Professional Documents
Culture Documents
SPIEDigitalLibrary.org/conference-proceedings-of-spie
Open-source full-waveform
ultrasound computed tomography
based on the angular spectrum
method using linear arrays
Rehman Ali
Event: SPIE Medical Imaging, 2022, San Diego, California, United States
ABSTRACT
We present full-waveform ultrasound computed tomography (USCT) for sound speed reconstruction based on
the angular spectrum method using linear transducer arrays. We first present a transmission scenario in which
plane-waves are emitted by a transmitting array and received by an array on the opposite side of the object
of interest. These arrays are rotated around the object of interest to interrogate the medium from di↵erent
view angles. Waveform inversion reconstruction is demonstrated on a numerical breast phantom, in which
sound speed is varied from 1486 to 1584 m/s. This example is used to isolate and examine the impact of each
view angles and frequency used in the reconstruction process. We also examine cycle-skipping artifacts as well
as optimization schemes that can be used to overcome them. The goal of this work is to provide an open-
source example and implementation of the waveform inversion reconstruction algorithm on Github: https://
github.com/rehmanali1994/FullWaveformInversionUSCT (DOI: 10.5281/zenodo.4774394). Next, we extend
the waveform inversion framework to perform sound speed tomography for pulse-echo ultrasound imaging with
a single linear array that transmits pulsed waves and receives signals backscattered from the medium. We first
demonstrate that B-mode image reconstructions can be achieved using the angular spectrum method; then, we
derive an optimization framework for estimating the sound speed in the medium by optimizing B-mode images
with respect to slowness, via the angular spectrum method. We demonstrate an initial proof of concept with
point targets in a homogeneous medium to demonstrate the fundamental principles of this new technique.
Keywords: Nonlinear inverse problems, waveform inversion, ultrasound, tomography, angular spectrum method
1. INTRODUCTION
Ultrasound computed tomography (USCT) is an imaging technique that uses the transmission of ultrasound
through tissue to reconstruct high-resolution images of tissue properties such as sound speed and attenuation.
The primary application of ultrasound tomography is in breast imaging1–6 where sound speed and attenuation
are key biomarkers for the identification of cancer in the breast. This report specifically investigates USCT for
sound speed estimation in heterogeneous sound-speed media.
Techniques for sound speed estimation strategies generally fall into two broad categories: bent-ray meth-
ods,7, 8 and waveform inversion.9, 10 Waveform inversion appears to be the most accurate technique for USCT
reconstruction and greatly outperforms bent-ray reconstructions in most applications. Despite the substantially
lower computation time of bent-ray methods, their key drawback is the inability to account for di↵raction,
which inevitably leads to poorer resolution in the reconstructed image. Waveform inversion is more accurate
because it uses iterative simulations of di↵raction to estimate the sound speed profile responsible for producing
the measured ultrasound signals.
This report specifically investigates waveform inversion for sound speed estimation in heterogeneous sound-
speed media using USCT. The primary goal of this work is to demonstrate and examine each step of waveform
inversion using the angular spectrum method.11, 12 We aim to examine the parameters that impact sound speed
reconstruction and consider challenges posed by cycle skipping on a robust reconstruction algorithm. Although
Further author information: (Send correspondence to Rehman Ali)
Rehman Ali: E-mail: rali8@stanford.edu
Medical Imaging 2022: Ultrasonic Imaging and Tomography, edited by Nick Bottenus, Nicole V. Ruiter,
Proc. of SPIE Vol. 12038, 120380R · © 2022 SPIE · 1605-7422 · doi: 10.1117/12.2601257
2. BACKGROUND
2.1 Fourier Split-Step Angular Spectrum Method
(a) (b)
0.5 ! ", $, %
0.6
! ", $, %
0.7
Bulk Propagation via
0.8
Δ$ ' ", $ Angular Spectrum Δ$ '̅ $ *!" +# , $, %
0.9
Method
0.5
0.7
0.9
' ", $ − '̅ $ *$" ", $, %
Correction 1
! ", $ + Δ$, %
1.2
1.1
1.2
1.3 1.3
1.4
1.4
Figure 1. Fourier Split-Step Angular Spectrum Method for Simulating Ultrasonic Pressure Fields. (a) Propagation from
1.5
1 2 3 4 5 6 7 8 9 10
depth z to z + z in a heterogeneous slowness medium. (b) Decomposition of propagation into bulk propagation via the
1.5
1 2 3 4 5 6 7 8 9 10
angular spectrum method followed by a correcting phase screen (see equations (1), (2), and (3)).
Ultrasonic wave-fields p(x, z, f ) as a function of location (x, z) and frequency f can be propagated from a
depth of z to z + z using the Fourier split-step form of the angular spectrum method
⋮
#⃑ ≝ #!,# ≝ # %, '
⋮
⋮ Receiver Array
Figure 2. Plane-wave Transmission Simulation Geometry and Mathematical Notation. (a) This schematic shows a vec-
torization of the wavefields at each layer in the medium. The distance between the transmitting and receiving arrays is
denoted as zsep . The signals measured at the receiving array are used in full-waveform inversion. (b) The transmitting
and receiving arrays are rotated around the medium. The received signals from each view angle are used in the complete
full-waveform inversion.
where ~s is the vectorization of slowness values over the grid, F is the discrete Fourier transform matrix,
DAS,k (~s, f ) is the angular spectrum method as a diagonal matrix acting in the kx domain, and DP S,k (~s, f )
is the phase screen implemented as a diagonal matrix acting over the x domain. These recursive relations de-
scribe the downward propagation of the ultrasonic wave-field from the transmitting array to the receiving array.
The entire Fourier split-step angular spectrum method for all layers in the medium can be summarized using
the following matrix formulation:
A(~s, f )~y (~s, f ) = ~b(~s, f ), (6)
where 2 3
I 0 0 ... 0 0
6 D1 (~s, f ) I 0 ... 0 0 7
6 7
6 0 D2 (~s, f ) I ... 0 0 7
6 7
A(~s, f ) = 6 .. .. .. .. .. ..7, (7)
6 . . . . . .7
6 7
4 0 0 0 ... I 0 5
0 0 0 ... DN (~s, f ) I
2 3 2 3
p~0 (f ) p~0 (f )
6 p~1 (~s, f ) 7 6 ~0 7
6 7 6 7
6 p~2 (~s, f ) 7 6 ~0 7
6 7 ~b(~s, f ) = 6 7
~y (~s, f ) = 6 .. 7, 6 .. 7. (8)
6 . 7 6 . 7
6 7 6 7
4 p~N 1 (~s, f ) 5 4 ~0 5
p~N (~s, f ) ~0
Note that the signal measured across the receiving arrays is p~N (~s, f ) = K~y (~s, f ) where K = [0, . . . , 0, I]; therefore,
the received signals can be modeled as
p~N (~s, f ) = K[A(~s, f )] 1~b(~s, f ), (9)
where the view angles around the imaged object are indexed i = 1, ..., Nviews , p~obs,i (f ) represents the observed
receive signals from each view angle, and p~N,i (~s, f ) represents the modeled receive signals based on the current
estimate of the slowness ~s.
