optimized noninvasive monitoring of thermal changes on digital b-mode renal sonography during...
TRANSCRIPT
-
8/14/2019 Optimized Noninvasive Monitoring of Thermal Changes on Digital B-Mode Renal Sonography During Revasculariza
1/13
Optimized Noninvasive Monitoring ofThermal Changes on Digital B-ModeRenal Sonography DuringRevascularization Therapy
Objective. Noninvasive real-time thermal change monitoring of human internal organs can play a crit-
ical role in diagnosis and treatment of many disorders, including reperfusion of renal arteries during
anticoagulation therapy. Methods. This article focuses on tissue temperature detection using ultra-
sound velocity changes in different structures and their related speckle shift from their primary loca-
tions on high-quality B-mode digital sonography. We evaluated different speckle-tracking techniques
and optimized them using appropriate motion estimation methods to determine the best algorithm
and parameters. Results. Performing thermal detection methods on simulated phantoms showed a
good correlation between speckle shifts and the ground truth temperature. For the simulated images,
average thermal error was 0.5C with an SD of 0.5C, where lower errors can be obtained in noiseless
(motionless) data. The proposed technique was evaluated on real in vivo cases during surgical occlu-
sion and reopening of the renal segmental artery and showed the potential of the algorithm for obser-
vation of internal organ changes using only digital ultrasound systems for diagnosis and therapy.
Conclusions. The adaptive Rood pattern search proved to be the best block-matching technique,
whereas the multiresolution Horn-Schunck technique was the best gradient optical flow method. The
extracted thermal change during in vivo revascularization therapy is promising. In addition, we present
an evaluation of several block-matching and optical flow motion estimation techniques. Key words:
block matching; optical flow; reperfusion therapy; speckle tracking; ultrasound thermometry.
Received February 4, 2009, from the Department ofBiomedical Engineering, Tehran University ofMedical Sciences, Tehran, Iran (M.D.A.); Departmentof Electrical and Computer Engineering, Universityof Louisville, Louisville, Kentucky USA (V.T.); andScience and Research Branch, Islamic AzadUniversity, Tehran, Iran (N.S.). Revision requested
March 16, 2009. Revised manuscript accepted forpublication April 21, 2009.
We thank Ali Najafian and Hossein Jandaghi forassistance with animal surgery and sonographicimage acquisition and Ehsan Mohammadi for
preparing agar gel phantoms. This work was sup-ported in part by the Research Center for Scienceand Technology in Medicine, Tehran University ofMedical Sciences.
Address correspondence to Vahid Tavakoli, MD,MS, Department of Electrical and ComputerEngineering, University of Louisville, Medical ImageComputing Laboratory, Lutz Hall, Room 308,Louisville, KY 40292 USA.
E-mail: [email protected]
AbbreviationsARPS, adaptive Rood pattern search; ES, exhaustive search;4SS, 4-step search; MAD, mean absolute difference; MSE,mean square error; PSF, point spread function; ROI, regionof interest; 3SS, 3-step search; 2D, 2-dimensional
Tissue Temperature Detection
Thermal changes in different parts of the human body
are directly related to the physiologic characteristics and
functions of the organs. The main supplier of the heat for
any organ is blood flow, and any disturbance in tissue
perfusion can be shown as a temperature gradient. Also,
many nonvascular disorders such as inflammations and
abscess formation are related to the changes in tempera-
ture. Therefore, superficial or mucosal thermal detection
(ordinary thermometer) is a major part of any physicalexamination in any clinic.1 However, for a deeper tem-
perature measurement, there is a lack of a standard tech-
nique. Moreover, ordinary electronic sensors (infrared
detectors and thermistors) are restricted to local use in
specific parts, are implanted, and are also invasive. In this
article, we introduce a novel method for assessing ther-
mal changes using speckle shifts on high-quality B-mode
sonography. This technique is not only noninvasive but
also real time and easily performable on an entire organ
even in 3-dimensional views.2
2009 by the American Institute of Ultrasound in Medicine J Ultrasound Med 2009; 28:15351547 0278-4297/09/$3.50
Technical Advance
Author: Please give degrees for names mentioned in acknowledgments.
Mohammad D. Abolhassani, PhD, Vahid Tavakoli, MD, MS, Nima Sahba, MS
-
8/14/2019 Optimized Noninvasive Monitoring of Thermal Changes on Digital B-Mode Renal Sonography During Revasculariza
2/13
The proposed technique is used to observe the
thermal changes during renal revascularization
therapy because revascularization (opening an
occluded artery) redirects the blood flow to the
tissue and elevates the temperature.
Thermal Detection Using Sonography
Use of laser and radio frequency knives, which
are based on thermal-controlling applications,
requires accurate thermal detection. Treatment
monitoring consists of real-time tissue damage
monitoring and real-time temperature estima-
tion.3,4Regarding ultrasound thermometry, different
methods have been proposed so far, such as analysis
of the frequency-dependent attenuation,5 speed
of sound, thermal expansion,69 and back-scattered
power.10 One novel technique is ultrasound-simulated acoustic emission, which was shown
to be temperature dependent.11 Konofagou et al12
used ultrasound-simulated acoustic emission in
detection and monitoring of thermal lesion for-
mation and thermal therapy.13,14
For many materials, the ultrasound velocity is
dependent on the temperature. The estimations
of the temperature variations are obtained by
calculating the axial displacement of the scatter-
ers due to tissue warm-up. The displacements
are caused primarily by the changes of the speed
of sound in the tissue when the temperature
changes. The second reason for the changes of
the signals is the thermal expansion of tissue. The
dependency of the speed of sound on the tem-
perature introduces apparent or virtual shifts in
the scatterer positions, and the thermal expan-
sion of the medium introduces physical shifts in
the scatterer positions. The variation of the speed
of sound with respect to the temperature is for-
mulated as Equation 1:
(1)
Where c is the ultrasound velocity; T is the
temperature at distance x; c0
(x) = c(x, T0); and
is the coefficient of speed change
with regard to temperature. Moreover,
(2)
where is a medium-dependent
parameter; is linear coefficient of thermal
expansion; and dis the displacement.
