Sunteți pe pagina 1din 10

Ultrasonics 44 (2006) 99–108

www.elsevier.com/locate/ultras

Comparison of shift estimation strategies in spectral elastography


a,b a,* c
Kenneth Hoyt , Flemming Forsberg , Jonathan Ophir
a
Department of Radiology, Thomas Jefferson University, Philadelphia, PA 19107, USA
b
Department of Electrical Engineering, University of Rochester, Rochester, NY 14627, USA
c
Department of Diagnostic and Interventional Imaging, University of Texas Medical School, Houston, TX 77030, USA

Received 25 April 2005; received in revised form 18 August 2005; accepted 19 August 2005
Available online 28 September 2005

Abstract

This paper compares the performance of various spectral shift estimators for use in spectral elastography, namely, the normalized
cross-correlation (NCC), sum squared difference (SSD) and sum absolute difference (SAD). Simulation and experimental results dem-
onstrate that the spectral SSD-based elastographic method exhibits no marked difference in performance compared to the more compu-
tationally costly NCC-based approach, which has conventionally been the preferred estimator in spectral elastography. The spectral
SAD-based strain estimator, despite being computationally less burdening, failed to exhibit performance comparable to that of the
NCC- and SSD-based techniques. Furthermore, though spectral subsample estimation techniques using a cosine-fit interpolation method
outperformed that of the parabolic-fit method in terms of both reduced bias errors and standard deviations, the latter was analyzed in
this study due to computational simplicity. The role of spectral density was evaluated without and with parabolic-based subsample inter-
polation. Based on minimizing computational complexity, it is concluded that a (low density) spectral SSD strain estimator coupled with
parabolic-based subsample estimation is the preferred choice for spectral elastography.
 2005 Elsevier B.V. All rights reserved.

Keywords: Elasticity imaging; Spectral elastography; Strain Filter; Subsample estimation; Ultrasound imaging

1. Introduction Conventional elastographic imaging is a time-domain


approach which estimates axial tissue displacement from
Throughout the last fifteen years, elastography has the cross-correlation function applied to congruent pre-
emerged as an effective imaging modality for depicting and post-compression gated RF echo signals [1]. By per-
local strain distributions in soft tissue [1]. Interpreted as a forming this estimation technique to overlapping windows
surrogate for the true tissue modulus distribution, the and, subsequently, to all A-lines in the image plane, an
fundamental premise behind elastographic imaging is that axial tissue displacement field is obtained. The gradient
local changes in soft tissue stiffness are generally correlated of this displacement field results in a strain image (termed
to pathological changes. Thus, early detection of these an elastogram). Despite exhibiting a high sensitivity for
changes may be important from a clinical perspective in extracting tissue strain information [6], conventional elas-
order to improve patient prognosis. Since its introduction tographic techniques have been shown to be susceptible to
to the research community in 1991 by Ophir et al. [1], elas- decorrelation noise sources that manifest by degrading
tography has found utility in many clinical applications the tissue displacement estimation process [7]. Though
ranging from (but not limited to) cancer imaging in breast strain estimation enhancement techniques such as global
[2] and prostate [3,4] to the intravascular detection and stretching [8], multicompression averaging [9] and local
characterization of vulnerable plaques [5]. adaptive processing algorithms [10,11] may reduce this
degradation, the effects of decorrelation noise still can
*
Corresponding author. Tel.: +1 215 955 4870; fax: +1 215 955 8549. remain a source of image (information) content ambigu-
E-mail address: flemming.forsberg@jefferson.edu (F. Forsberg). ity, which should be further minimized if elastography is

