maps
of ultrasound images over a range of assumed beamforming speeds and extract the map that
maximizes the associated local sharpness metrics.
6.3 Detailed Description of the Algorithm
The algorithm in [27] combines two different measures of sharpness, hereafter referrerd
to as the Spectral Measure of Sharpness (Si) and the Spatial Measure of Sharpness (S2), to
create a single sharpness map as follows [27]:
S -Y (8)
where 0 y 5 1 (y = 0.5 chosen for this study). The following convention will be used
throughout the descriptions: Let X denote the M1 x M2 input grayscale image. X is divided into
48
blocks of size m x m with d pixels of overlap between neighboring blocks. Let x denote a
block of X.
S,: Spectral Measure of Sharpness [271
Given x, let y(f, 0) denote the 2-D Discrete Fourier Transform (DFT), where f is the radial
frequency and 0 is the orientation. From y(y, 0), the magnitude of the spectrum across all
orientations can be computed as follows:
z(f)= Iy(f,0)1 (9)
The slope of the spectrum of x is then given by the following:
a* = arg min ||f -" - z(f)112 (10)
where the L 2 norm is taken over all radial frequencies. Finally, the spectral measure of sharpness
is given by
1S1(x) = 1 - (11)
1 + exp(T(a* - T 2 ))
where T1 = -3 and T2 = 2. The sigmoid function enables the metric to characterize regions
with a slope of 0 a 1 as sharp and regions with a '> 1 as blurred. In applying the metric,
a block size of m = 32 is chosen with an overlap of d = 24. A Hanning window is used to
suppress edge effects. The block size of 32 stems from the fact that it is sufficient to determine
the slope of the DFT co-efficients within any block.
Sq: Spatial Measure of Sharpness [271
Given a block x, the total variation within the block is defined as:
v(x) = |xi - xjl (12)i,)
49
where xi and xj are 8-neighbor pixels in x. Instead of obtaining the spatial measure of sharpness
directly from (11), the total variation is computed as the maximum of the total variation of
smaller blocks of x:
S2 (x) = max v( ) (13)Ex
where is a 2 x 2 block of x. The overall sharpness of the block can then be approximated by
the maximum sharpness of its sub-blocks. For the computation of the spatial sharpness measure,
blocks of size m = 8 with no overlap are utilized.
6.4 Preliminary Results (Ex Vivo)
The algorithm proposed by the authors of [27] was tested on the CIRS Liver Phantom.
As was mentioned earlier, the CIRS Liver Phantom is manufactured to have a constant
longitudinal speed of sound (1540 m/s). In order to construct a longitudinal speed of sound map,
B-mode images were gathered at the following beamforming speeds: 1400 m/s, 1420 m/s, 1460
m/s, 1500 m/s, 1540 m/s, 1580 m/s and 1620 m/s. The GE Logiq E9 in Figure 4 and the
associated ground-truth measurement setup was utilized during data collection. S3 maps were
generated for the ultrasound images associated with each of the beamforming speeds and locally
optimized to determine the best prediction of the spatial variation of the speed of sound within
the phantom. The following figure shows the B-mode image of the CIRS Liver Phantom
obtained at a beamforming speed of 1540 m/s using the setup in Figure 4.
50
Figure 13: B-mode image of the CIRS Liver Phantom acquired using the GE Logiq E9 at a beamforming
speed of 1540 m/s.
As can be seen in the image above, the phantom is designed to have point scatterers within its
volume and a constant longitudinal speed of sound (1540 m/s). The following image shows the
optimized speed of sound map obtained by optimizing the S 3 sharpness metric over the above
mentioned range of beamforming speeds.
51
1620
1600
50 A 1580
1560
100 ~ 1540
1520
150 1500
1480
200 1460
1440
250 r -1420
140050 100 150 200 250 300 350
Figure 14: Speed of sound map for the CIRS Liver Phantom obtained by optimizing the S3 sharpness map
over a range of beamforming speeds.
The following figure shows the speed of sound map obtained after being convolved with a low-
pass, Gaussian filter (o = 2):
52
1620
4 1600
50 71580
1560
100 1540
1520
150 1500
1480
200 1460
1440
250 1420
50 100 150 200 250 300 350
Figure 15: Speed of sound map (blurred with a Gaussian filter (a = 2)) for the CIRS Liver Phantom
obtained by optimizing the S3 sharpness map over a range of beamforming speeds.
As can be seen in the figure above, the vast majority of the phantom has a speed of roughly 1540
m/s with localized variations ranging from -120 m/s to + 80 m/s. A spatial average of the predicted
map yields a bulk value of 1514 m/s which represents a deviation of 1.68% from the ground-truth
value of 1540 m/s. At the current stage, ground-truth values for the speed of sound maps do not
exist, but the goal is to use these maps in conjunction with corresponding elastography maps
(which are available on commercial scanners) to generate a composite framework that will allow
for the non-invasive diagnosis of NAFLD.