Details of the formulation have been describedpreviously. In a novel method, Abolhassani et al2
proved the correlation between the temperature
estimation and the thermal speckle shifts on
high-quality B-mode digital sonography in tissue-
mimicking phantoms through which a heater
electrode and an array of thermistors were
inserted. The average error of their study was
0.2C; the maximum error was 0.53C; and the SD
was 0.53C.
Speckle-Tracking Methods
To date, there have been many methods ofspeckle tracking between 2 ultrasound images.
Block matching is one of them, which focuses on
the gray scale similarity between 2 sliding blocks
of an image. There are many methods of observ-
ing the similarity between 2 blocks; for instance,
the mean absolute difference (MAD), mean
square error (MSE), maximum likelihood, and
entropy maximization. Moreover, there are sev-
eral approaches to using these block-matching
techniques. An old and famous one is exhaustive
block matching, but its time consumption has
led to newer methods with comparable results,
such as the 3-step search (3SS), 4-step search
(4SS), cross-diamond search, adaptive Rood pat-
tern search (ARPS), and simple and efficient
algorithms. These techniques have been dis-
cussed in detail previously.1523
Another group of motion estimation methods,
introduced in 1981, was the gradient optical flow
methods such as Lucas-Kanade (which pre-
sumes intensity constancy between the consec-
utive frames) and Horn-Schunck (a velocity
constancy assumption is added to the previousmethod). Many other optimizations have been
performed on these techniques, such as phase
matching, discrete cosine transform preprocess-
ing, uniform or gaussian prefiltering, invariant
moment implementation, histogram equaliza-
tion, a multiresolution approach (Suhling-Unser
strategy), and new iterations or warping flow
charts.2325
1536 J Ultrasound Med 2009; 28:15351547
Monitoring of Thermal Changes During Revascularization Therapy
-
8/14/2019 Optimized Noninvasive Monitoring of Thermal Changes on Digital B-Mode Renal Sonography During Revasculariza
3/13
Moment Implementation and Multiresolution
Strategy
During motion, pixels of an image not only
translate but also shear and rotate. Based on this
fact, many motion estimation methods makeuse of moments (Zernike or spline) to achieve a
high degree of invariance and robustness to
shear and noise.2628
On the other hand, moment implementation is
not the whole story. In each frame, pixels have
different motion vectors in each region; ie, each
region of an image translates, rotates, or shears
into different sizes. To overcome this problem, we
proposed a multiresolution strategy inspired by
the Suhling-Unser multiresolution method.29,30
In the Suhling-Unser technique, coarse kernels
are obtained on the whole image, and then finerkernels are implemented until a certain amount
of error is satisfied. It means that areas of large
motion are computed using coarser windows,
and areas of fine motions are computed using
small moments. Using this technique, we can
capture smaller regions of motion that would be
invisible to a large-kernel method.
From Speckle Tracking to Thermal Change
Detection
Equations 1 and 2 relate the speckle displace-
ment to the thermal change. Each speckle shift
shows its own thermal shift in addition to the
thermal shift of the proximal speckles (with
respect to the probe). Therefore, there is an accu-
mulation effect in the thermal speckle shift on
B-mode ultrasound images.
To recover the thermal change of each speckle,
it has been proven that vertical displacement
derivation is necessary. The derivation step can
enhance the noise, and it is crucial to preserve
the robustness of the proposed method. Some
articles have discussed the noise sensitivity ofthe thermal change detection during the vertical
derivation and showed that an order statistics fil-
ter can extract the thermal changes from the
speckle displacements.31
Handling Motion Artifacts in the z-Axis
Several studies have shown that the optical
flowbased methods can handle large motion arti-
facts in the z-axis. Several studies extracted motion
vectors on 2-dimensional (2D) B-mode echocar-
diographic images from the 3-dimensional heart
motion.3234As will be discussed later, this error is
called an out-of-plane error and can be overcome
using temporal windowing and confidence mea-
surement.
Materials and Methods
In this section, we present different motion esti-
mation techniques that were used in this project.
Several methods were implemented and opti-
mized using different parameters. The algo-
rithms were evaluated on the simulated thermal
change in the simulated tissue to evaluate their
effectiveness and performances.
Motion Estimation Methods for Speckle TrackingThere are many different methods for extracting
motion between different images, such as block-
matching techniques, gradient-based methods
(optical flow), and feature correlation. Figure 1
illustrates different types of motion estimation
methods.
The block-matching methods try to find a sim-
ilar block in 2 frames. Optical flow is inspired by
the block-matching method, whereas both of
them assume intensity constancy in a region or
between several pixels. Nevertheless, optical flow
algorithms use advanced variational calculus
techniques that make them very fast. The feature
matching and correlation techniques extract 2
similar features in 2 frames. These techniques
suffer from loss of features, finding good features
to track, and providing dense motion fields.