⋮ ⋮
⋮ ⋮ ( %, ' − Δ', +
(b)
'⃑! (%)
⋮ ⋮ '⃑" (#,
⃑ %)
∗
#⃑ ≝ #!,# ≝ # %, ' "⃑ #,
⃑ % ≝ '⃑# (#,
⃑ %) Δ' #̅ ' -$% .' , ', +
⋮
⋮ ⋮ '⃑$ (#,
⃑ %)
0.5
⋮
0.6
Thin-Film
⋮ 0.7
∗
2⃑! (%)
Phase Screen 0.8
Δ# %, ' -(% %, ', +
Correction
0.9
⋮ ( %, ', +
1
2⃑" (#,
⃑ %)
4⃑ #,
⃑ % ≝ 2⃑ (#,
⋮ 1.1
# ⃑ %)
1.2
1.3
⋮
⋮
1.4
⃑
2$ (#,
⃑ %) ⋮ 1.5
1 2 3 4 5 6 7 8 9 10
2⃑$ #,
⃑ % = '⃑$ #,
⃑ % − '⃑()* (%) ' +, , = z%&' , % ≝ '⃑$ (#,
⃑ %)
Receiver Array
Figure 3. Upward Propagation of Adjoint Wave-Field. (a) This schematic shows a vectorization of the upward propagation
of the error in the receive signals. (b) The angular spectrum method used in the upward propagation is mathematically
adjoint (reversed in order and complex conjugated) to the downward propagation.
The goal of this section is to show a complete derivation of each step in the waveform inversion process. We
first consider the partial objective function Ji (~s, f ) = 12 k~
pN,i (~s, f ) p~obs,i (f )k22 at view i and frequency f . The
gradient of this partial objective function with respect to the slowness sm,n at pixels (m, n) is
7 8 9
⇢
@Ji (~s, f ) H @~
pN,i (~s, f )
= Re (~pN,i (~s, f ) p~obs,i (f )) , (11)
@sm,n @sm,n
where
@~
pN,i (~s, f ) @ ⇣ ⌘ @[A(~s, f )] 1 ~
= K[A(~s, f )] 1~b(~s, f ) = K b(~s, f )
@sm,n @sm,n @sm,n
@A(~s, f ) 1 @A(~
s, f )
= K[A(~s, f )] 1 [A(~s, f )] 1~b(~s, f ) = K[A(~s, f )] ~y (~s, f ).
@sm,n @sm,n
where
~v (~s, f ) = [AH (~s, f )] 1
K T (~
pN,i (~s, f ) p~obs,i (f )) . (13)
While ~y (~s, f ) is the wave-field projected forward from the transmitting array, ~v (~s, f ) is a back-projection of the
error between the modeled and measured signals at the receiver array upwards into the medium (see Figure 3(a)).
Equation (13) shows that the error p~N,i (~s, f ) p~obs,i (f ) is injected at the receiver array using the K T operator
(K represents sampling at the receiver array, so K T is an injection operator). Then, the resulting injected error
is propagated upwards using an adjoint form of the angular spectrum method via [AH (~s, f )] 1 . To see this,
equation (13) can be re-written as
According to the Rytov approximation,25, 26 the phase screen can absorb all perturbations in the slowness so
that
@Dk (~s, f ) @DP S,k (~s, f ) ⇤
= F DAS,k (~s, f )F. (22)
@sm,n @sm,n
The mathematical expression for the matrix form of the phase screen DP S,k (~s, f ) is
@DP S,k (~
s,f )
Based on equation (24), @sm,n is a diagonal matrix with at most 1 non-zero element, which means that
@Dk (~
s,f )
@sm,n is a diagonal matrix with at most 1 non-zero element. Ultimately, this means that @A(~ s,f )
@sm,n is a matrix
with exactly 1 non-zero element corresponding to the voxel for sm,n . This implies that equation (12) represents
a point-wise multiplication of the downward-going transmitted wave-field ~y (~s, f ) and the upward-going back-
projected wave-field ~v (~s, f ) of the error in the received
2 signals.
3 2 Following
3 2 this line3 of reasoning, if we denote
a1 b1 a 1 b1
as the point-wise multiplication of two vectors (i.e., 4 a2 5 4 b2 5 = 4 a2 b2 5), then
a3 b3 a 3 b3
However, because ~b(f ) is zero wherever a slowness pixel exists, the gradient of the partial objective function is
Transmitting Array
! ", $ = 0, ' ≝ !⃑! (')
! " ∇"⃑$ &⃑ ≈ ()2+",-)/ 0 &,
⃑" #$ (2
⃑ &,
⃑ " ∘ ∇"⃑$ &⃑ ) ! ", $ = Δ$, ' ≝ !⃑" (-,
⃑ ')
! ", $ = 2Δ$, ' ≝ !⃑# (-,
⃑ ')
Figure 4. Visual Interpretation of L(f )r~s J(~s) Representing the Linearized Forward Projection of Gradient r~s J(~s) on the
Received Ultrasound Signal.
A fast and robust conjugate gradient (CG) algorithm for the minimization problem (10) requires a fast and
closed-form solution for the step length of r~s J(~s). This requires a linearization Li (f ) around the current slowness
@~
pN,i (~s, f ) 1 @A(~
s, f )
Li (f ) = = K[A(~s, f )] ~y (~s, f ),
@~s @~s
and
1
Li (f )r~s J(~s) = (j2⇡f z)K[A(~s, f )] (~y (~s, f ) r~s J(~s)) .
Figure 4 shows a visual interpretation of equation (2.2): Li (f )r~s J(~s) represents a linearized prediction of the
the e↵ect that perturbing ~s by r~s J(~s) would have on receive signals p~N,i (~s, f ).
Composing all the key components of the gradient calculation and the linearized forward projection of the
gradient, the CG algorithm for the spatial reconstruction of slowness is described in Algorithm 1. Note that the
CG algorithm is intentionally dampened by the parameter to overcome issues related to cycle-skipping, which
we demonstrate and discuss in more details later on.
Algorithm 1 Conjugate Gradient (CG) Algorithm for Waveform Inversion Reconstruction of Slowness
s0 to a uniform profile: ~s0 = sinit~1
1: Initialize slowness ~ 1
. sinit = 1540 m
s
~
2: Initialize search direction d0 = r~s J(~
s0 )
3: Initialize step length parameter 1 . << 1 to dampen CG for cycle-skipping problem
4: for k = 0, ..., N 1 do . N Iterations of CG
d~T
k r~
s J(~
sk )
5: Compute Step Size: ↵k = P PNviews ~ 2
.
f i=1 kLi (f )dk k2
ptx(i) (x, z + z, f ) = DP S (x, z, f )Fkx1 x {DAS (kx , z, f )Fx kx {p(x, z, f )}}, (29)
prx(i) (x, z + z, f ) = DP⇤ S (x, z, f )Fkx1 x {DAS
⇤
(kx , z, f )Fx kx {p(x, z, f )}}. (30)
If we vectorize Ii (x, z), ptx(i) (x, z, f ), and prx(i) (x, z, f ) as I~i (~s), ~ytx(i) (~s, f ), and ~yrx(i) (~s, f ), the image formation
process can be expressed as Z 1
~
Ii (~s) = ~y ⇤ (~s, f ) ~yrx(i) (~s, f )df, (31)
tx(i)
0
@ I~i (~s)
I~i (~s + ~s) ⇡ I~i (~s) + ~s, (34)
@~s
Z ⇤
@ I~i (~s) 1 @~ytx(i) (~s, f ) ⇤ @~yrx(i) (~s, f )
= ~yrx(i) (~s, f ) + ~ytx(i) (~s, f ) df, (35)
@~s 0 @~s @~s
where
@~ytx(i) (~s, f ) 1
@[A(~s, f )] ~btx(i) (~s, f )
=
@~s @~s
1 @A(~
s, f )
= [A(~s, f )] 1~btx(i) (~s, f )
[A(~s, f )] (36)
@~s
@A(~s, f )
= [A(~s, f )] 1 ~ytx(i) (~s, f ),
@~s
@~yrx(i) (~s, f ) @[A⇤ (~s, f )] 1 ~
= brx(i) (~s, f )
@~s @~s
@A⇤ (~s, f ) ⇤
= [A⇤ (~s, f )] 1 [A (~s, f )] 1~brx(i) (~s, f ) (37)
@~s
@A⇤ (~s, f )
= [A⇤ (~s, f )] 1 ~yrx(i) (~s, f ).