53
1- 1
CHAPTER
7CONCLUSIONS AND FUTURE WORK
In the context of NAFLD, which has an estimated prevalence of 30% in the US population,
ultrasound imaging remains the likely choice for an imaging biomarker. Despite this fact, the use
of conventional B-mode ultrasound imaging is plagued by several shortcomings. Conventional
ultrasound imaging is only accurate when steatosis is moderate-to-severe (i.e., defined as
histologic degree >33%) [28, 29] and is insensitive to lesser grades of hepatic steatosis [28, 29, 30,
31, 32]. In addition, ultrasound evaluation is subjective, with significant inter-observer variability,
rendering it ineffective for the evaluation of early therapeutic responses in patients undergoing
NAFLD treatment. The development and validation of a robust and accurate ultrasound method to
accurately quantify steatosis at its earliest stages would be of great clinical importance. In this
study, we have aimed to use several markers embedded in conventional ultrasound images to
predict the true speed of sound within the liver parenchyma and use this ability to quantify the
degree of fatty infiltration.
The proposed method involved the iterative estimation of the true speed of sound in the liver
parenchyma using two image quality metrics: image brightness around the expected focal zone
and image sharpness around the intended focal zone (FFT and PCA metrics as explained above).
The method's capability to extract the true speed of sound was tested in tissue-mimicking
phantoms, ex-vivo sheep and mice livers; comparison to ground truth measurements indicated that
the proposed method was able to predict the speed of sound to within 0.5% error. The method's
54
potential to non-invasively and quantitatively diagnose NAFLD was tested by using it to predict
the speed of sound in the livers of mice of two kinds: 1) healthy and 2) those subjected to a high
fat diet for a prolonged duration of time. Not only was the method able to accurately capture the
true speeds of sounds in the livers (less than 0.5% error), but it was also able to capture the
downshift in the speed of sound of the liver parenchyma due to fatty infiltration; lastly, when tested
on mice livers with fat percentages ranging from 3% to 25%, the algorithm was able to capture the
proposed inverse correlation between fat content and the estimated speed of sound. Another
strength of the proposed method lies in the fact that it can account for the transmission of
ultrasound waves through layers of subcutaneous fat; this is a widespread limitation in estimating
the bulk speed of sound in vivo. The proposed method utilizes a straightforward approach that
segments the underlying RF data to predict the speed of sound in successively deeper layers of
tissue. The validity and robustness of this approach was tested by predicting the speed of sound in
healthy sheep liver submerged in homogenous oil (to simulate a layer of fat). The speeds of sound
in both layers were accurately predicted to within 1 %.
Even though the proposed method was validated in simulations and static ex vivo
experiments with reasonably accurate results, a considerable amount of work remains to be done
before the proposed algorithm can be utilized in a clinical setting. The presented results were
representative of static configurations i.e. the ultrasound transducer and samples were not in
motion. This will not be in the case in human subjects due to the consistent motions associated
with breathing. While the subjects can be asked to hold their breaths, this can be uncomfortable
for prolonged periods of time. As such, a dynamic motion compensation technique has to be
developed to negate the associated detrimental effects of unconstrained motion. The scope of this
work has been limited to simple steatosis i.e. fatty infiltration into the liver without the presence
55
of inflammation or fibrosis. Given the scope, the proposed method was designed to provide a bulk
speed of sound within the liver parenchyma. A logical extension would be to extend the S3 metric
to provide accurate spatial distributions of the longitudinal speed of sound within the parenchyma,
eventually leading to a full-volume reconstruction of the liver that can be used in conjunction with
elastography maps to stage and diagnose the entire spectrum of NAFLD.
While much work remains to be done, the initial results are promising and demonstrate the
viability of using quantitative ultrasound as a non-invasive technique to diagnose NAFLD.
56
REFERENCES
[1] Williams CD, Stengel J, Asike MI, Torres DM, Shaw J, Contreras M, Landt CL, Harrison SA.
Prevalence of nonalcoholic fatty liver disease and nonalcoholic steatohepatitis among a largely
middle-aged population utilizing ultrasound and liver biopsy: a prospective study,
Gastroenterology, 2011, vol. 140, pp. 12 4 - 13 1.
[2] Chalasani N, Younossi Z, Lavine JE, Diehl AM, Brunt EM, Cusi K, Charlton M, Sanyal AJ.
The diagnosis and management of non-alcoholic fatty liver disease: practice Guideline by the
American Association for the Study of Liver Diseases, American College of Gastroenterology,
and the American Gastroenterological Association. Hepatology, 2012a, pp. 2005-2023.