However, the optical flow techniques are very
fast, but their sensitivities to noise can induce
inaccurate results.
The best approach to find an efficient motion
estimation technique is the implementation of
different methods on simulated or real images.The simulated images provide us with accurate
measurements, whereas it is usually very difficult
to obtain accurate and dense results from the
real sequences.
To find out an optimal technique for ultra-
sound thermal detection, we implemented dif-
ferent algorithms on the simulated ultrasound
images and then validated the results on the real
ones. Figure 1 shows different types of motion
estimation techniques.
J Ultrasound Med 2009; 28:15351547 1537
Abolhassani et al
-
8/14/2019 Optimized Noninvasive Monitoring of Thermal Changes on Digital B-Mode Renal Sonography During Revasculariza
4/13
There are several major differences between
echocardiographic series and camera sequences.
Camera movies primarily suffer from occlusion
and aperture and transparency problems. On the
other hand, echocardiographic series mainly sufferfrom an out-of-plane error, a low signal to noise ratio,
and violation of the Nyquist theorem in the spa-
tiotemporal dimension. Echocardiographic motion
estimation is especially dependent on spatiotem-
poral aliasing and an out-of-plane error; there-
fore, higher frame rates can dramatically increases
the accuracy.
Block-Matching Algorithms
The main idea toward this type of motion esti-
mation is the fact that each block of a frame
moves toward a similar block in another frame ifthe time interval is small enough. The general
strategy is to move a microblock over the other
blocks and compute the most similar blocks. To
find the best block matching, it is necessary to
have a similarity-matching function called the
cost function, through which similarity between
consecutive images is extracted. There are sever-
al famous block-matching algorithms including
the MAD and MSE:
(3)
(4)
whereNis the size of the macroblock, and Cand
R define pixel shifting.
Exhaustive Search Method (Full Search Method)
This method computes the cost function of any
block of the macroblock (area in which we arelooking for a smaller block). This is a very time-
consuming but an exact method.
Another variation is the cross-correlation
method. In this algorithm, 1 block of the image
slides over the next image, and the most corre-
lated block to the primary block is considered
a matched block with which motion can be
estimated:
(5)
The maximum peak of c(m, n) indicates the
speckle displacement, where SP1
and SP2
are the
sliding microblock and macroblock, respectively.
The usual correlation methods are computation-
ally expensive, and this problem has led to manyoptimizations in this area, such as gradient
descent-based approaches.
Other Variations of Block Matching
To cope with the computational expense of the
exhaustive search (ES), several searching strategies
are proposed, such as the 3SS, 4SS, simple and effi-
cient search, new 3SS, diamond search, and ARPS.
Gradient-Based (Optical Flow) Methods
Horn-SchunckThe general differential formulation of the opti-
cal flow computation assumes that the intensity
of a particular point in a moving pattern does not
change with time.6 This constant intensity
assumption is written as
(6)
where E is the pixel intensity with the gradient
toward the spatial and temporal dimension, and
is a weighting parameter.23
Locally Constant Motion (Lucas-Kanade)
As an alternative to the brightness constancy
model in Equation 6, the Lucas-Kanade method
(Equation 7) proposed to solve the problem on
small local windows where the motion vector is
assumed to be constant. We ought to minimize
this equation:
(7)
where Eis the pixel intensity, and Wis a weight-
ed vector.
Results will be
ATW2AV = ATW2b ,
where W = diagonal [W(x1),...,W(x
n)]; b = (E
t
(x1),...,(Et (xn)); and u and vare our motion vec-
tors inxandydirections.25
1538 J Ultrasound Med 2009; 28:15351547
Monitoring of Thermal Changes During Revascularization Therapy
-
8/14/2019 Optimized Noninvasive Monitoring of Thermal Changes on Digital B-Mode Renal Sonography During Revasculariza
5/13
Multiresolution Approach
Primarily coarse kernels are obtained on the
whole image, and then finer kernels are imple-
mented until a certain amount of error is satis-
fied. Therefore, large motions are computedusing coarse windows, and fine motions are com-
puted using small moments. Multiresolution
moment implementation for ultrasound images
has been discussed in detail previously.2630
Motion Extraction in Simulated Data
Renal tissue movement consists of 2 major sub-
group motions: (1) motion artifacts due to respi-
ration and patient displacement and (2) thermal
speckle shift. The steps resulting to the thermal
estimation are shown in Figures 2 and 3.
It is obvious that for accurate motion estima-tion, we ought to remove the first group, which
is more prominent and destructive. B-mode
sonography has a periodic nature because of res-
piratory motion; therefore, noise removal is cru-
cial for obtaining a motionless frame. The
following steps are implemented to extract the
thermal change motion:
1. Region of interest (ROI) where renal tissue is
extracted. The pixels in this area define the renal
tissue, but those outside the ROI pertain to
abdominal organs.
2. Region of interest neighborhood motion esti-
mation. The motion vectors around the ROI area
are extracted, and on the basis of these vectors,
the interpolation is performed in the ROI using a
cubic B-spline to approximate the pure ROI
motion artifact vectors.
3. Warping strategy. The fact that there are
always some degrees of motion in the proximity
of renal tissue, even in carefully fixed frames,
necessitates ROI adjustment. Note that we have
a combination of the motion artifact and the
thermal speckle shift inside the renal tissue.A motion interpolation step using cubic B-spline
interpolation (called Ispline) was implemented
in the previous step with the neighboring renal
pixels (ROI) to extract the motion artifact in the
ROI. After that, the motion artifact vectors adjust
the primary image pixel shifting after a simple
subtraction.