@~s
⇣ ~ ⌘H
The following equations show how to apply @ I@~ i (~
s
s)
to an image I(~ ~ s):
!H Z ⇤
!T
@ I~i (~s) ~ s) =
1
@~ytx(i) (~s, f ) ⇤
@~yrx(i) (~s, f )
~ s)
I(~ ~yrx(i) (~s, f ) + ~ytx(i) (~s, f ) df I(~
@~s 0 @~s @~s
(38)
Z 1 ✓ ◆T ⇣ ⌘ ✓ ◆H ⇣ ⌘
@~ytx(i) (~s, f ) ~ s) ~y ⇤ (~s, f ) + @~yrx(i) (~s, f ) ~ s) ~ytx(i) (~s, f ) df,
= I(~ rx(i) I(~
0 @~s @~s
where ✓ ◆T ✓ ◆T
@~ytx(i) (~s, f ) T @A(~s, f )
= ~ytx(i) (~s, f ) [A(~s, f )] 1 , (39)
@~s @~s
✓ ◆H ✓ ◆T
@~yrx(i) (~s, f ) H @A(~s, f )
= ~yrx(i) (~s, f ) [A(~s, f )] 1 . (40)
@~s @~s
Given a slowness perturbation ~s, the image perturbation I~i (~s) with respect to slowness may be computed as:
! Z 1 @~y ⇤ (~s, f ) ! ✓ ◆
@ ~i (~s)
I tx(i) @~yrx(i) (~s, f )
~
Ii (~s) = ~s = ~s ⇤
~yrx(i) (~s, f ) + ~ytx(i) (~s, f ) ~s df, (41)
@~s 0 @~s @~s
where ✓ ◆
@~ytx(i) (~s, f ) 1 @A(~s, f )
~s = [A(~s, f )] ~s ~ytx(i) (~s, f ), (42)
@~s @~s
One approach to WEMVA for pulse-echo ultrasound is to maximize the B-mode image brightness. In
pulse-echo ultrasound, the images Ii (x, z) from each transmit event are coherently compounded, amplitude
PNtx
detected, and log-compressed into a B-mode ultrasound image: IBM ode (x, z) = 20 log10 | i=1 Ii (x, z)|. Maxi-
mizing IBM ode (x, z) is conceptually equivalent to maximizing the power of the coherent sum of the images (as
PNtx 2 PNtx ~ 2
is often done in seismic imaging): Ipower (x, z) = 12 i=1 Ii (x, z) or Jpower (~s) = 12 i=1 Ii (~
s) . In this case,
2
the gradient of the power of the coherent sum of migrated images with respect to the slowness is
8 ! 0N 19
Ntx
X < @ I~ (~s) H X tx =
i
r~s Jpower (~s) = Re @ I~j (~s)A . (44)
: @~s ;
i=1 j=1
However, the objective function here is not a least-squares minimization problem. In fact, the objective function
is the maximization of a squared norm. Because the maximization of image brightness is not a least-squares
problem, there is no forward model to linearize that would enable the closed form solution for the step size in a
CG algorithm.
Algorithm 2 Conjugate Gradient (CG) Algorithm for Linearized WEMVA30 Reconstruction of Slowness
1: Initialize slowness ~s0 to a uniform profile: ~s0 = sinit~1 1
. sinit = 1540 m
s
2: Initialize step length parameter 1 . << 1 to dampen CG for cycle-skipping problem
3: for n = 0, ..., N 1 do . N Updates to Slowness Model
4: Compute all wavefields ~ytx(i) and ~yrx(i) , and images I~i based on current slowness model ~sn
5: Initialize slowness update ~s0 to zero: ~s0 = ~0
6: Initialize search direction d~0 = r ~s J( ~s0 ) . Jpower could be substituted for J
7: for k = 0, ..., K 1 do ⇣ ~ ⌘ . K Iterations of CG
d~T
kr ~s J( ~sk ) @ Ii+1 @ I~i
8: Compute Step Size: ↵k = PNtx ~ 2 where Li = @~s @~ s .
i=1 kLi dk k2
Increment Slowness Update: ~
9: ⇣ ⇣ ~sk+1 = ~skT + ↵k dk ⌘ ⌘
r ~s J( ~ sk+1 ) r ~s J( )T r
10: Momentum: k = min max sk+1 ) (r ~s J( ~
sk )T r ~s J( ~
r~s J( ~ sk )
~
sk ))
, 0 , r r~s J(~s J(~sk+1
sk )T r
~
s J(
~ ~
sk+1 )
s J( ~
~ sk )
11: Update Search Direction: d~k+1 = r ~s J( ~sk+1 ) + k d~k
12: end for
13: Update Slowness Model: ~sn+1 = ~sn + ~sK
14: end for
15: return ~sN
A linearized approach to WEMVA30 attempts to model the di↵erences between migrated images from neigh-
boring transmission events based on the underlying slowness. Specifically, this requires the assumption that
I~i (~s) ⇡ I~i+ 1 (~s) ⇡ I~i+1 (~s) so that any di↵erence between I~i (~s) and I~i+1 (~s) is solely due to an erroneous slowness
2
model. The I~i+ 12 (~s) term is simply a notation device to express the similarity of images from neighboring transmit
events. Using this notation, the di↵erences I~i+1 (~s) I~i (~s) can be modeled by the slowness in the medium:
Algorithm 2 summarizes a complete implementation of WEMVA30 based on the conjugate gradient algorithm.
However, because of the computational cost of WEMVA, this work only demonstrates the results of the first
gradient computation (equations (44) and (49)) in the algorithm as an initial proof of concept. In a future work,
we aim to provide a complete demonstration with a computationally-efficient implementation of WEMVA.