0041-624X/$ - see front matter  2005 Elsevier B.V. All rights reserved.
doi:10.1016/j.ultras.2005.08.006
100 K. Hoyt et al. / Ultrasonics 44 (2006) 99–108

to find acceptance and widespread use in the clinical the amplitude of Eq. (1) is taken as the spectral shift esti-
setting. mate and is related to tissue strain s as follows [15]:
Spectral elastography was introduced as an alternative k
to the conventional elastographic technique and functions s¼ ð2Þ
k0
by estimating local tissue strain directly from the relative
shift between sets of pre- and post-compression power where k0 is the center frequency of the pre-compression
spectra [12,13] or from the spectral scaling factor [14]. scattering function.
More importantly, spectral elastography has been reported In order to compute tissue strain as described by Eq. (2),
to be more robust to decorrelation processes such as axial the spectral estimates must be derived using windowed RF
jitter and out-of-beam tissue motion than conventional echo signals. Furthermore, to ensure that strain analysis
elastography [12,15]. It has also been shown that by utiliz- was applied to windowed pre- and post-compression sig-
ing adaptive spectral strain estimation techniques, marked nals composed of the same echo information, the latter seg-
improvements in estimator performance may be obtained ments were shifted and scaled as a function of the applied
when compared to temporal elastographic methods com- strain prior to performing spectral estimation [12,14,15].
pounded with strain enhancement techniques, namely, Note that this post-compression data windowing operation
local adaptive processing [16]. Despite these promising re- is regarded as an indirect approach to incorporate signal
sults in terms of spectral elastographic performance, com- phase information into the strain estimation process and,
putational complexity still constitutes a source of concern consequently, allows for a more accurate description of
for real-time applications. strain information when investigating tissues subjected to
An initial set of reports detailing spectral elastography small applied compressions. The frequency content of each
utilized a high spectral density (i.e., Fourier transform data window was estimated using an N-length fast Fourier
length) in order to provide sufficient data samples as input transform (FFT) computation whereas the power spectrum
to the discrete cross-correlation function and for the subse- for these same data windows was taken as the magnitude
quent spectral shift estimation [12,13]. Our group later squared of the FFT output normalized by the length of
showed that by utilizing curve-fitting interpolation about the spectral window.
the spectral cross-correlation peak (in order to estimate Simulations were implemented in order to compare the
the subsample spectral shift) that the spectral density performance of the NCC to that of two other less compu-
(and computational complexity) may be significantly re- tationally burdensome pattern matching functions, namely,
duced while sustaining the elastographic signal-to-noise the sum squared difference (SSD)
ratio (SNRe) [15]. A concern of this latter report is that
X
N =2
parabolic interpolation was employed as the curve-fitting RSSD ðkÞ ¼ ðS 0 ðmÞ  S 1 ðm þ kÞÞ
2
ð3Þ
method, and despite being computationally simple, it has m¼1
been shown to exhibit subsample dependent bias and vari-
and the sum absolute difference (SAD)
ability in time-domain applications [17].
Overall, the motivation for this research project was to X
N =2

investigate spectral elastography from a computational RSAD ðkÞ ¼ jS 0 ðmÞ  S 1 ðm þ kÞj ð4Þ
m¼1
complexity perspective and to evaluate potential means
for its minimization while maintaining spectral elasto- where the argument that minimizes the amplitude of Eq.
graphic performance. (3) or (4) is taken as the spectral shift estimate and related
to tissue strain via Eq. (2). The selection of the SSD and
2. Simulation methods and results SAD methods for evaluation in this study was premised
on previous reports which concluded superior performance
2.1. Spectral shift estimators compared to other time-delay estimators [18]. Note the
overall reduction in computational complexity when mov-
The shift in spectral elastography has typically been esti- ing from Eq. (1) to (3) to (4).
mated using the normalized cross-correlation function An elastographic simulation program was developed
(NCC), which is defined as using Matlab (Mathworks, Inc., Natick, MA, USA). A
PN =2 1-D model was implemented that assumes Gaussian mod-
m¼1 S 0 ðmÞ  S 1 ðm þ kÞ ulated sine and cosine pulses (7.5 MHz center frequency
RNCC ðkÞ ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
PN =2 PN =2 ffi ð1Þ
S ðmÞ S ðmÞ with a 60% fractional bandwidth) for the impulse response
m¼1 0 m¼1 1
of the ultrasound system (or point spread function; PSF)
where S0 and S1 denote the pre- and post-compressed and tissue scattering function, respectively. Additionally,
power spectra, respectively, and N represents the spectral the speed of sound was assumed to be 1540 m/s. In general,
window length (typically described as the Fourier trans- a pre-compression RF echo A-line sequence (40 mm in
form length or spectral density). Note the upper limits of length and sampled at 60 MHz) was created by convolving
summation, which are equal to the folding frequency of the PSF with the scattering function. The post-compression
the power spectra or N/2. The argument k that maximizes echo sequence was subsequently created by linearly com-
K. Hoyt et al. / Ultrasonics 44 (2006) 99–108 101

