Professional Documents
Culture Documents
8, AUGUST 2015
1969
AbstractBackground: Speckle noise is an inherent characteristic of dynamic contrast-enhanced ultrasound (DCEUS) movies and
ultrasound images in general. Speckle noise considerably reduces
the quality of these images and limits their clinical use. Currently,
temporal compounding and maximum intensity persistence (MIP)
are among the most widely accepted processing methods enabling
the visualization of vasculature using DCEUS. Goal: A different
approach has been used in this study, in order to improve the noise
removal, while enabling the investigation of CEUS dynamics. Methods: A multiplicative model for the formation of DCEUS speckled
images is adopted and the log-transformed cines are processed. A
preprocessing step was performed, locally removing low value outliers. Due to the fast-changing spatial distribution of microbubbles
inside the vasculature, the noise in consecutive DCEUS frames is
independent, facilitating its removal by temporal denoising. Noise
reduction is efficiently achieved by wavelet denoising, in which the
signals wavelet coefficients are thresholded and small-value noiserelated coefficients are discarded. The main advantage of using
wavelet denoising in the present context is its ability to estimate
ultrasound contrast agents (UCA) concentration over time adaptively, without assuming a model or predefining the signals degree
of smoothness. The performance of wavelet denoising was compared against MIP, temporal compounding, and Log-normal model
fitting. Results: Phantom experiments showed improved SNR, using wavelet denoising over a wide range of UCA concentrations
(MicroMarker, 0.0011%). In the in vivo tests, improved noise removal was achieved, reflected by a significantly lower coefficient of
variation in homogeneous vascular regions (p < 0.01).
Index TermsContrast-enhanced ultrasound, denoising, multiplicative noise, outlier-resistant estimation.
I. INTRODUCTION
T is well known that bloodtissue contrast in ultrasound
scans can be improved by injecting the patient with ultrasound contrast agents (UCA) and using contrast-specific pulse
sequences [1]. The increased echogenicity and typical radii
Manuscript received October 13, 2014; revised January 13, 2015; accepted
February 14, 2015. Date of publication February 27, 2015; date of current
version July 15, 2015. This work was supported by the Canadian Institutes of
Health Research, the Terry Fox Foundation, the Ontario Research Fund, and
VisualSonics Inc. Asterisk indicates corresponding author.
A. D. Bar-Zion is with Department of Biomedical Engineering,
TechnionIsrael Institute of Technology, Haifa 32000, Israel (e-mail:
barz@tx.technion.ac.il).
C. Tremblay-Darveau is with the Department of Medical Biophysics, University of Toronto.
M. Yin is with the Sunnybrook Health Sciences Center.
D. Adam is with the TechnionIsrael Institute of Technology.
F. S. Foster is with the Sunnybrook Health Sciences Center, and also with the
University of Toronto.
This paper has supplementary downloadable material available at http://
ieeexplore.ieee.org (File size: 7.17 MB).
Digital Object Identifier 10.1109/TBME.2015.2407835
range makes UCAs ideal for the imaging of blood flow. When
imaged at low intensities (mechanical index < 0.05), microbubbles oscillate stably and circulate in the vasculature for long
periods of time, typically many minutes, before collapsing or
being absorbed via endocytosis. The extended circulation of
contrast agents inside the vasculature increases the importance
of the wash-out phase in scans incorporating a bolus injection
of UCA, and enables the use of targeted microbubbles that
bind to specific antibodies [2]. The use of DCEUS imaging
for the quantification of blood perfusion relies on the linear
correlation between UCA concentration and the received harmonic ultrasound intensity. As presented in [3], with a sufficient dynamic range, this linearity holds over three orders of
dilution.
The ability to detect echoes of nonlinear bubble oscillations
and to suppress linear signals from tissue is essential for perfusion imaging. Several multipulse sequences were proposed for
that purpose [4]. In the pulse inversion technique, the polarity of
every two consecutive pulses is altered, preserving the even harmonics. Alternatively, amplitude modulation can be used for the
separation of odd harmonics. A method that manipulates both
amplitudes and phases has been shown to improve UCA detection sensitivity [4]. Using these multipulse sequences, UCA can
be imaged with sufficient contrast to tissue ratio.
Speckle noise is an inherent feature of DCEUS and of ultrasound imaging in general, as it is a coherent imaging modality.
This noise results from the presence of many small scatterers
within each resolution cell, defined by the 6 dB of the main
beam profile in the axial, lateral, and out of plane orientations.