3. EXPERIMENTAL METHODS
3.1 USCT Simulation and Reconstruction
Figure 5 shows the ground-truth sound speed map from a numerical breast phantom and transmit pulse used to
simulate receive signals in k-Wave. This sound speed map was adapted and modified from a contrast-enhanced
cone-beam breast CT image in work from the Diagnostic Breast Center Göttingen (https://doi.org/10.1016/
j.tranon.2017.08.010)31 under the Creative Commons Attribution-NonCommercial-No Derivatives License
(CC BY NC ND) (https://creativecommons.org/licenses/by-nc-nd/4.0/). Plane waves (pulse bandwidth:
0.5-1.5 MHz) are transmitted through the tissue from a transmitting linear array (192 elements, 0.6 mm pitch) to
a receiving linear array (192 elements, 0.6 mm pitch). The distance between these arrays is zsep =120 mm. These
arrays are rotated 360 degrees around the medium in 2-degree steps. The reconstruction algorithms described
in Algorithm 1 and section 2.2 of the Theory are used to reconstruct the speed of sound in the medium. The
Transmitting Array 76
Observed Signals at 40 degrees
76
Observed Signals at 80 degrees
76
Observed Signals at 120 degrees
Sound Speed [m/s]
-0.06
1580
Time [ s]
Time [ s]
Time [ s]
78 78 78
-0.04
1560
Z Coordinate [m]
-0.02 80 80 80
1540
0 82 82 82
-0.05 0 0.05 -0.05 0 0.05 -0.05 0 0.05
0.02 1520 Lateral Element Position [m] Lateral Element Position [m] Lateral Element Position [m]
Observed Signals at 160 degrees Observed Signals at 200 degrees Observed Signals at 240 degrees
76 76 76
0.04 1500
Time [ s]
Time [ s]
Time [ s]
78 78 78
0.06
-0.06 -0.04 -0.02 0 0.02 0.04 0.06
X Coordinate [m] 80 80 80
Receiving Array
82 82 82
Pulse Timing -0.05 0 0.05 -0.05 0 0.05 -0.05 0 0.05
5
Signal (Mega-Pascals)
Lateral Element Position [m] Lateral Element Position [m] Lateral Element Position [m]
Observed Signals at 280 degrees Observed Signals at 320 degrees Observed Signals at 360 degrees
76 76 76
Time [ s]
Time [ s]
Time [ s]
0 78 78 78
80 80 80
-5 82 82 82
-4 -2 0 2 4 -0.05 0 0.05 -0.05 0 0.05 -0.05 0 0.05
Time [micro-seconds] Lateral Element Position [m] Lateral Element Position [m] Lateral Element Position [m]
Figure 5. Numerical breast phantom. Sound speed map [m/s], transmit pulse, and k-Wave simulated receive signals
from 9 di↵erent view angles. The sound speed map was adapted and modified from a contrast-enhanced cone-beam
breast CT image in work from the Diagnostic Breast Center Göttingen (https://doi.org/10.1016/j.tranon.2017.
08.010)31 under the Creative Commons Attribution-NonCommercial-No Derivatives License (CC BY NC ND) (https:
//creativecommons.org/licenses/by-nc-nd/4.0/)
4. NUMERICAL RESULTS
Axial [m]
Axial [m]
Axial [m]
Axial [m]
Axial [m]
5 6.55 6.55
0 0
For Specific
0 0 0 0 0
6.5 6.5
0 0
View Angles 6.45 -2 6.45
0.05 0.05 0.05 0.05 6.4
0.05 6.4
0.05
-5 -5 -4
6.35 6.35
Figure 6. Partial Gradients r~s Ji (~s, f ) from Specific View Angles Using Equation (27) Compared to Parallel-Beam CT
6.7 6.7
5
-0.05 -0.05 10 -0.05 -0.05 6.65 6.65
0 6.6
6.6
Axial [m]
Axial [m]
Axial [m]
Axial [m]
5 6.55 6.55
-5
0 0 0 0
6.5 6.5
0 -10
6.45 6.45
-0.05 -0.05
0 0
0.05 0.05 10 5 -0.05 -0.05
0 0
0.05 0.05
Lateral [m]Lateral [m] Lateral [m]Lateral [m]
Figure Complete Gradient Based on Accumulation of Partial Gradients from All View Angles and Frequencies: r~s J(~s) =
P PN7.views
f i=1 r~s Ji (~s, f ). Including higher frequencies in the complete gradient results in cycle-skipping artifacts (examples
are circled in red). Ground-truth slowness is given in units of seconds/meter
!
Modeled Signal 1
!
Ideal Update Actual Update
Direction Direction due to
Modeled Signal 2 Cycle Skipping
Figure 8. Visual Illustration of Cycle Skipping. Modeled Signal 1 simulates a medium with a slightly lower slowness
(higher speed of sound) that than the ground-truth medium responsible for the Observed Signals. The di↵erence between
Modeled Signal 1 and the Observed Signal results in a positive update to the slowness estimate to advance the simulated
wavefront. However, because Modeled Signal 2 simulates a much lower slowness than the ground truth medium, Modeled
Signal 2 leads the Observed Signal by more than a half cycle (transparent blue boxes around central peak). As a result,
the di↵erence between Modeled Signal 2 and the Observed Signal results in a negative update to the slowness in the
medium, aligning incorrect peaks in Modeled Signal 2 and the Observed Signal, even though the ideal update to the
slowness should be positive.
These gradient images are e↵ectively the backprojection of the error between the simulated and measured receive
signals for these particular view angles at a 0.5 MHz transmit frequency. However, in comparison to parallel-beam
CT where backprojection always occurs along straight line paths from sources to receivers, the backprojection
in USCT is based on di↵raction, which does not occur on straight-line paths. For this reason, if sources and
receiver switch places, the backprojection image from USCT can change, unlike parallel-beam CT where sources
and receivers are interchangeable.
The partial gradient images r~s Ji (~s, f ) in Figure 6 would be summed across view angles and frequencies to
P PNviews
form the complete gradient r~s J(~s) = f i=1 r~s Ji (~s, f ) as shown in Figure 7. However, when these gradient
images are accumulated in frequency, cycle-skipping causes parts of the image to invert. Recall that according
to equations (13) and (27), the gradient is computed by back-propagating the error between the simulated and
measured receive signals. When simulated and measured receive signals are o↵set by more than half a cycle of
the waveform, the error between the simulated and measured signals will induce updates to the slowness model
in the direction opposing the optimal or preferred alignment of simulated and received signals (see Figure 8).
Figures 9 and 10 show the frequency-dependent e↵ect of cycle skipping on the gradient computation. As
frequency increases, cycle skipping causes the same misalignment between simulated and measured signals to
increasingly update parts of the image in the opposite direction (Figure 9). Figure 10 is used to highlight cycle
skipping in a particular region of the image based on the backprojection from a selected transmission angle.