Table 1 (and corresponding SNRe) described by this same range


Default values used to generate the reference and delay spectral signals denotes sensitivity.
Parameter Value The mean strain estimates are illustrated in Fig. 1 for the
Center frequency (fc) 7.5 MHz various spectral shift estimation methods. Fig. 1a demon-
Fractional bandwidth (B) 60% strates that all estimators follow a linear trend for applied
Sampling frequency (fs) 60 MHz strains less than 10%. Inspection of expanded results for an
Signal-to-noise ratio (SNR) 30 dB
Temporal window length (L) 3.0 mm
applied strain range from 0.5% to 5% (Fig. 1b) indicates
Spectral window length (N) 1024 that the SAD spectra strain estimator tends to slightly
underestimate the true (applied) strain value, which is accu-
rately estimated using both the spectral NCC and SSD
pressing the scattering distribution by a specified strain elastographic methods.
level and then convolving with the PSF in the same manner The strain estimate variability (i.e., standard deviation
as done for the pre-compressed signal. For the simulation of the strain estimates) for the spectral NCC, SSD and
results presented in this section, strain estimates were ob- SAD strain estimators are illustrated in Fig. 2. Inspection
tained from 20 strain levels (dynamic range of 0.1–50%) of Fig. 2a indicates that for applied strain greater than
and averaged over 50 independent realizations. The default 5%, an abrupt increase in strain estimate standard devia-
values used for the simulations described in this paper are tion ensues for all estimators evaluated. Additionally, there
summarized in Table 1 unless otherwise noted. All statisti- is no discernible discrimination between the NCC and SSD
cal results introduced were obtained using a two-sample methods. Fig. 2b demonstrates that both the spectral NCC
t-test (assume unequal variances) where a p-value less than and SSD strain estimators exhibit the same estimation var-
0.05 was considered statistically significant. iance while being markedly lower than that demonstrated
Strain estimator performance was quantified using the by the SAD method.
SNRe which is defined for a homogenous strain region as Simulation Strain Filters are illustrated in Fig. 3. The
spectral SAD strain estimation method exhibits a lower
l
SNRe ¼ s ð5Þ SNRe (most noticeable for applied strains between 1.0%
rs and 5.0%) when compared to the Strain Filter responses
where ls denotes the mean strain estimates (signal) and rs derived using the NCC- and SSD-based methods (albeit
the standard deviation of the strain estimates (or strain esti- not statistically significant, p = 0.73, when considering the
mate noise). Additionally, simulation results were obtained entire Strain Filter response), whereas comparison of the
depicting the range of strains from the compressed tissue as latter two approaches (i.e., Strain Filter responses) indi-
a function of the SNRe. Termed the Strain Filter [6], it can cates analogous performance (p > 0.99). Overall, these re-
be regarded as a measure of estimator performance, which sults indicate that the spectral NCC and SSD strain
allows investigation into the tradeoffs between various estimation techniques perform virtually identically. Addi-
strain processing technique and parameters. Specifically, tionally, inspection of Eqs. (1) and (3) for the respective
the width of the Strain Filter for a given SNRe is an indi- pattern matching functions indicates that a reduction in
cator of the strain dynamic range over which the SNRe ex- computation load is achieved when employing the SSD
ceeds a particular value; whereas the lowest strain value method over that of the NCC due to the fact the former

2
10
NCC
SSD
SAD
Mean Strain Estimate µs

1
Mean Strain Estimate µs

10
0
10

0
10

-1 -1
10 10
-1 0 1 2 -1 0
10 10 10 10 10 10
(a) Applied Strain (%) (b) Applied Strain (%)