[3] Rinella ME. Nonalcoholic fatty liver disease: a systematic review, JAMA American Medical
Association, 2015, vol. 313, pp 2263-2273.
[4] Idowu MO, Chhatrala R, Siddiqui MB, Driscoll C, Stravitz RT, Sanyal AJ, Bhati C, Sargeant
C, Luketic VA, Sterling RK, Contos M, Matherly S, Puri P, Siddiqui MS. De novo hepatic steatosis
drives atherogenic risk in liver transplantation recipients, Liver Transplantations, 2015, vol. 21,
pp. 1395-1402.
[5] Athyros VG, Tziomalos K, Katsiki N, Doumas M, Karagiannis A, Mikhailidis DP.
Cardiovascular risk across the histological spectrum and the clinical manifestations of non-
alcoholic fatty liver disease: An update, World Journal of Gastroenterology, 2015, vol. 21, pp.
6820-6834.
[6] Schwimmer JB, Pardee PE, Lavine JE, Blumkin AK, Cook S. Cardiovascular Risk Factors and
the Metabolic Syndrome in Pediatric Nonalcoholic Fatty Liver Disease, American Heart
Association, Inc, 2008, vol.118 (3), pp. 277-83.
57
[7] Kelishadi R, Cook SR, Adibi A, Faghihimani Z, Ghatrehsamani S, Beihaghi A, et al.
Association of the components of the metabolic syndrome with non- alcoholic fatty liver disease
among normal-weight, overweight and obese children and adolescents, Diabetology and Metabolic
Syndrome, 2009, vol. 1(1).
[8] Chalasani N, Younossi Z, Lavine JE, Diehl AM, Brunt EM, Cusi K, Charlton M, Sanyal AJ,
The diagnosis and management of non-alcoholic fatty liver disease: practice guideline by the
American Gastroenterological Association, American Association for the Study of Liver Diseases,
and American College of Gastroenterology, Gastroenterology, 2012b, pp. 1592-1609.
[9] McPherson S, Hardy T, Henderson E, Burt AD, Day CP, Anstee QM. Evidence of NAFLD
progression from steatosis to fibrosing-steatohepatitis using paired biopsies: implications for
prognosis and clinical management, Journal of Hepatology, 2015, vol. 62, pp. 1148-1155.
[10] Vajro P, Lenta S, Socha P, Dhawan A, McKiernan P, Baumann U, Durmaz 0, Lacaille F,
McLin V, Nobili V. Diagnosis of nonalcoholic fatty liver disease in children and adolescents:
position paper of the ESPGHAN Hepatology Committee, Journal of Pediatric Gastroenterology
and Nutrition, 2012, vol. 54, pp. 700-713.
[11] Yan E, Durazo F, Tong M, Hong K. Nonalcoholic fatty liver disease: pathogenesis,
identification, progression, and management, Nutrition Reviews, 2007, vol. 65, pp. 376-384.
[12] Kumagai H, Yokoyama K, Katsuyama K, Hara S, Yamamoto H, Yamagata T, Taniguchi N,
Hirota and Itoh K. A new method for measuring the speed of sound in rat liver ex vivo using an
ultrasound system: correlation of sound speed with fat deposition, Ultrasound in Medicine and
Biology, 2014, vol. 40, pp. 2499-2507.
[13] Chalasani N, Younossi Z, Lavine JE, Diehl AM, Brunt EM, Cusi K, Charlton M, Sanyal AJ.
The diagnosis and management of non-alcoholic fatty liver disease: practice Guideline by the
58
American Association for the Study of Liver Diseases, American College of Gastroenterology,
and the American Gastroenterological Association, Hepatology, 2012a, pp. 2005-2023.
[14] Ghoshal G, Lavarello RJ, Kemmerer JP, Miller RJ, and Oelze ML. Ex vivo study of
quantitative ultrasound parameters in fatty rabbit livers, Ultrasound in Medicine and Biology,
2012, vol. 38, pp. 2238-2248.
[15] Gyongy, M, Kollar, S. Variation of ultrasound image lateral spectrum with assumed speed of
sound and true scatterer density, Ultrasonics, 2015, vol. 56, pp. 370-380.
[16] Qu, X, Azuma T, Liang JT, Nakajima Y. Average sound speed estimation using speckle
analysis of ultrasound data, International Journal of Computer Assisted Radiology and Surgery,
2012, vol. 7, pp. 891-899.
[17] Teitelbaum JE, Rodriguez RJ, Ashmeade TL, Yaniv I, Osuntokun BO, Hudome S, Fanaroff
A. Quantitative ultrasound in the evaluation of bone status in premature and full-term infants,
Journal of Clinical Densitometry, 2006, vol. 9(3), pp. 358-362.