The slowness update in the circled region gradually inverts as the frequency increases. Figure 11 annotates each
part of the simulated and measured signals for this transmission angle and relates errors between the simulated
and measured signals to the backprojection image at that angle. Although parts of the slowness update behave
Axial [m]
Axial [m]
Axial [m]
Axial [m]
Axial [m]
Axial [m]
6.55 6.55
-2
0 0 0 0 -2 0 0
6.5 6.5
0.02 0.02 0.02 0.02 0.02 -4 0.02
-500 -4 6.45 6.45
0.04 0.04 0.04 0.04 0.04 0.04
-6 6.4 -6 6.4
0.06 0.06 0.06 0.06 0.06 0.06
6.35 6.35
-8 -8 -4
0.08 -1000
0.08 of Error 0.08 0.08 0.08Slowness -4 0.08 -4
BackprojectionBackprojection
of Error Ground-Truth
Ground-Truth Slowness
Backprojection of Error Backprojection
4
Ground-Truth
of Error
10 10
Slowness
5
Ground-Truth
10 Slown
-0.08 -0.05 -0.08 0 0.05 -0.08 4-0.05 0
-0.08 -0.05
0.05 -0.08
-0.08 0 -0.05
10 0.05 -0.08
0 0.05
-0.05 0 10 0.05 -0.08 -0.05 0
1.1 MHz Lateral [m] 1.3 MHz Lateral500
[m] 1.55[m]
Lateral MHz Lateral [m] 6.7 Lateral6.7
500
[m] 6.7 Lateral [m]
-0.06 -0.06 -0.06 2 -0.06 -0.06
-0.06 -0.06 -0.06
0 6.65 6.65
0 6.65
-0.04 -0.04 -0.04 -0.04 -0.04
-0.04 0 -0.04 -0.04
0 6.6
6.6 6.6
Slowness [s/m]
-500 -500
-0.02 -0.02 -0.02 -0.02 -0.02
-0.02 -0.02 -0.02
Axial [m]
[m]
Axial [m]
Axial [m]
Axial [m]
Axial [m]
Axial [m]
Axial [m]
-2 6.55 6.55 6.55
-1000 -5 -1000
0 0 0 0 00 0 0
Axial
-4 6.5 6.5 6.5
0.02 0.02 0.02 -1500
0.02 0.02
0.02 0.02 -1500 0.02
-6 -10 6.45 6.45 6.45
0.04 0.04 0.04 -2000
0.04 0.04
0.04 0.04 -2000 0.04
-8 6.4 6.4 6.4
0.06 0.06 0.06 0.06
-2500 0.06
0.06 0.06 -2500 0.06
-15 6.35 6.35 6.35
-10
0.08 0.08 0.08 0.08 0.08
0.08 0.08 0.08
5 4
-0.05 0 -0.05 0.05 0 0.05
-0.05
10 0 0.05
-0.05 0 10-0.05
-0.05 0.05 00 0.05
0.05
-0.05 0 0.05 -0.05 0
Lateral [m] Lateral [m] Lateral [m] Lateral [m] Lateral
Lateral [m]
[m] Lateral [m] Lateral [m]
Figure 9. Frequency Decomposition of Gradient Image r~s J(~s). As the frequency increases, the e↵ect of cycle skipping on
the gradient image increases.
-4 -4
Backprojection of Error Backprojection of Error Backprojection of Error
Ground-Truth Slowness Ground-Truth
10 Slowness Ground-Truth
10 Slown
-0.08 -0.08 -0.08 -0.08 -0.08 -0.08
0.5 MHz 0.7 MHz 0.9800
MHz 6.7 10000 6.7
-0.06 -0.06 -0.06 -0.06 -0.06 -0.06
10 6.65 8000 6.65
600
-0.04 -0.04 -0.04 -0.04 -0.04 6000
-0.04
400 6.6 6.6
-0.02 -0.02 -0.02 -0.02 -0.02 4000 -0.02
Axial [m]
Axial [m]
Axial [m]
Axial [m]
5
Axial [m]
Axial [m]
200 6.55 6.55
0 0 0 0 0 2000 0
0 6.5 6.5
0.02 0 0.02
0.02 0.02 0 0.02 0.02
-200 6.45 -2000 6.45
0.04 0.04 0.04 0.04 0.04 0.04
-400 6.4 -4000 6.4
0.06 0.06 -5 0.06 0.06 0.06 0.06
6.35 -6000 6.35
-600
0.08 0.08 4 0.08 0.08 -4 0.08 0.08 -4
Backprojection of Error 10Backprojection Ground-Truth Slowness
of Error 10
Backprojection of Error
Ground-Truth Slowness Ground-Truth
10 Slown
-0.08 -0.05 0 0.05 -0.08 -0.05
1.5 0
-0.08 -0.05
0.05 -0.08 0 -0.05 0.05 -0.08
0 -0.05
0.05 0 30 0.05 -0.08 -0.05 0
1.1 MHz Lateral [m] 1.3 MHz Lateral [m] 1.52000
Lateral MHz Lateral [m]
[m] 6.7 Lateral [m] 6.7 Lateral [m]
-0.06 -0.06 -0.06 -0.06 1500 -0.06 20 -0.06
1 6.65 6.65
-0.04 -0.04 -0.04 -0.04 1000 -0.04 10 -0.04
6.6 6.6
0.5 500
-0.02 -0.02 -0.02 -0.02 -0.02 0 -0.02
Axial [m]
Axial [m]
Axial [m]
Axial [m]
Axial [m]
Axial [m]
0 6.55 6.55
0 0 0 0 0 -10 0
0 -500 6.5 6.5
0.02 0.02 0.02 0.02 -1000 0.02 -20 0.02
6.45 6.45
-0.5 -1500
0.04 0.04 0.04 0.04 0.04 -30 0.04
6.4 6.4
-2000
0.06 0.06 -1 0.06 0.06 0.06 -40
0.06
-2500 6.35 6.35
0.08 0.08 0.08 0.08 0.08 0.08
-0.05 0 0.05 -0.05 0 -0.05
0.05 0 -0.05 0.05 0 0.05
-0.05 0 0.05 -0.05 0
Lateral [m] Lateral [m] Lateral [m] Lateral [m] Lateral [m] Lateral [m]
Figure 10. Frequency-Dependent Cycle-Skipping Artifact in Partial Gradient Image rJi (~s, f ) for Selected Transmission
Angle. A specific region where cycle skipping becomes dominant as frequency increases is circled in red.
ideally according to whether the simulated signals lag or lead the measured signals, if the lag or lead is too large,
parts of the image will update in the opposite direction based on cycle skipping.
When the conjugate gradient algorithm (Algorithm 1) is used to reconstruct the speed of sound in the
medium, the damping factor plays a crucial role in stabilizing the reconstruction algorithm in the presence of
cycle skipping artifacts. Cycle skipping is the main source of non-convexity in the waveform inversion problem.
Time [ s]
Time [ s]
-0.08 -0.08
2.5 → negative
79 79
Transmitting Array 6.7
-0.06
2
slowness update
-0.06 80
6.65
80
-0.04 -0.04 81 81
Observed
1.5 waves lag 6.6
82 82
-0.02 1behind projected
-0.02 -0.05 -0.04 -0.03 -0.02 -0.01 0 0.01 0.02 0.03 0.04 0.05 -0
Axial [m]
Axial [m]
6.55 [m]
0 wave
0.5
→ positive
0
Lateral Element Position
76
Observed Signals slowness update 76
Forward Projected6.5
Signals
0.02 0 0.02
77 77 6.45
-0.5
0.04
78
Even though0.04 trailing 78 6.4
Time [ s]
Time [ s]
0.06 waves lag behind
-1
0.06
79 79
Receiver Array projected waves,
-1.5 6.35
0.08 80 0.08 80
cycle skipping
81 -0.05 0 0.05 -0.05 0 0.05
causes → negative 81
Lateral [m] Lateral [m]
82
-0.05 -0.04 -0.03 -0.02 -0.01 0 0.01
slowness updates
0.02 0.03 0.04 0.05
82
-0.05 -0.04 -0.03 -0.02 -0.01 0 0.01 0.02 0.03 0.04 0.05
Lateral Element Position [m] Lateral Element Position [m]
Figure 11. Explanation of Cycle-Skipping Artifact in Gradient Image for the Selected Angle of Transmission. When the
measured signals appear sooner than the simulated signals, this indicates a negative update to the slowness image (circled
in yellow). When the measured signals lag behind than the simulated signals, this indicates a positive update to the
slowness image (circled in orange). However, if measured signals lag too far behind simulated signals (circled in red),
cycle skipping induces negative updates to the slowness image even though positive updates would be ideal.