The combination of these reflections creates complicated constructive and destructive interference patterns in the echo signal,
named speckle noise. Speckle noise further reduces the contrast
and resolution of DCEUS and B-mode images and limits their
diagnostic value.
Therefore, speckle reduction has been the focus of many studies aiming to improve the quality of B-mode ultrasound imaging. Following the multiplicative model of speckle formation
first presented in [5], the log transform can be used to convert
the multiplicative noise into additive noise. This additive noise
component can be removed using a variety of classical denoising
methods, for example, wavelet denoising [6], [7]. The underlying assumption of this family of methods is that the resulting
log-compressed images contain an independent identically distributed (i.i.d.) Gaussian noise. In [8], the authors emphasized
both theoretically and experimentally that these assumptions are
generally wrong for log-transformed B-mode images. They presented a preprocessing step that reduced the spatial correlation
of the noise and removed noise outliers. This step produced noise
0018-9294 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications standards/publications/rights/index.html for more information.
1970
BAR-ZION et al.: DENOISING OF CONTRAST-ENHANCED ULTRASOUND CINE SEQUENCES BASED ON A MULTIPLICATIVE MODEL
(1)
r= =
m2 m21
where m is the moment of the corresponding distribution. Inserting the moments of the Rayleigh distribution, we
get r = 1.91 [20]. This constant ratio between the signal (expected value) and the noise (standard deviation) indicates that
the speckle noise is multiplicative for high UCA concentrations.
For the cases where low doses of UCA are injected, or poorly
perfused tissues are scanned, the microbubble concentrations
can be quite low. Many different statistical distributions were
used to describe cases in which the concentration of scatterers
was low, or they could not be assumed to be independently
distributed. The K-distribution will be used in this study, as
proposed in [9], for the modeling of CEUS signals over a large
range of concentrations
bA
2b
K 1 (bA) , A 0
(2)
p( A| , b) =
() 2
where is the shape parameter of the K-distribution, E{A2 } is
the second moment of the amplitude, is the gamma function,
and the function Kx () is the modified Bessel
function of the
second kind with order x. Parameter b = 2 /E{A2 } serves
as a scale parameter. In [21], parameter was shown to be
associated with the concentration of effective reflectors
= ( + 1) Ns ,
> 1
(3)
. (4)
m = E {A } = 2
/2 ()
Inserting (4) to (1), one obtains the amplitude SNR for the
K-distribution
1
. (5)
r=
( ()) / 1 + 12 + 12 1
It should be noted that according to (4), the expected value
of the intensity signal m2 is equal to 2 2 . Thus, the linear correspondence between the ultrasound mean intensity and UCA
concentration, even at low concentrations, can be explained by
the K-distribution model, since 2 is proportional to the number
of scatterers in the resolution cell [23].
B. Model for Speckled CEUS Images
There is no unanimously accepted model for the formation
of speckled CEUS images; therefore, the multiplicative speckle
noise model, widely used for the denoising of B-mode images,
was adopted here. A general multiplicative model for a speckle
noise was proposed in [5]
g [n, m] = f [n, m] u [n, m] + [n, m]
(6)
1971
where g and f are the measured and the original image, respectively. The multiplicative component of the noise is given by u,
represents the additive noise component. Here, n and m are the
radial and angular coordinates, respectively, for a phased-array
transducer, or axial and lateral coordinates for a linear transducer. The multiplicative component of the noise relates to the
physics of coherent imaging, the additive noise component represents electrical measurement noise and phenomena not fully
described by the model.
In [8], it was claimed that compared to the stronger specklerelated multiplicative noise component, the additive noise in
B-mode ultrasound images is negligible. The same strategy is
used here in the context of DCEUS scans despite the lower SNR
of the harmonic signal. Applying the log transform to (6), the
multiplicative speckle noise u is converted into an additive term.
Defining g, f, and u
as the logarithms of g, f, and u, respectively,
one obtains
g [n, m, l] = f [n, m, l] + u
[n, m, l]
(7)
1972
for which the noise does not fit the FisherTippet model and
can be described using the K-distribution.
t > t0
(9)
where AUC is the area under curve of the fitted model, t0 is
the time delay before the beginning of the wash-in phase, I0
is the intensity offset. On a logarithmic time scale, and
are the location and scale parameters, respectively. Due to its
popularity, this model is an important reference used for the
evaluation of new DCEUS processing methods.