4. After the ROI adjustment, the motion estima-
tion is performed in the ROI. Many speckle-
tracking techniques were implemented on
images to evaluate their efficacy and time con-
sumption in favor of the novel optimized tech-
nique. Shifts in B-mode frames show thermal
velocity changes if motion artifacts are desirably
removed. We implemented block-matchingmethods (3SS, 4SS, cross-diamond search, sim-
ple and efficient search, and ES using the MSE
and MAD) and gradient-based optical flow
methods (Lucas-Kanade, Horn-Schunck, and
iterative Lucas-Kanade with different numbers
of iteration). These methods were performed
using Zernike and spline moments to observe
the accuracy of different sizes, orders, and types.
J Ultrasound Med 2009; 28:15351547 1539
Abolhassani et al
Figure 1. Different types of motion estimation techniques.
Figure 2. Schema of thermal extraction steps.
-
8/14/2019 Optimized Noninvasive Monitoring of Thermal Changes on Digital B-Mode Renal Sonography During Revasculariza
6/13
Transformation From the Speckle Shift to the
Thermal Change
According to Equation 2, knowing k, the medium
dependent parameter, which can be obtained
experimentally for different tissues, and the pixel
displacement, we can estimate the local temper-
ature change.
Renal B-Mode Sonographic Simulation
We simulated abdominal sonography in the renal
tissue region regarding the gaussian speckle archi-
tecture and randomness. Also, we added respira-
tory motion with respect to the diaphragmatic
motion, modeled as a sinusoidal expansion and
contraction of the abdomen with varying frequen-
cies from 14 to 22 per minute, as in the normal
human body. A small part of these images had
very small changes to simulate thermal speckle
shifts, and the amounts of their speckle shifts were
proportional to k, the renal pixel displacement
coefficient, and other equations and coefficients
that were analyzed in the previous sections. The
amounts of the speckle shifts were tuned from 0
to 15 pixels, which were in correlation to the
amounts of thermal changes during the vascular
occlusion and revascularization (from 0C to
20C). The model presumes that the system form-
ing the image has a linear and space invariant
point spread function (PSF). If the biological tissue
is modeled by a continuous distribution of pointscatterers, whose impulse response is T(x,y), then
the radio frequency image I(x,y) can be obtained
by spatial convolution of them:
(8) I(x, y) = PSF(x, y) T(x, y).
By performing a spatial Hilbert transformation,
the final image was obtained as the envelope of
the radio frequency image I(x, y). In particular,
the PSF has been defined in terms of the acoustic
wavelength, , and the axial and lateral resolu-
tion of the imaging system by
(9)
We inserted a transducer frequency of 4 MHz,
sound velocity of 1540 m/s, wavelength of
0.385 mm, spatial frequency of sound of 5.195
cycles/mm, SD along the x- and y-axes of
0.426 mm, SD along the z-axis of 0.213 mm, x-
and y-axis range of 5 to 5 mm, and z-axis range
0 to 5 mm. The pattern was added to a mobile cir-
cular or multichamber cardiac model. A detailed
description of the B-mode ultrasound modeling
was published previously.24
Thermoacoustic Lens Effect
The ultrasound beam traversing the heated
region deviates as the thermal change alters its
velocity. The steep temperature gradient can
worsen this effect, which is usually negligible.
This phenomenon results in a regional ripple dis-
tal to the heated region. There are software and
hardware methods to reduce it; the former,
which is in our interest, is 2D spatial filtering. A
time-varying 2D gaussian filter as described
below will be used for postprocessing the ther-
mal change data:
(10)
where k= K/C, and n is the order of the filter.8
Converting Speckle Displacements to Thermal
Changes
Because the motion of each speckle is an accu-
mulated motion, the vertical motion vector is
derived once. Moreover, because derivationmakes the algorithm noise sensitive, we imple-
mented an order statistic filter of size 20, as
described previously.31,32According to Equations
1 and 2, displacements are converted to a thermal
shift. Finally, a gaussian kernel size of 10 pixels
and variance of 0.5 can smooth the result.
Motion Artifacts in the z-Axis
As we have discussed previously, optical flow has
been able to recover exact motion fields in the 2D
1540 J Ultrasound Med 2009; 28:15351547
Monitoring of Thermal Changes During Revascularization Therapy
Figure 3. Different motions of renal tissue and their extraction.
-
8/14/2019 Optimized Noninvasive Monitoring of Thermal Changes on Digital B-Mode Renal Sonography During Revasculariza
7/13
images even in the presence of large motion arti-
facts in the z-axis (such as cardiac motion detec-
tion in B-mode echocardiographic series).
The motion in the z direction is called an out-
of-plane error. There are 2 techniques to cope with the out-of-plane error: (1) temporal win-
dowing (spatiotemporal windowing) and (2)
confidence measurement.
1. Temporal windowing (spatiotemporal win-
dowing). This technique applies a gaussian kernel
among the consecutive frames (using a weighted
summation of the consecutive frames). This strat-
egy reduces the effect of the out-of-plane error,
especially when the temporal kernel is large.
Moreover, temporal windowing dramatically
decreases the power of noise and increases the sig-
nal to noise ratio, which is a helpful phenomenonin noise-sensitive optical flow techniques.
2. Confidence measurement. The robustness
and sensitivity of the optical flow methods are
not the same among different regions of an
image. Considering a confidence measurement
for the motion estimation increases the accura-
cy, especially for the out-of-plane error regions.