Fig. 1. Comparison of spectral strain estimators for the measured strain versus the true applied strain for the NCC, SSD and SAD spectral shift
estimation methods. Graph (b) depicts an expanded view of (a). The power spectra were computed using a 4096 FFT and 3.0 mm temporal windows with
80% overlap.
102 K. Hoyt et al. / Ultrasonics 44 (2006) 99–108

-3
x 10
0.05 5
NCC
Strain Estimate Standard Deviation σs

Strain Estimate Standard Deviation σs


SSD
0.04 SAD 4

0.03 3

0.02 2

0.01 1

0 0
-1 0 1 2 -1 0
10 10 10 10 10 10
(a) Applied Strain (%) (b) Applied Strain (%)

Fig. 2. Comparison of strain estimator performances (strain estimate variability) for the NCC, SSD and SAD spectral shift estimation methods. Graph (b)
depicts an expanded view of (a). The power spectra were computed using a 4096 FFT and 3.0 mm temporal windows with 80% overlap.

25 the focus of the comparison made in this paper will be


NCC on the parabolic and cosine curving fitting methods due
SSD to their relative computational simplicity. Specifically, the
20 SAD
subsample spectral shift ok is computed using parabolic-
fit interpolation about the peak of the spectral pattern
15
matching curve R(0) as follows:
SNRe

10
Rð1Þ  Rð1Þ
ok ¼ ð6Þ
2ðRð1Þ  2Rð0Þ þ Rð1ÞÞ
5
where R(1) and R(1) denote the two nearest-neighbor
samples. Conversely, the subsample spectral shift com-
0
-1 0 1 2
puted cosine interpolation is given by the following
10 10 10 10
expression:
Applied Strain (%)
b
Fig. 3. Illustration of simulated Strain Filters for the spectral NCC, SSD ok ¼  ð7Þ
and SAD strain estimators investigated. The power spectra were computed
a
using a 4096 FFT and 3.0 mm temporal windows with 80% overlap. The where
NCC and SSD exhibit nearly identical performance and both methods  
outperform the SAD method in terms of sensitivity, dynamic range and Rð1Þ þ Rð1Þ
a ¼ arccos
SNRe. 2Rð0Þ
  ð8Þ
Rð1Þ  Rð1Þ
b ¼ arctan
2Rð0Þ sin a
does not require normalization by the variances of the
respective signals. Note that the cosine interpolation method is more com-
putationally demanding than the parabolic method due to
2.2. Subsample estimation the geometric functions.
A subsample simulation program was developed using
Most shifts encountered in spectral elastography are not Matlab (Mathworks, Inc., Natick, MA, USA) in order to
exact integer values and thus require some form of interpo- evaluate the performance (in terms of subsample bias error
lation strategy in order to increase the density of the sam- and standard deviation) of the spectral SSD shift estimator
ples. Though increasing the Fourier transform length will coupled with the curve-fitting subsample estimation tech-
improve the accuracy of the shift estimator, this requires nique described by Eqs. (6) and (7). A set of 250 synthetic
a large N-length Fourier transform and thus is computa- RF echo base signals were generated by convolving the
tionally burdensome since it essentially provides interpo- PSF (for a given set of system parameters) with the scatter-
lated samples throughout the entire spectrum. Typically, ing function (as described in the previous section). Power
subsample estimation in time-domain applications is per- spectral estimates for these base signals were computed as
formed via some curve-fitting method about the pattern the magnitude squared of an N-length fast Fourier trans-
matching function. Though several methods have evolved form (FFT). Analogous to the method described by Viola
over the years and have appeared in the literature [17,19], and Walker [17], referenced and delayed spectral signals
K. Hoyt et al. / Ultrasonics 44 (2006) 99–108 103