The diffusion-with-drift model is based on a fluid-dynamic
interpretation of the indicator-dilution process. The combination
of diffusive and convective transport is described by
2 C (x, t)
C (x, t)
C (x, t)
=D
v
t
x2
x
(10)
(M 1)/2
g [l i]
i=(M 1)/2
N
1
[k] H [k] exp
G
k =0
i2kl
N
(11)
N
1
a [k]
g , [k] [k]
(12)
k =0
N
1
k =0
ak (
g , [k])
g , [k] [k].
(13)
BAR-ZION et al.: DENOISING OF CONTRAST-ENHANCED ULTRASOUND CINE SEQUENCES BASED ON A MULTIPLICATIVE MODEL
(15)
where k is the threshold for level k, defining a small percentage p of the coefficient gk as outliers. Next, the detected
residuals can be subtracted from gk and the next level of the
wavelet decomposition can be performed. As discussed in [19],
short decomposition wavelet filters reduce the leakage of outlier
1973
patches into the smooth coefficients, isolating them and facilitating their removal. On the other hand, long reconstruction filters
are desirable in order to ensure a sufficient level of smoothness.
Following [19] the b-spline biorthogonal wavelets with three
and nine vanishing moments were used in this study.
In this study, the noise contaminating the DCEUS movies
contains isolated outliers in the time domain. In [24], the Fisher
Tippet noise containing isolated outliers was processed by a
single use of the outlier removal procedure. It was shown that
this outlier removal step successfully removed the heavy tail
of the distribution, producing the semi-Gaussian noise. At the
same time, the short moving median had a negligible effect
on the signal. Similarly, a single use of the outlier removal
procedure was used in the current study. The resulting semiGaussian noise can then be removed using the standard threestep algorithm with the universal threshold rule. This processing
scheme is robust to the noise outliers and also incorporates the
numerical efficiency of wavelet denoising.
IV. VALIDATION OF THE PROPOSED MODEL
All the phantom and in vivo experiments were performed using the Vevo 2100 (VisualSonics Inc., Toronto, Canada) scanner
and the linear array MS250 transducer in contrast mode. Amplitude modulation is used in the Vevo2100 for the separation
of CEUS signal. The cines were exported from the scanner as
linear raw data files (without log compression).
A. Phantom Experiments
In order to evaluate the performance of the proposed denoising algorithm when compared to established methods, a series
of flow phantom experiments were performed.
1) Ultrasound Contrast Agents: MicroMarker UCA was used
in all the phantom and in vivo experiments. Before each experiment, a new vial was activated according to the companys
protocol. Following the activation, the size distribution of the
microbubbles was measured using a Beckman Coulter Multisizer 3 system. In order to obtain reliable quantification, 10 L
of UCA was diluted in 10 mL of Isoton II diluent [9]. In each
test, 100 L of this solution was drawn by the Multisizer and
each measurement was repeated three times. The aperture used
detects bubbles with radii as small as 0.7 m.
2) Continuous-Flow Phantom Setup: The purpose of the
continuous-flow phantom experiments is threefold: to detect the
linear region of the relationship between the UCA concentration
and the ultrasound intensity, to validate the K-distribution
model, and to compare the effectiveness of the MIP, temporal
compounding, and wavelet denoising algorithms using movies
containing homogenous concentrations of UCA. The setup of
this phantom is described in Fig. 1(a). Before each injection, a
solution of deionized water and UCA was mixed using a magnetic stirrer. Then, the solution was loaded into a 3-mL syringe
and injected using a syringe pump (model NE-4000, New Era
Pump Systems, Inc., Farmingdale, NY, USA) into a polyethylene tube (internal diameter of 1.4 mm) submerged inside a small
water tank. The injection rate was set to 7.7 L/s. Cines of 300
frames at 25 frames/s were recorded. The scans were repeated
using four different vials of MicroMarker. The transducer was
1974
TABLE I
CONTRAST AGENTS CONCENTRATIONS USED IN CONTINUOUS-FLOW
PHANTOM EXPERIMENT
Relative
concentration
A
A/2
A/5
A/10
A/20
A/50
A/100
A/200
A/500
A/1000
A/2000
A/5000
Contrast agent
concentration
(%)
Scatterers per
ultrasound resolution
cell
5%
2.5%
1%
0.5%
0.25%
0.1%
0.05%
0.025%
0.01%
0.005%
0.0025%
0.001%
743.5
371.7
185.9
74.35
37.17
18.59
7.44
3.72
1.86
0.74
67500 3500
33750 1750
13500 700
6750 350
3375 175
1350 70
675 35
337.5 17.5
135 7
67.5 3.5
33.75 1.75
13.5 0.7
The number of effective Scatterers per ultrasound resolution cell was estimated by fitting
the measured amplitude SNR to the K-distribution model (see Fig. 2).