A detailed discussion of confidence measure-
ment was published previously.30
Clinical Evaluation
Any medical system should be implemented on
real tissue before practical application. We selected
3 healthy sheep and then occluded a renal seg-
mental artery in each to observe the heat changes
in that special lobule (each lobule is supplied by
a segmental artery). Theoretically, we should have
had a decrease in the renal lobular temperature
after occluding the segmental artery. The occlusion
was done surgically, and a gradual reopening of the
segmental arterial branch was performed in nearly
2 hours to model anticoagulation reperfusion med-
ications ideally. The animals were anesthetizedwith midazolam and morphine. The incision was
done as a great midline incision. The whole surgery
lasted 1 hour and was finalized with a bolus of
potassium chloride injected intravenously.
Results
B-Mode Sonographic Simulation
The coefficients described in Renal B-Mode
Sonographic Simulation above were used in the
sonographic simulation. Finally a 20-frame
sonographic sequence was obtained, in which 1
point was considered the hot point. The temper-
ature of the hot point varied from 0C to 50C.
The appropriate thermal coefficient and motionartifacts were assigned to the simulation.
Different motion estimation techniques were
implemented and optimized using different
parameters. In this study, the evaluation was per-
formed on simulated images to find out the best
algorithm and optimized coefficients.
Evaluation of Motion Estimation Methods on
the Simulated Data (Finding the Optimized
Method)
According to the algorithm, which was discussed pre-
viously, we synthesized the computerized B-modesonographic movies. The speckle-tracking tech-
niques were evaluated on these simulated images.
All sequences were prefiltered using a gaussian
kernel of size 20 pixels and variance 0.5. For the
Lucas-Kanade method, we experimentally found
that a local window of size of 5 pixels was the best
one. All other parameters were set as described
by Barron et al.35 For the Horn-Schunck method,
an iteration number equal to 300 was found to
be the best. The number was optimized
between 0 to 1000, and 600 was found to be the
best. Gaussian spatial and temporal windows
were considered 15 pixels and 10 frames, respec-
tively. All the applications were performed in
MATLAB 7.1 (The MathWorks, Natick, MA) with a
3-GHz processor (Intel Corporation, Santa Clara,
CA), and 1028 MHz of RAM.
It is a very popular evaluation to implement the
motion estimation techniques on the simulated
sequences.35 Errors were computed with the
widely used amplitude (Equation 11) and angu-
lar errors (Equation 12):
(11)
(12)
where V, Vdefine the real and estimated vectors.33
J Ultrasound Med 2009; 28:15351547 1541
Abolhassani et al
-
8/14/2019 Optimized Noninvasive Monitoring of Thermal Changes on Digital B-Mode Renal Sonography During Revasculariza
8/13
Figures 47 show the results of the implementa-
tion of different motion estimation techniques on
the simulated images with respect to the average
amplitude and angular error. The gradient-based
optical flow methods show better results than theblock-matching methods except the ES method.
According to the images, newer block-matching
techniques have better performance because of
their adaptive behavior with respect to the previ-
ous frame and their better searching strategy in
the diamond search and 3SS. The performance of
older block-matching techniques such as the 3SS
is lower than that of the newer ones. The exhaus-
tive and correlation-based methods are accurate
but more computationally expensive. The corre-
lation algorithm is the most time-consuming
technique, and implementing the gradientdescent cannot dramatically remove this prob-
lem. There was no significant difference between
the MAD and MSE, but the MSE takes more time.
Among the optical flow techniques, the moment-
based approaches act better because of their
invariant hypergeometry. Horn-Schunck outper-
forms the popular Lucas-Kanade due to its higher
robustness to the out-of-plane error. The imple-mentation of the phase matching has a small
effect on increasing the accuracy. Increasing the
optical flow iterations can increase the accuracy
but also increases the time consumption.
Table 1 shows the results of implementation of
the optimized optical flow (multiresolution Horn-
Schunck) and optimized block-matching (ARPS)
techniques for thermal detection. To determine
the thermal changes, motion vectors are derived
once with respect to the vertical axis. Therefore,
the robustness to noise and achieving smooth
results will be very important. Theoretically, theHorn-Schunck energy function has a smoothing
term that achieves smooth and visually attrac-
tive results. Therefore, thermal detection is more
accurate based on the Horn-Schunck algorithm.
Moreover, the multiresolution moment imple-
mentation of the Horn-Schunck technique has
the ability to achieve invariant features and
reduce noise (contrary to their high-pass nature).
Figure 8 shows 2 frames of the synthetic
sequence before and after the thermal shift. Figure
9 is the result of the implementation of the most
accurate algorithm (Horn-Schunck) using the
optimized coefficients on the mentioned frames.
The thermal motion is extracted after subtracting
the interpolated motion from the total motion.
The left image is the total motion vector and is pre-
dominantly made up from the motion artifact.
The right image shows the net thermal shift after
motion artifact removal using the proposed
method.
Figure 10 is the color-coded result of the ther-
mal change in the same frames. The top image
shows the real thermal change (ground truth), which was used in the model, and the bottom
images show the result of the optimized method
in the presence of a motion artifact (left image)
and without a motion artifact (right image).
In Vivo Results Based on the Optimized Method
So far, it has been shown that the multiresolu-
tion Horn-Schunck method is the optimized
technique for simulated ultrasound images.
We implemented this algorithm on the real data
1542 J Ultrasound Med 2009; 28:15351547
Monitoring of Thermal Changes During Revascularization Therapy
Figure 4. Comparison of different block-matching speckle-tracking methods usedon the simulated data.