were formed by decimating the base signals at 100 times the standard deviation when compared to the parabolic-
desired final sampling rate, then the base signals were sub- based method [19]. Since these marginal improvements
sequently decimated starting at different samples to pro- in performance are at the expense of an increase in
duce reference and delayed signals with known subsample computational complexity, the parabolic-fit interpolated
spectral delay. Subsample delays were varied from 0 to method is employed for the remainder of this section,
0.95 samples in increments of 0.05 samples. Finally, the which is dedicated to a detailed analysis into the relative
mean spectral subsample bias error and standard devia- performance of this spectral subsample estimator (using
tions were computed as follows: simulation).
The role of data sampling frequency on subsample esti-
b ¼ 1 X 250
b i  DkÞ mation performance is illustrated in Fig. 5. As the sampling
bð DkÞ ð Dk ð9Þ
250 i¼1 rate is increased, subsample estimation performance deteri-
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
!2ffi orates in terms of increased bias errors and subsample shift
u
u 1 X 250
1 X250
variability. This is opposite to what has been reported for
b ¼t
rð DkÞ Dk i  b j
Dk ð10Þ
250 i¼1 250 j¼1 subsample estimation in the time-domain [17,19]. However,
it should be remembered that ultrasound signals described
where Dk and Dkb denote the true and estimated spectral de- in the frequency domain are bandlimited. Thus, given a
lays, respectively. fixed FFT length (i.e., spectral window), the impact of
The results presented in Fig. 4 compare the bias errors increasing the sampling rate is that the signal spectrum as-
(Fig. 4a) and standard deviations (Fig. 4b) of the sub- sumes a smaller proportion of the overall spectrum length.
sample interpolated estimates obtained using either the This effectively decreases the SNR and leads to a decrease
parabolic or cosine curve-fitting techniques. The cosine- in subsample estimation accuracy. Note this analysis as-
based curve-fitting method exhibits moderate perfor- sumes the sampling frequency is not lowered to the extent
mance improvements in terms of lower bias error and so as to produce aliasing effects.

-3
x 10
0.02 1
Parabolic Fit Parabolic Fit
Standard Deviation (Samples)

Cosine Fit Cosine Fit


0.8
Bias Error (Samples)

0.01

0.6
0
0.4

-0.01
0.2

-0.02 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
(a) Subsample Delay (Samples) (b) Subsample Delay (Samples)

Fig. 4. Comparison of parabolic and cosine curve-fitting techniques. Performance is illustrated for both subsample (a) bias error and (b) standard
deviation.

-3
x 10
0.03 2
40 MHz 40 MHz
Standard Deviation (Samples)

0.02 60 MHz 60 MHz


Bias Error (Samples)

80 MHz 1.5 80 MHz


0.01

0 1

-0.01
0.5
-0.02

-0.03 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
(a) Subsample Delay (Samples) (b) Subsample Delay (Samples)

Fig. 5. Comparison of the (a) bias error and (b) standard deviation of subsample spectral estimates as a function of the data sampling frequency.
104 K. Hoyt et al. / Ultrasonics 44 (2006) 99–108

-3
x 10
0.03 2
2.0 mm 2.0 mm

Standard Deviation (Samples)


0.02 3.0 mm 3.0 mm
Bias Error (Samples)

4.0 mm 1.5 4.0 mm


0.01

0 1

-0.01
0.5
-0.02

-0.03 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
(a) Subsample Delay (Samples) (b) Subsample Delay (Samples)

Fig. 6. Comparison of the (a) bias error and (b) standard deviation of subsample spectral estimates as a function of the spatial window length.

-3
x 10
0.03 2
1024 1024
0.02 2048 Standard Deviation (Samples) 2048
Bias Error (Samples)

4096 1.5 4096


0.01

0 1

-0.01
0.5
-0.02

-0.03 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
(a) Subsample Delay (Samples) (b) Subsample Delay (Samples)

Fig. 7. Comparison of the (a) bias error and (b) standard deviation of subsample spectral estimates as a function of the Fourier transform length.

(a) (b) (c)


30 30 30
NCC
25 SSD 25 25
SAD
20 20 20
SNRe

SNRe
SNRe

15 15 15

10 10 10

5 5 5

0 0 0
-1 0 1 2 -1 0 1 2 -1 0 1 2
10 10 10 10 10 10 10 10 10 10 10 10
Applied Strain (%) Applied Strain (%) Applied Strain (%)

30 30 30

25 25 25

20 20 20
SNRe

SNRe

SNRe

15 15 15

10 10 10

5 5 5

0 0 0
-1 0 1 2 -1 0 1 2 -1 0 1 2
10 10 10 10 10 10 10 10 10 10 10 10
Applied Strain (%) Applied Strain (%) Applied Strain (%)