As opposed to true linear least-squares minimization problems, where conjugate gradient is typically implemented
with a = 1 in its most standard form, waveform inversion inversion requires us to deliberately dampen or slow
down the conjugate gradient method to prevent overshooting when cycle skipping occurs. Figure 12 shows the
result of naively implementing the conjugate gradient method without damping ( = 1). The result is that
each step of the conjugate gradient method results in overshooting to compensate for the cycle skipping when
determining the appropriate step size based on the linearized forward model. The instability of the conjugate
gradient algorithm without damping causes the reconstruction to diverge. However, by damping the conjugate
gradient method ( = 1), the sound speed reconstruction converges to within 2.2 m/s RMS error after the same
12 iterations of the conjugate gradient method (Figure 13). By cutting down the size of each step in the conjugate
gradient method, the reconstruction works around the parts of the image where cycle skipping occurs, and the
cycle-skipping artifacts gradually disappear.
-0.06 -0.06 1570 -0.06 -0.06 1570 -0.06 1570 -0.06 1570 -0.06 1570 1570 -0.06 1570
Axial [m]
Axial [m]
Axial [m]
Axial [m]
Axial [m]
Axial [m]
Axial [m]
Axial [m]
1540 1540 1540 1540 1540 1540 1540
0 0 0 0 0 0 0 0
1530 1530 1530 1530 1530 1530 1530
0.04 0.04 0.04 0.04 0.04 1510 0.04 0.04 1510 1510 0.04 1510
1510 1510 1510
0.06 0.06 1500 0.06 0.06 1500 0.06 1500 0.06 1500 0.06 1500 1500 0.06 1500
-0.06 -0.06 1570 -0.06 -0.06 1570 -0.06 1570 -0.06 1570 -0.06 1570 1570 -0.06 1570
Axial [m]
Axial [m]
Axial [m]
Axial [m]
Axial [m]
Axial [m]
Axial [m]
Axial [m]
1540 1540 1540 1540 1540 1540 1540
0 0 0 0 0 0 0 0
1530 1530 1530 1530 1530 1530 1530
0.06 0.06 1500 0.06 0.06 1500 0.06 1500 0.06 1500 0.06 1500 1500 0.06 1500
-0.06 -0.06 1570 -0.06 -0.06 1570 -0.06 1570 -0.06 1570 -0.06 1570 1570 -0.06 1570
Axial [m]
Axial [m]
Axial [m]
Axial [m]
Axial [m]
Axial [m]
Axial [m]
0.04 0.04 0.04 0.04 1510 0.04 1510 0.04 1510 0.04 1510 1510 0.04 1510
1510
0.06 0.06 1500 0.06 0.06 1500 0.06 1500 0.06 1500 0.06 1500 1500 0.06 1500
Figure 12. Conjugate Gradient Reconstruction of Sound Speed Without Damping ( = 1). A naive implementation of
the conjugate gradient algorithm without any damping factor results in the divergence of the reconstructed sound speed
image. After 12 iterations, the algorithm continues to diverge from the ground-truth speed of sound.
produce a maximally focused point target at its current position in the image. In pulse-echo ultrasound imaging,
the primary challenge of applying WEMVA to migrated signals when there is a bulk error in the speed of sound
is the time-to-depth ambiguity. WEMVA will have initial difficulties in the depth placement of migrated targets
because of this latent ambiguity. However, the hope is that successive iterations of WEMVA with multiple
imaging targets using the conjugate gradient method should first resolve any laterally-varying component of
the speed of sound in the medium and then slowly resolve axially-varying components in the speed of sound
responsible for the depth placement of those imaging targets.
Figure 14(c) shows the same point target reconstruction and slowness gradient images when the linearized
form of WEMVA is used. Recall that the linearized form of WEMVA is based on modeling the di↵erence between
migrated images from neighboring transmit elements. Because linearized WEMVA models the di↵erences between
neighboring transmit elements, the resulting gradient images are less smooth and have stronger edge e↵ects due
to the truncation of the aperture. However, the same e↵ects seen in Figure 14(b) are seen Figure 14(c): when
cbf m < ctrue or sbf m > strue , there is a negative slowness update through the central portion of the gradient
flanked by positive updates; when cbf m > ctrue or sbf m < strue , there is a positive slowness update through
the central portion of the gradient flanked by negative updates. Despite the rough appearance of the gradient
images in Figure 14(c), previous work by30 indicates that the integration of these gradients across multiple
imaging targets leads to meaningful updates to the slowness model.
-0.06 -0.06 1570 -0.06 -0.06 1570 -0.06 1570 -0.06 1570 -0.06 1570 1570 -0.06 1570
Axial [m]
Axial [m]
Axial [m]
Axial [m]
Axial [m]
Axial [m]
Axial [m]
1540 1540 1540 1540 1540 1540 1540
0 0 0 0 0 0 0 0
1530 1530 1530 1530 1530 1530 1530
0.04 0.04 1510 0.04 0.04 1510 0.04 1510 0.04 1510 0.04 1510 1510 0.04 1510
0.06 0.06 1500 0.06 0.06 1500 0.06 1500 0.06 1500 0.06 1500 1500 0.06 1500
-0.06 -0.06 1570 -0.06 -0.06 1570 -0.06 1570 -0.06 1570 -0.06 1570 1570 -0.06 1570
Axial [m]
Axial [m]
Axial [m]
Axial [m]
Axial [m]
Axial [m]
Axial [m]
1540 1540 1540 1540 1540 1540 1540
0 0 0 0 0 0 0 0
1530 1530 1530 1530 1530 1530 1530
0.04 0.04 1510 0.04 0.04 1510 0.04 1510 0.04 1510 0.04 1510 1510 0.04 1510
0.06 0.06 1500 0.06 0.06 1500 0.06 1500 0.06 1500 0.06 1500 1500 0.06 1500
-0.06 -0.06 1570 -0.06 -0.06 1570 -0.06 1570 -0.06 1570 -0.06 1570 1570 -0.06 1570
Axial [m]
Axial [m]
Axial [m]
Axial [m]
Axial [m]
Axial [m]
Axial [m]
1540 1540 1540 1540 1540 1540 1540
0 0 0 0 0 0 0 0
1530 1530 1530 1530 1530 1530 1530
0.04 0.04 1510 0.04 0.04 1510 0.04 1510 0.04 1510 0.04 1510 1510 0.04 1510
0.06 0.06 1500 0.06 0.06 1500 0.06 1500 0.06 1500 0.06 1500 1500 0.06 1500
Figure 13. Conjugate Gradient Reconstruction of Sound Speed With Damping ( = 0.25). By damping the step size, the
conjugate gradient algorithm converges to the correction speed of sound after 12 iterations. The RMS error in between
the final reconstruction and the ground-truth speed of sound is 2.2 m/s.