Figure 5. Comparison of different block-matching speckle-tracking methods usedon the simulated data.
-
8/14/2019 Optimized Noninvasive Monitoring of Thermal Changes on Digital B-Mode Renal Sonography During Revasculariza
9/13
during occlusion and reopening of the renal
artery, only the optimized motion estimation
technique, ie, multiresolution Horn-Schunck,
was applied to the sequences, and the coeffi-
cients were set the same as for the simulatedsequence validation. It is very difficult to quan-
titatively analyze the thermal change on real
animal frames, and we validated the in vivo
results on the basis of the clinical knowledge of
physicians.
The real sequences were obtained using a
probe frequency of 7 MHz and a temporal reso-
lution of 25 frames per second. Three sheep were
anesthetized successfully using midazolam and
morphine. All of the subjects remained alive dur-
ing the whole surgery, and no significant operat-
ing room side effect happened.Figure 11 shows 2 frames of the sequence,
which were taken 0 and 10 minutes after the
reopening. The occluded artery in not visible
in the ultrasound images, but the thermal
change results of the arterial occlusion can be
detected.
1. In the first step, the total motion is extracted
(combined motion artifact and thermal change)
for the ROI and ROI neighborhood using the
optimized method. The ROI (a 200 200 box) is
defined manually for the first frame. The ROI
neighborhood pixels are defined as 100 pixels
around the primary ROI. Figure 12 shows imple-
mentation of the optimized algorithm (mul-
tiresolution Horn-Schunck) on the real
ultrasound images in the ROI and ROI neighbor-
hood pixels for the mentioned frames.
2. In the next step, the interpolated motion vec-
tors of the ROI are computed on the basis of the
motion vectors of the ROI neighborhood pixels.
We implemented cubic Ispline interpolation in
2D followed by mild gaussian smoothing (vari-
ance, 0.1; size, 5 5). Figure 13, left image, showsthe total motion vector for the mentioned frames.
3. To extract the pure thermal change, we sub-
tracted the total ROI motion vectors (Figure 13, left
image) from ROI motion artifact (Figure 13, cen-
ter image). Figure 13, right image, shows the total
motion vector in the ROI region.
4. The vertical thermal motion vectors are
derived with respect to the vertical axis to deter-
mine the net thermal change. Figure 14 shows
the thermal elevation in the ROI.
As proved, the lower renal lobe temperature
increases. It is evident that there is a warm-up
gradient in the temperature in the bottom of the
renal tissue, which correlates well with the
anatomy of the lower renal lobe perfusion. With
respect to the fact that the lower renal segment
was being occluded and reopened in that case,
the lower renal part warm-up is noteworthy after
the reperfusion.
Figure 15 shows the thermal change between 0
and 5, 0 and 10, 0 and 15, and finally 0 and 20minutes in the sonographic series. In the last 2
J Ultrasound Med 2009; 28:15351547 1543
Abolhassani et al
Figure 6. Results of implementation of optical flow techniques on the simulateddata.
Figure 7. Comparison of different block-matching speckle-tracking methods usedon the simulated data.
Table 1. Thermal Detection Results of the Best Motion EstimationMethod of Each Category
Thermal Change Average Variance of ThermalMethod Error, C Change Error, C
ARPS 1.2 2.4Horn-Schunck 1.02 1.9Multiresolution Horn-Schunck 0.50 0.502
-
8/14/2019 Optimized Noninvasive Monitoring of Thermal Changes on Digital B-Mode Renal Sonography During Revasculariza
10/13
color images, even branching of the central lobar
artery is substantial (although sonography can-
not even show the main artery). Note that ther-
mal change estimation cannot extract the
primary temperature but only the thermal
change between 2 frames.
We averaged the temperature change of the ROI
of each frame. Figure 16 shows that the averagedtemperature in the ROI increases after reopening
of the arterial segment (warming section). The
results of the real data were validated by clinical
experts. The plots in Figures 15 and 16 are well
related to the clinical knowledge of the lower seg-
mental renal artery.
Discussion
We have proposed a novel real-time and nonin-
vasive method for thermal detection based on
ultrasound velocity changes in tissue matter.
This technique was used to follow revasculariza-
tion therapy in renal tissue after renal arterial
occlusion. The motion artifacts were carefully
removed. Implementation of the algorithm on
the real and the simulated data showed a good
correlation among them.
1544 J Ultrasound Med 2009; 28:15351547
Monitoring of Thermal Changes During Revascularization Therapy
Figure 8. Simulated thermal change in renal tissue. The boxesshow the thermal detection ROI.
Figure 11. Two frames of the renal sonographic series immedi-ately and 10 minutes after reopening. Note that the renal lower
lobe artery cannot be seen. The boxes show the ROI, and the
line indicates a slight vertical shift.
Figure 9. Left image, Motion vectors of the abdominal sono-graphic model in the ROI of the aforementioned frames. Right
image, Thermal change vectors after subtracting the motion arti-
fact using the optimized method (multiresolution Horn-Schunck).
Figure 10. Top image, Thermal change in the renal tissue.Bottom, right image, Result of implementation of the multireso-
lution Horn-Schunck technique (the most accurate method)
using the optimized coefficients in the presence of a motion arti-
fact. Bottom, left image, Same implementation without a motion
artifact. Obviously, the motion artifact decreases accuracy.
Figure 12. Motion vectors of the ROI and ROI neighborhood pixels,which are a summation of thermal changes and a motion artifact.