Fig. 8. Simulation Strain Filters depicting spectral densities of (a) 1024, (b) 2048 and (c) 4096 without and with parabolic-fit interpolation for spectral
subsample shift estimation, top and bottom rows, respectively.
K. Hoyt et al. / Ultrasonics 44 (2006) 99–108 105

The function of spatial window length on subsample maintaining the overall SNR (in contrast to what was ob-
estimation accuracy is illustrated in Fig. 6. Though larger served when increasing the sampling frequency).
spatial windows are known to improve the resolution of
the spectral estimates, the results of Fig. 6 indicate that this 2.3. Spectral elastographic analysis
increase in spectral resolution is partially offset by an in-
crease in the subsample bias errors and standard devia- Using the simulation program described in Section 2.1,
tions. For a fixed set of parameters, increasing the Strain Filters were derived using the spectral NCC-, SSD-
spectral resolution in ultrasound signals/images corre- and SAD-based strain estimation techniques in order
sponds to an increase in fidelity (given a peaky spectrum to analyze the tradeoffs between spectral density with-
that is typical in ultrasound backscattered echo signals), out and with subsample estimation using a parabolic-fit
which can decrease the spectral correlation for a given
lag value. This decrease in spectral correlation results in a
degradation of the subsample estimator performance,
which is analogous to that observed in subsample time-de- B
lay estimators [17].
The results of Fig. 7 illustrate the effects of spectral den-
sity on subsample estimation performance. As these results I
demonstrate, increasing the Fourier transform length (and
corresponding spectral window size) increases the perfor-
mance of the subsample estimator in terms of minimizing
both the bias errors and standard deviation of the subsam-
ple estimates. This is as expected since increasing the spec- Fig. 9. Illustration of a typical inhomogeneous phantom image and the
tral density ultimately results in an increase in the corresponding background (B) and inclusion (I) regions employed for the
correlation between the reference and delayed spectra while CNRe analysis.

Fig. 10. Elastographic results obtained using an isoechoic phantom. Elastograms were derived using the spectral (a) NCC, (b) SSD and (c) SAD strain
estimators and for 0.5%, 1.0% and 5.0% applied strains, top to bottom, respectively. All elastograms were processed using spatial window lengths of 3 mm
with 80% overlap, a spectral density of 1024 and no spectral subsample shift estimation.
106 K. Hoyt et al. / Ultrasonics 44 (2006) 99–108

interpolated curve (Fig. 8). The results depicted in the top quick-release holder that was mounted to the distal end
row of Fig. 8 demonstrate the overall improvement in SNRe of a mechanized arm. This set-up allowed 3-D motion of
as the spectral density is increased without the use of sub- transducer and application of precise axial compressions
sample estimation (albeit not statistically significant, controlled through a laptop computer and was physically
p > 0.29). Likewise, the results found in the bottom row attached to a portable rack-based system for stability. All
of Fig. 8 show the same increase in spectral density but com- collected RF echo data were digitized at 30 MHz and then
pounded with the use of subsample estimation. These re- upsampled to 60 MHz prior to processing.
sults show that there is no significant difference between Spectral strain estimates were computed from windowed
the Strain Filter responses obtained using the various spec- pre- and post-compression (following applied strains of
tral densities and parabolic-fit interpolation in order to im- 0.5%, 1.0% and 5.0%) RF A-line pairs using the approach
prove spectral shift estimation accuracy (p > 0.92). described in Section 2.1. Following the processing of all A-
Collectively, the results indicate that when a subsample esti- lines in the image plane, a strain image was displayed
mation method is not utilized, a high spectral density is (termed an elastogram). All elastograms were post-pro-
needed in order to maximize strain estimator performance. cessed using a 5-by-5 median spatial filter in order to sup-
Furthermore, an additional enhancement in performance press any extreme strain estimates.
(in terms of SNRe) can be achieved by employing spectral A gelatin-based heterogeneous phantom (10 · 10 ·
subsample shift estimation even at low spectral densities. 10 cm3) was used to evaluate the spectral shift-based strain
estimation techniques described in Section 2.1. The phan-
3. Experimental methods and results tom contained a 0.5 cm (in radius) isoechoic stiff cylindrical
inclusion that was approximately three times stiffer than
A modified HDI-1000 scanner (Philips Medical Systems, the surrounding homogeneous material and was produced
Bothell, WA, USA) allowing access to RF data was used using a previously described technique [20]. Estimator
with a 7.5 MHz (60% fractional BW) linear array trans- performance was quantified using the CNRe as follows
ducer. The transducer was secured and stabilized using a [21]:

Fig. 11. Elastographic results obtained using an isoechoic phantom. Elastograms were derived using the spectral (a) NCC, (b) SSD and (c) SAD strain
estimators and for 0.5%, 1.0% and 5.0% applied strains, top to bottom, respectively. All elastograms were processed using spatial window lengths of 3 mm
with 80% overlap, a spectral density of 1024 and parabolic-fit spectral subsample shift estimation.
K. Hoyt et al. / Ultrasonics 44 (2006) 99–108 107

Table 2 graphic method, though computationally the least burden-


Corresponding elastographic contrast-to-noise ratios (CNRe) (dB) for the some method evaluated, failed to exhibit comparable
images presented in Fig. 11
performance to that of the NCC- and SSD-based tech-
Applied strain (%) Spectral shift estimator niques. Based on minimizing computational complexity,
NCC SSD SAD it is concluded that a (low density) spectral SSD strain esti-
0.5 15.5 15.5 13.8 mator coupled with parabolic-based subsample estimation
1.0 17.9 17.9 17.2 is the preferred choice for spectral elastography.
5.0 16.4 16.0 16.1

Acknowledgement
2
2ðlB  lI Þ This work was supported by the National Institutes of
CNRe ¼ ð11Þ
r2B þ r2I Health (USA) Program Project Grant P01-CA64597.
where l denotes the mean strain estimates, the variance of
References
the strain estimates, and the subscripts I and B indicate
homogenous image regions of the inclusion and back- [1] J. Ophir, E.I. Céspedes, H. Ponnekanti, Y. Yazdi, X. Li, Elastogra-
ground, respectively. All CNRe values were obtained from phy: a quantitative method for imaging the elasticity of biological
phantom regions corresponding to those illustrated in tissues, Ultrasonic Imaging 13 (1991) 111.
Fig. 9 where the regions of interest were positioned to [2] B.S. Garra, E.I. Céspedes, J. Ophir, S.R. Spratt, R.A. Zuurbier, C.M.
encompass the same A-lines and the effects of attenuation Magnant, M.F. Pennanen, Elastography of breast lesions: initial
clinical results, Radiology 202 (1997) 79.
were assumed negligible. [3] A. Pesavento, A. Lorenz, Real time imaging and in vivo applications
Elastograms were processed without (Fig. 10) and with in prostate cancer, Proc. IEEE Ultrason. Symp. 2 (2001) 1647.
(Fig. 11) employing spectral subsample estimation (i.e., [4] R. Souchon, O. Rouviere, A. Gelet, V. Detti, S. Srinivasan, J. Ophir,
parabolic-fit interpolation). Notice that without using sub- J.Y. Chapelon, Visualisation of HIFU lesions using elastography of
sample estimation, the strain estimates are limited to dis- human prostate in vivo: preliminary results, Ultrasound Med. Biol. 29
(2003) 1007.
crete values defined by the spectral density. Though these [5] J.A. Schaar, C.L. de Korte, F. Mastik, S. Chaylendra, G. Pasterk-
quantization-based strain estimation errors become less amp, E. Boersma, P.W. Serruys, A.F.W. van der Steen, Character-
pronounced for higher applied strain owing to in part the izing vulnerable plaque features with intravascular elastography,
improved SNRe (see Fig. 8a), image quality is still discern- Circulation 108 (2003) 2636.
ibly compromised (see Fig. 10). Conversely, when para- [6] T. Varghese, J. Ophir, A theoretical framework for performance
characterization of elastography: the Strain Filter, IEEE Trans.
bolic-fit interpolation is utilized in order to provide Ultrason. Ferroelec. Freq. Control 44 (1997) 164.
subsample spectral shift estimation, no such degradation [7] F. Kallel, J. Ophir, Three-dimensional tissue motion and image noise