12 12
(a) 12 12 12 10
5
12
Element 1
12
1
12 12 10
5
Element 20
12 12 12
2
Element 39
12 12 12 12
Element 58
12 12 12
Element 77
12 12 12 12
Element 96
12 12 12 12 12 12 1
10 Gradient
ness 10 Gradient
Slowness
Slowness 10 Gradient
10 Gradient
Image Slowness
Gradient
Slowness
Reconstruction 10 Gradient
10 Gradient
Slowness
Slowness
Slowness 10 Gradient
Gradient
Slowness
Slowness Slowness
Slowness
Gradient 10
10 Gradient
10 0.5Gradient 10 Gradient
10 Gradient
Slowness
Slowness
Slowness 10 Gradient
Gradient
Slowness Slowness
Slowness
Slowness
Gradient 10
10 Gradient
10 Gradient Slowness 10Slowness
10 Gradient
Slowness
Slowness Gradient
Slowness
Gradient
Gradient 10 Gradient
10 Gradient
Slowness
Slowness 10
10 Gradient 10 Gradient
10 Gradient
Slowness
Slowness
Slowness 10 Gradient
Gradient
Slowness Slowness
Slowness 10
10 Gradient
10 Gradient 10
10 Grad
Slowness
Slowness
3 0 0 3 00 0 3
3 Reconstruction
0 0 030 3 00 0 30 3 3 015 30 30 30 0 0 30 3 3 0 30 300 3 0 0 30 3 3 0 30 30 3 0 0 30 3 3 0 30 3 3
z Axial Distance (mm)
(mm)
(mm)
Distance (mm)
15 15 15
15 15-4 -2 0 15
2 15
4 15 35 1515 15 15 1535 15 15 15 15 15 15 151 15 15
15 15 15 15 15 15 15 15 15 15 15 15
1 1 1
x Azimuthal 1
Distance (mm) 1
-20 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
-1.5 -2
40 40 0 0
20 20 20
20 20 20 20 20 2020 20 20 20 20 20 20 20 20 20 20 20 20
20 20 20 20 20 20 20 20 20 20 20 20
AxialDistance
AxialDistance
45 -2 45
-3
0
0 25 25 0 25
25 25 0 0 25 25 0
25
-30 0 2525 25 0
25 0 0 25 25 0
0-0.5 25 25
0 25 25 0
25 0 0 25 0
25 025
25 0 25 25 0
25 0 0-0.5 25 0
25 0
25 0 25 25 0
25 0 0 25 0
25 0 0
-10 0 10 10 12 -10 0 10 10 12
30 30 30
30 30 30 30 30 3030
x Azimuthal 30(mm)
30Distance 30 x30 30Distance30
Azimuthal (mm)30 30 30 30-1 30 30
30 30 30 30 30 30 30 30 30 30 30 30
z zAxial
z zAxial
-1 -1
-1 -1 -1 -1 -1
-40 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1
35 35 35
35 35 35 35 35 3535 35 35 35 35 35 35 35 35 35 35 35 35
35 35 35 35 35 35 35 35 35 35 35 35
-1.5 -1.5
40 40 40
40 40 40 40 40 4040 40 40 40 40 40 40 40 40 40 40-2 40 40
40 40 40 40 40 40 40 40 40 40 40 40
-2 -2 -2 -2 -2
-50 -2 -2 -2 -2 -2 -2 -2 -2 -2 -2 -2 -2 -2 -2 -2 -2 -2 -2 -2 -2 -2 -2 -2 -2 -2
45 45 45
45 45 45 45 45 4545 45 45 45 45-2 45 45 45 45 45 45 45 45
45 45 45 45 -2 45 45 45 45 45 45 45 45
-3
-3 -3 50 -3 -3 -3
-60 -3 -3 -3 -3 -3 -3 -3 -3 -3 -3 -3 -3 -3 -3 -3 -3 -3 -3 -3 -3 -3 -3 -3 -3 -3
0 10 -10 -10 0 0 10 -10
10
-10 -10 00 0 10
10 10
-10 -10 0 -10 0 10 0-10
10
-10 -10
100 0 -10 0 1010 0 10 10 12 -10 0 -10 0 10
10
-10 -10 0-10
10 0-10 100 10 0 1010 12 -10
100 -10 10 -100 10 0-10
-10 0 -10 010 1010
-10 12
0 -10 0 10 10 0 10 10
-10 -10 0 -10 0 10 0-10
10 -10
100 -10 0 10 0 10 10
-10 -10 0 0
x Azimuthal
hal Distance (mm) Distance
x Azimuthal (mm)
Distance
x Azimuthal
(mm)
x AzimuthalDistance
x Azimuthal
Distance (mm)
Distance
(mm) x Azimuthal
(mm)
x Azimuthal Distance
x Azimuthal
Distance (mm) (mm)
Distance
x Azimuthal
x Azimuthal (mm)
x Azimuthal Distance
x Distance
Azimuthal
Distance (mm)
(mm) (mm)
Distance x Azimuthal
(mm)
x Azimuthal Distance
x Azimuthal
Distance (mm)
Distance
x(mm) x Azimuthal
x Azimuthal
Azimuthal (mm)
Distance Distance
x Azimuthal
Distance
(mm) (mm)
Distance
(mm) x Azimuthal
(mm)
x Azimuthal Distance
x Azimuthal
Distance (mm)
x Azimuthal (mm)
Distance
x Azimuthal
x Azimuthal
Distance (mm)
(mm) Distance
x Azimuthal
Distance (mm) (mm)
Distance x Azimuthal
(mm)
x Azimuthal Distance
x Azimuthal
Distance (mm) (mm)
Distance
x Azimuthal
x Azimuthal(mm) Distance
x Azimuthal
Distance (mm)
Distance
(mm) x Azimuthal
(mm)
x Azimuthal D
Distanc
5 5 5 5 5 5 5 5 5 5 5
5
Images
15 15 15 15 15 15 15 15 15 15 15 15
-20 -20 -20 1 -20 -20 2 1
-20 2 1 2
Migrated
20 20 20 20 20 20 20 20 20 20 20 20
30 30 30 30 30 30 30 30 30 30 30 30
Sum ofz Axial
-2 -2 -40 -40 -2 -1
-40 -40 -1 -40 -40 -1
35 35 35 35 35 35 35 35 35 35 35 35
40 40 40 40 40 -440 40 40 40 -440 40 40 -4
-50 -50 -2 -50 -50 -2 -50 -50 -2
45 45 45 45 45 45 45 45 45 45 45 45
-6 -6 -6
-60 -60 -3 -60 -60 -3 -60 -60 -3
-10 -10 0 0 10 10 -10 -10 0 0 10 10 -10 -10 0 0 10 10 -10 -10 0 0 10 10 -10 -10 0 0 10 10 -10 -10 0 0 10 10
x Azimuthal Distance
x Azimuthal (mm) (mm)
Distance x Azimuthal Distance
x Azimuthal (mm) (mm)
Distance x Azimuthal
x Azimuthal Distance
Distance (mm) (mm) x Azimuthal
x Azimuthal Distance (mm) (mm)x Azimuthal
Distance Distance
x Azimuthal (mm) (mm)
Distance x Azimuthal Distance
x Azimuthal (mm) (mm)
Distance
5 5
Linearized WEMVA Based
5 5 5 5 5 5 5 5 5 5
on Differences Between
15 15 15 15 15 15 15 15 15 15 15 15
-20 -20 -20 2 1 -20 -20 2 1
-20 2 1
20 20 20 20 20 20 20 20 20 20 20 20
30 30 30 30 30 30 30 30 30 30 30 30
-2 -2 -2
Successive
40 40 40 40 40 -440 40 40 40 -440 40 40 -4
-50 -50 -2 -50 -50 -2 -50 -50 -2
45 45 45 45 45 45 45 45 45 45 45 45
-6 -6 -6
-60 -60 -3 -60 -60 -3 -60 -60 -3
-10 -10 0 0 10 10 -10 -10 0 0 10 10 -10 -100 0 10 10 -10 -10 0 0 10 10 -10 -10 0 0 10 10 -10 -10 0 0 10 10
x Azimuthal Distance
x Azimuthal (mm) (mm)
Distance x Azimuthal Distance
x Azimuthal (mm) (mm)
Distance x Azimuthal
x Azimuthal Distance
Distance (mm) (mm) x Azimuthal
x Azimuthal Distance (mm) (mm)x Azimuthal
Distance Distance
x Azimuthal (mm) (mm)
Distance x Azimuthal Distance
x Azimuthal (mm) (mm)
Distance
Figure 14. Slowness Gradients for the Pulse-Echo Reconstructions of a Point Target. (a) Point target image and wave
paths induced between each transmit element and the point target when the image is focused at a sound speed cbf m equal
to the ground-truth sound speed ctrue = 1540 m/s in the medium. (b) Point target reconstructions and slowness update
direction images based on equation (44) when cbf m = ctrue , cbf m = 0.96ctrue , and cbf m = 1.04ctrue . This WEMVA
implementation is based on the coherent sum of migrated images. (c) Point target reconstructions and slowness update
direction images based on equation (49) when cbf m = ctrue , cbf m = 0.96ctrue , and cbf m = 1.04ctrue . Here, WEMVA is
based on modeling the di↵erences between migrated images from neighboring transmit elements.