-
8/14/2019 Optimized Noninvasive Monitoring of Thermal Changes on Digital B-Mode Renal Sonography During Revasculariza
11/13
Among all of the motion estimation techniques
analyzed, the ARPS was the best block-matching
method, whereas the moment-based multireso-
lution Horn-Schunck technique was the best
gradient-based method. However, the latter out-
performed the former; therefore, we used the
multiresolution Horn-Schunck technique for the
thermal change extraction.The ARPS was the best block-matching method
because of its adaptive strategy, which adjusts
the next motion vectors according to the previ-
ous vectors. The multiresolution approaches
were able to extract different values of thermal
change, overcome the local minima problem,
and achieve more reliable outputs. Fine motions
were preserved using finer moments, and large
and coarse motions were preserved. For the sim-
ulated images, the average thermal error was
0.5C with an SD of 0.5C, although much lower
errors can be obtained in noiseless (motionless)
data. This method can be an efficient technique
for revascularization therapy evaluation and
follow-up based on digital B-mode sonographic
methods. This study was also a good evaluation
of several motion estimation techniques for
speckle tracking.
We are currently working on multigrid meth-
ods to achieve faster techniques. Moreover, we
are interested in implementing this algorithm in
other clinical problems. Many vital changes in
the human body also change the tissue temper-
ature; therefore, thermal detection can be a ref-
erence standard for tracking them. The most
attractive but complex one is detecting the
myocardial temperature during streptokinase-
based coronary revascularization, which is very
difficult because of the great myocardial contrac-
tion and relaxation.
J Ultrasound Med 2009; 28:15351547 1545
Abolhassani et al
Figure 13. Left image, Total motion vectors of the ROI. Centerimage, Interpolated ROI motion vectors using ROI neighborhood
pixels. Right image, Subtraction of the left image from the cen-
ter images, which is the thermal change.
Figure 15. Renal B-mode sonograms and their extracted thermal changes afterarterial reopening in the ROI. The first image (a) was taken 10 minutes after the
revascularization therapy, whereas the others were taken 20 (b), 30 (c), and 40 (d)minutes after the revascularization therapy respectively. There is a gradual ROIwarm-up when the artery is opened, and even branching of the vessel is evident
in the last frame. e, Close-up of the thermal change in d.
Figure 14. Thermal change after derivation of accumulatedthermal displacements in Figure 13, right image. The color bar
shows the thermal change in degrees Celsius.
Figure 16. Renal lower segment thermal elevation after reperfusion.The x-axis is the time in minutes, and the y-axis is the temperature
of the tissue. Notice the close correlation with the clinical situation.
Author: In Figure 14 legend, please verify Figure 13, right
image, as changed from Figure 9c.
-
8/14/2019 Optimized Noninvasive Monitoring of Thermal Changes on Digital B-Mode Renal Sonography During Revasculariza
12/13
References
1. Kasper DL, Braunwald E, Fauci AS, Hauser S, Longo D,
Jameson JL. Principles of Internal Medicine. 16th ed. New
York, NY: McGraw-Hill; 2005.
2. Abolhassani MD, Norouzi A, Takavar A, Ghanaati H.
Noninvasive temperature estimation using sonographic
digital images. J Ultrasound Med 2007; 26:215222.
3. Gertner MR, Worthington AE, Wilson BC, Sherar MD.
Ultrasound imaging of thermal therapy in in vitro liver.
Ultrasound Med Biol 1998; 24:10231032.
4. Adam D, Raukhverger R. Real-time monitoring hyperther-
mia therapy by ultrasound imaging. Paper presented at:
Mediterranean Conference on Medical and Biological
Engineering; July 31August 5, 2004; Ischia, Italy.
5. Ueno S, Hashimoto M, Fukukita H, Yano T. Ultrasound
thermometry in hyperthermia. IEEE Ultrason Symp Proc
1990; 3:16451652.
6. Varghese T, Daniels MJ. Real-time calibration of tempera-
ture estimates during radiofrequency ablation. Ultrason
Imaging 2004; 26:185200.
7. Varghese T, Zagzebski JA, Chen Q, et al. Ultrasound moni-
toring of temperature changes during radio frequency
ablation: preliminary in-vivo results. Ultrasound Med Biol
2002; 28:321329.
8. Simon C, Vanbaren P, Ebbini ES. Two-dimensional temper-
ature estimation using diagnostic ultrasound. IEEE Trans
Ultrason Ferroelectr Freq Control 1998; 45:10881099.
9. Seip R, Vanbaren P, Cain A. Noninvasive real-time multi-
point temperature control for ultrasound phased array
treatments. IEEE Trans Ultrason Ferroelectr Freq Control1996; 43:10631073.
10. Straube WL, Arthur RM. Theoretical estimation of the tem-
perature dependence of backscattered ultrasonic power
for noninvasive thermometry. Ultrasound Med Biol 1994;
20:915922.
11. Karjalainen T, Thierman JS, Hynynen K. Ultrasound acous-
tic stimulated emission for controlling thermal surgery. IEEE
Ultrason Symp Proc 1999; 2:13971400.
12. Konofagou E, Thierman J, Hynynen K. The use of ultra-
sound-stimulated acoustic emission in the monitoring of
modulus changes with temperature. Ultrasonics 2003;
41:337345.
13. Konofagou E, Thierman J, Karjalainen T, Hynynen K. The
temperature dependence of ultrasound-stimulated acous-
tic emission. Ultrasound Med Biol 2002; 28:331338.