[18] Boyd NF, Martin LJ, Bronskill M, Yaffe MJ, Duric N, Minkin S. Breast tissue composition
and susceptibility to breast cancer, Journal of the National Cancer Institute, 2010, vol. 102(16),
pp. 1224-1237.
[19] Chalasani N, Younossi Z, Lavine JE, Diehl AM, Brunt EM, Cusi K, Charlton M, Sanyal AJ,
The diagnosis and management of non-alcoholic fatty liver disease: practice guideline by the
American Gastroenterological Association, American Association for the Study of Liver Diseases,
and American College of Gastroenterology, Gastroenterology, 2012b, pp. 1592-1609.
[20] Anderson ME, McKeag MS, and Trahey GE. The impact of sound speed errors on medical
ultrasound imaging, Journal of the Acoustical Society of America, 2000, vol. 106, pp. 3540-3548.
59
[21] Weijers G, Wanten G, Thijssen JM, van der Graaf M, and de Korte CL. Quantitative
ultrasound for staging of hepatic steatosis in patients on home parenteral nutrition validated with
magnetic resonance spectroscopy: a feasibility study, Ultrasound in Medicine and Biology, 2016,
vol. 42, pp. 637-644.
[22] Wagner RF, Smith SW, Sandrik JM, and Lopez H. Statistics of speckle in ultrasound b-scans,
IEEE Transactions on Sonics and Ultrasonics, 2000, vol. 30, pp. 1 56 -16 3 .
[23] Nock L, Trahey GE, Smith SW. Phase aberration correction in medical ultrasound using
speckle brightness as a quality factor, The Journal of the Acoustical Society of America 1989,
vol. 85, pp. 1819-1833.
[24] De K, and Masilamani V., Image sharpness measure for blurred images in frequency domain,
Procedia Engineering, 2013, vol. 64, pp. 149-158.
[25] Zhu X, and Milanfar P, A no-reference sharpness metric sensitive to blur and noise, 2009
International Workshop on Quality of Multimedia Experience, 2009, vol. 2 (3), pp 64-69.
[26] McGough RJ. Focus: Fast Object Oriented C++ Ultrasound Simulator. Michigan State. 2016.
Matlab Software.
[27] Vu CT, and Chandler DM. S 3 : A Spectral and Spatial Sharpness Measure, IEEE Computer
Society: 2009 First International Conference on Advances in Multimedia, pp. 37-43.
[28] Lee SS, Park SH, Kim HJ, Kim SY, Kim MY, Kim DY, Suh D-J, Kim KM, Bae MH, Lee
JY, Lee SG, Yu ES. Non-invasive assessment of hepatic steatosis: Prospective comparison of the
accuracy of imaging examinations, Journal of Hepatology, 2010, vol. 52, pp. 579-585.
[29] van Werven JR, Marsman HA, Nederveen AJ, Smits NJ, Kate ten FJ, van Gulik TM, Stoker
J. Assessment of hepatic steatosis in patients undergoing liver resection: comparison of US, CT,
60
TI-weighted dual-echo MR imaging, and point-resolved 1 H MR spectroscopy, Radiology, 2010,
vol. 256, pp. 159-168.
[30] de Moura Almeida A, Cotrim HP, Barbosa DBV, de Athayde LGM, Santos AS, Bitencourt
AGV, de Freitas LAR, Rios A, Alves E. Fatty liver disease in severe obese patients: diagnostic
value of abdominal ultrasound, World Journal of Gastroenterology, 2008, vol.14, pp. 1415-
1418.
[31] Kinner S, Reeder SB, Yokoo T. Quantitative Imaging Biomarkers of NAFLD, Digestive
Diseases and Sciences Springer US, 2016, vol. 61, pp. 1337-1347.
[32] Palmentieri B, de Sio 1, La Mura V, Masarone M, Vecchione R, Bruno S, Torella R, Persico
M. The role of bright liver echo pattern on ultrasound B-mode examination in the diagnosis of
liver steatosis, Digestive and Liver Disease, 2006, vol. 38:485-489.
[33] Kaliannan K, Hamarneh SR, Economopoulos KP, Alam SN, Moaven 0, Patel P, Malo NS,
Ray M, Abtahi SM, Muhammad N, Raychowdhury A, Teshager A, Mohamed MMR, Moss AK,
Ahmed R, Hakimian S, Narisawa S, Millan JL, Hohmann E, Warren HS, Bhan AK, Malo MS,
Hodin RA. Intestinal alkaline phosphatase prevents metabolic syndrome in mice. PNAS, 2013,
vol. 110(17) pp. 7003-7008.
[34] Pai VJ, Wang B, Li X, Wu L, Kang JX. Transgenic Mice Convert Carbohydrates to Essential
Fatty Acids. PLoS ONE, 2014, vol. 9(5): e97637.
61