is observed in the elastograms (see Fig. 11) owing to a in elastography, IEEE Trans. Ultrason. Ferroelec. Freq. Control 44
decrease in strain estimation noise for the applied strain (1997) 1286.
values investigated (see Fig. 8a). The CNRe values corre- [8] E.I. Céspedes, J. Ophir, Reduction of image noise in elastography,
Ultrasonic Imaging 15 (1993) 89.
sponding to the elastograms presented in Fig. 11 are sum- [9] T. Varghese, J. Ophir, E.I. Céspedes, Noise reduction in elastograms
marized in Table 2. Although the SAD-based spectral using temporal stretching with multicompressional averaging, Ultra-
elastograms (Fig. 11c) exhibit a lower strain contrast than sound Med. Biol. 22 (1996) 1043.
the NCC- or SSD-based approaches (albeit not statistically [10] S.K. Alam, J. Ophir, E.E. Konofagou, An adaptive strain estimator
significant, p > 0.51), the spectral NCC and SSD elasto- for elastography, IEEE Trans. Ultrason. Ferroelec. Freq. Control 45
(1998) 461.
graphic methods produce elastograms with comparable [11] S. Srinivasan, F. Kallel, R. Souchon, J. Ophir, Analysis of an
image contrast (see Table 2). Notice that the relative per- adaptive strain estimation technique in elastography, Ultrasonic
formance of the spectral shift estimators utilized in these Imaging 24 (2002) 109.
experiments was accurately predicted using the simplified [12] E.E. Konofagou, T. Varghese, J. Ophir, S.K. Alam, Power spectral
1-D elastographic simulation (see Section 2). strain estimators in elastography, Ultrasound Med. Biol. 25 (1999)
1125.
[13] T. Varghese, E.E. Konofagou, J. Ophir, S.K. Alam, M. Bilgen, Direct
4. Conclusions strain estimation in elastography using spectral cross-correlation,
Ultrasound Med. Biol. 26 (2000) 1525.
A comparison of various spectral strain estimators, cou- [14] S.K. Alam, F.L. Lizzi, T. Varghese, E.J. Felella, S. Ramachandran,
Adaptive spectral strain estimators for elastography, Ultrasonic
pled without and with subsample shift estimation, was per- Imaging 26 (2004) 131.
formed using simulations and experiments. It was shown [15] K. Hoyt, F. Forsberg, J. Ophir, Investigation of parametric spectral
that results obtained using the spectral SSD-based elasto- estimation techniques for elasticity imaging, Ultrasound Med. Biol.
graphic method demonstrated no marked difference when 31 (2005) 1109.
compared to those obtained using the more computation- [16] K. Hoyt, Spectral strain estimation techniques for tissue elasticity
imaging, Ph.D. dissertation, Drexel University, 2005.
ally costly NCC-based approach. To date, the latter has [17] F. Viola, W.F. Walker, A spline-based algorithm for continuous time-
been the preferred function in several spectral elastographic delay estimation using sampled data, IEEE Trans. Ultrason. Ferro-
reports [13,15]. Moreover, the spectral SAD-based elasto- elec. Freq. Control 52 (2005) 80.
108 K. Hoyt et al. / Ultrasonics 44 (2006) 99–108

[18] F. Viola, W.F. Walker, A comparison of the performance of time- [20] T. Hall, M. Bilgen, M. Insana, T. Krouskop, Phantom materials for
delay estimators in medical ultrasound, IEEE Trans. Ultrason. elastography, IEEE Trans. Ultrason. Ferroelec. Freq. Control 44
Ferroelec. Freq. Control 50 (2003) 392. (1997) 1355.
[19] E.I. Céspedes, Y. Huang, J. Ophir, S. Spratt, Methods for the [21] T. Varghese, J. Ophir, An analysis of elastographic contrast-to-noise
estimation of subsample time-delays of digitized echo signals, ratio, Ultrasound Med. Biol. 24 (1998) 915.
Ultrasonic Imaging 17 (1995) 142.

S-ar putea să vă placă și