14. Arthur RM, Straube WL, Trobaugh JW, Moros EG. Non-
invasive estimation of hyperthermia temperatures with
ultrasound. Int J Hyperthermia 2005; 21:589600.
15. Furht B, Greenberg J, Westwater R . Motion Estimation
Algorithms for Video Compression. Waltham, MA: Kluwer
Academic Publishers; 1997.
16. Li R, Zeng B, Liou ML. A new three-step search algorithm
for block motion estimation. IEEE Trans Circuits Syst Video
Technol 1994; 4:438442.
17. Lu J, Liou ML. A simple and efficient search algorithm for
block-matching motion estimation, IEEE Trans Circuits Syst
Video Technol 1997; 2:429433.
18. Po L, Ma WC. A novel four-step search algorithm for fast
block motion estimation, IEEE Trans Circuits Syst Video
Technol 1996; 6:313317.
19. Zhu S, Ma KK. A new diamond search algorithm for fast
block-matching motion estimation. IEEE Trans Image Process
2000; 9:287290.
20. Nie Y, Ma KK. Adaptive rood pattern search for fast block-
matching motion estimation. IEEE Trans Image Process
2002; 11:14421448.
21. Cheung C, Po L. A novel small cross-diamond search algo-
rithm for fast video coding and video conferencing appli-
cations. In: Proceedings of the IEEE International
Conference on Image Processing. Piscataway, NJ: Institute
of Electrical and Electronics Engineers; 2002.
22. Cheung C, Po L. A novel cross-diamond search algorithmfor fast block motion estimation. IEEE Trans Circuits Syst
Video Technol 2002; 12:11681177.
23. Horn BKP, Schunck BG. Determining optical flow. Artif Intel
1981; 17:185283.
24. Baraldi P, Sarti A, Lamberti C, Prandini A, Sgallari F,
Evaluation of differential optical flow techniques on syn-
thesized echo images. IEEE Trans Biomed Eng 1996;
43:259272.
25. Lucas BD, Kanade T. An iterative image registration tech-
nique with an application to stereo vision. In: Proceedings
of the International Joint Conference on Artificial
Intelligence. San Francisco, CA: International Joint
Conferences on Artificial Intelligence; 1981:674679.26. Sahba N, Tavakoli V, Ahmadian A, Abolhassani MD,
Fotouhi M. Hybrid local/global optical flow and spline
multi-resolution analysis of myocardial motion in B-mode
echocardiography images. Tech Acoust 2008. http://
www.ejta.org.
27. Tavakoli V, Sahba N, Ahmadian A, Amini AA. Adaptive
Multi-resolution myocardial motion estimation in B-mode
echocardiography images. In: Proceedings of the IEEE
International Conference on Bioinformatics and Biomedical
Engineering. Piscataway, NJ: Institute of Electrical and
Electronics Engineers; 2008:20332036.
28. Tavakoli V, Sahba N, et al. An evaluation of different optical
flow techniques for myocardial motion analysis. In:
Proceedings of the IEEE International Conference on
Bioinformatics and Biomedical Engineering. Piscataway, NJ:
Institute of Electrical and Electronics Engineers; 2008:2528.
29. Tavakoli V, Sahba N. A new variational technique for com-
bining affine registration and optical flow in echocardiog-
raphy images. In: Proceedings of the 30th Annual
International IEEE Engineering in Medicine and Biology
Society Conference. Piscataway, NJ: Institute of Electrical
and Electronics Engineers; 2008:205208.
30. Shling M, Arigovindan M, Jansen C, Hunziker P, Unser M,
Myocardial motion analysis from B-mode echocardio-
grams, IEEE Trans Image Process 2005; 14:525536.
1546 J Ultrasound Med 2009; 28:15351547
Author: Please verify volume numbers in reference 5, 11, and 1618. Please give page
numbers for reference 21. Please verify reference 23; unable to find in our sources.
-
8/14/2019 Optimized Noninvasive Monitoring of Thermal Changes on Digital B-Mode Renal Sonography During Revasculariza
13/13
31. Mehrabani BM, Tavakoli V, Abolhassani MD, et al. An effi-
cient temperature estimation using optical-flow in ultra-
sound B-mode digital images, In: Proceedings of the 30th
Annual International IEEE Engineering in Medicine and
Biology Society Conference. Piscataway, NJ: Institute of
Electrical and Electronics Engineers; 2008:8689.
32. Sahba N, Tavakoli V, Nambakhsh MS. Ultrasound thermal
change detection based on steerable filters. In:
Proceedings of the 30th Annual International IEEE
Engineering in Medicine and Biology Society Conference.
Piscataway, NJ: Institute of Electrical and Electronics
Engineers; 2008:29842987.
33. Tavakoli V, Sahba N, Hajebi N, et al. A fast and accurate
method for echocardiography strain rate imaging. In:
Proceedings of the Annual Meeting of the International
Society for Optical Engineering. Bellingham, WA:
International Society for Optical Engineering; 2009:7259
72592.
34. Tavakoli V, Sahba N, et al. Two-dimensional echocardiogra-
phy motion analysis based on optical flow techniques.
Biomed Imaging Interv J. In press.
35. Barron JL, Fleet DJ, Beauchemin S, Performance of optical
flow techniques. Int J Comput Vis 1994; 12:4377.
J Ultrasound Med 2009; 28:15351547 1547
Abolhassani et al
Author: Please give all authors last names and initials for references 28,
31, and 33. Please give correct page numbers for reference 33. Also,
please give all authors last names and initials for reference 34 and
update all information if now available.