This work also translates the same modeling framework used in waveform inversion to derive slowness updates
for pulse-echo ultrasound imaging. The initial work presented here uses a point target to demonstrate two di↵er-
ent implementations of wave-equation migration velocity analysis from seismic imaging. In each implementation,
this work examines challenges such as the time-to-depth ambiguity involved in scaling the initial demonstration
in point targets to more complex media using an iterative conjugate gradient algorithm (Algorithm 2). A com-
plete demonstration of WEMVA for medical pulse-echo ultrasound imaging in more complex media is left to
future work.
ACKNOWLEDGMENTS
The author would like to acknowledge Dr. Biondo Biondi and Joseph Jennings, from the Department of Geo-
physics at Stanford University, for their expertise and advice on waveform inversion and migration velocity
analysis. The author would also like to acknowledge feedback from various faculty within the Stanford Radiolog-
ical Sciences Laboratory (RSL) such as Dr. Jeremy Dahl and Dr. Adam Wang. Finally, the author would like
REFERENCES
[1] Li, C., Duric, N., Littrup, P., and Huang, L., “In-vivo breast sound-speed imaging with ultrasound tomog-
raphy,” Ultrasound in Medicine and Biology 35(10), 1615–1628 (2009).
[2] Duric, N., Littrup, P., Poulo, L., Babkin, A., Pevzner, R., Holsapple, E., Rama, O., and Glide, C., “Detection
of breast cancer with ultrasound tomography: First results with the computed ultrasound risk evaluation
(CURE) prototype,” Medical physics 34(2), 773–785 (2007).
[3] Schreiman, J., Gisvold, J., Greenleaf, J. F., and Bahn, R., “Ultrasound transmission computed tomography
of the breast.,” Radiology 150(2), 523–530 (1984).
[4] Wiskin, J., Borup, D., Johnson, S., and Berggren, M., “Non-linear inverse scattering: high resolution
quantitative breast tissue tomography,” The Journal of the Acoustical Society of America 131(5), 3802–
3813 (2012).
[5] Kolb, T. M., Lichy, J., and Newhouse, J. H., “Occult cancer in women with dense breasts: detection with
screening US–diagnostic yield and tumor characteristics.,” Radiology 207(1), 191–199 (1998).
[6] Kolb, T. M., Lichy, J., and Newhouse, J. H., “Comparison of the performance of screening mammography,
physical examination, and breast US and evaluation of factors that influence them: an analysis of 27,825
patient evaluations,” Radiology 225(1), 165–175 (2002).
[7] Hormati, A., Jovanović, I., Roy, O., and Vetterli, M., “Robust ultrasound travel-time tomography using the
bent ray model,” in [Medical Imaging 2010: Ultrasonic Imaging, Tomography, and Therapy ], 7629, 76290I,
International Society for Optics and Photonics (2010).
[8] Quan, Y. and Huang, L., “Sound-speed tomography using first-arrival transmission ultrasound for a ring
array,” in [Medical Imaging 2007: Ultrasonic Imaging and Signal Processing ], 6513, 651306, International
Society for Optics and Photonics (2007).
[9] Wang, K., Matthews, T., Anis, F., Li, C., Duric, N., and Anastasio, M. A., “Waveform inversion with source
encoding for breast sound speed reconstruction in ultrasound computed tomography,” IEEE transactions
on ultrasonics, ferroelectrics, and frequency control 62(3), 475–493 (2015).
[10] Krebs, J. R., Anderson, J. E., Hinkley, D., Neelamani, R., Lee, S., Baumstein, A., and Lacasse, M.-D.,
“Fast full-wavefield seismic inversion using encoded sources,” Geophysics 74(6), WCC177–WCC188 (2009).
[11] Vyas, U. and Christensen, D., “Ultrasound beam simulations in inhomogeneous tissue geometries using
the hybrid angular spectrum method,” IEEE transactions on ultrasonics, ferroelectrics, and frequency con-
trol 59(6), 1093–1100 (2012).
[12] Stolk, C. C. and de Hoop, M. V., “Modeling of seismic data in the downward continuation approach,” SIAM
journal on applied mathematics 65(4), 1388–1406 (2005).
[13] Wiskin, J., Borup, D., Iuanow, E., Klock, J., and Lenox, M. W., “3-d nonlinear acoustic inverse scatter-
ing: Algorithm and quantitative results,” IEEE transactions on ultrasonics, ferroelectrics, and frequency
control 64(8), 1161–1174 (2017).
[14] Pérez-Liva, M., Herraiz, J., Udı́as, J., Miller, E., Cox, B., and Treeby, B., “Time domain reconstruction of
sound speed and attenuation in ultrasound computed tomography using full wave inversion,” The Journal
of the Acoustical Society of America 141(3), 1595–1604 (2017).
[15] Ruiter, N. V., Zapf, M., Hopp, T., Gemmeke, H., and van Dongen, K. W., “Usct data challenge,” in [Medical
Imaging 2017: Ultrasonic Imaging and Tomography ], 10139, 101391N, International Society for Optics and
Photonics (2017).
[16] Ruiter, N. V., Zapf, M., Hopp, T., Gemmeke, H., van Dongen, K. W., Camacho, J., Herraiz, J. L., Liva,
M. P., and Udı́as, J. M., “Usct reference data base: conclusions from the first spie usct data challenge
and future directions,” in [Medical Imaging 2018: Ultrasonic Imaging and Tomography ], 10580, 105800Q,
International Society for Optics and Photonics (2018).
[17] Sanabria, S. J., Ozkan, E., Rominger, M., and Goksel, O., “Spatial domain reconstruction for imaging speed-
of-sound with pulse-echo ultrasound: simulation and in vivo study,” Physics in Medicine & Biology 63(21),
215015 (2018).