Fig. 1. Phantom setup. (a) Continuous-flow setup during the scanning phase.
(b) Indicator-dilution setup.
aligned to the center of the flow channel with an in-plane angle of about 30 to minimize reverberations. A syringe pump
was used in this experiment in order to reduce flow velocity
fluctuations.
Several measures were made to improve the quality of the
measurements. Before mixing each solution, the vial of the
contrast agents was shaken to achieve stable microbubble concentration over time. Before each injection, deionized water was
pumped through the tubing using a peristaltic pump (Minipuls 3,
Gilson Inc., Middleton, WI, USA) until no traces of UCA
were seen by the scanner. In addition, increasing concentrations of UCA were injected in order to reduce the influence
of residual microbubbles between consecutive injections. Finally, great care was taken to remove static air bubbles from the
tubing.
For each vial, a range of contrast agent dilutions spanning
four orders of magnitude were scanned (see Table I). This range
of dilutions includes both tens of microbubbles in each resolution cell, complying with the Rayleigh distribution approximation, and resolution cells containing single bubbles. Assuming
that the average blood volume of a laboratory mouse is about
1.52.5 mL; a 50 L bolus injection of UCA would results in a
preclinical concentration of around 2.5%.
The recorded cines were loaded into a personal computer
and analyzed using dedicated MATLAB code (The Mathworks,
Natic, MA, USA). A ROI was selected in each cine, containing
a cross section of the tube with a negligible signal originating
from the walls of the tube. Since three tests were performed
using this phantom setup, three different processing methods
were applied to the data. First, in order to detect the linear region of the UCA concentration versus ultrasound intensity, the
intensity of the pixels inside the ROIs was averaged throughout the movie. Next, in order to validate the K-distribution
model, the amplitude SNR was calculated inside the ROI for
every frame and averaged throughout the cine. Then, the SNR
values were fitted to the K-distribution SNR model of (5).
The K-distribution model enabled the normalization of the concentration axis so that the SNR can be presented as a function of
the number of effective scatterers per resolution cell. Finally, the
log-transformed cines which were processed using the different
algorithms were compared in order to evaluate the effectiveness
of the algorithms in estimating the TIC and the noise removal.
The moving average filter of the temporal compounding algorithm was chosen to be 21 samples long, to achieve a reasonable
noise reduction, while avoiding oversmoothing the data during
the wash-in phase. Following the considerations described in the
previous section, wavelet denoising was performed using the 3,
9 biorthogonal wavelets. The robust residual removal procedure
was set so that the length of the moving-median window was 7
and p was set to 7%.
The spatial coefficient of variation was used to evaluate the
ability of the different algorithms to remove the speckle noise.
The underlying assumption was that the intensity within a small
ROI should be uniform. The spatial coefficient of variation (CV )
was defined as the ratio of the standard deviation to the mean
of log-intensity value in each processed movie within the
ROI.
3) Indicator-Dilution Phantom Setup: An indicator-dilution
phantom was used to test the effectiveness of the different algorithms on real scans containing changing concentrations of
UCAs. The setup of this phantom, based on [27], is described in
Fig. 1(b). The three-way stopcock was used to connect both the
reservoir and the syringe pump to the phantom. The peristaltic
pump was used to maintain a steady flow of 36 L/s within the
system. The syringe pump, filled with undiluted UCA, was used
to inject a bolus of 50 L into the phantom. Cines of 920 frames
at 25 frames/s were recorded using the Vevo 2100 scanner in
contrast mode. The alignment of the MS250 probe was carried
out as described in the previous section.
The recorded cine was loaded into a personal computer and
analyzed using MATLAB. A ROI was selected as described
above. Regional intensity measurements were calculated by averaging the intensity of the pixels included inside the entire
5 5 pixel ROI. Local TIC were estimated from single pixels.
BAR-ZION et al.: DENOISING OF CONTRAST-ENHANCED ULTRASOUND CINE SEQUENCES BASED ON A MULTIPLICATIVE MODEL
1975
Fig. 2. Continuous-flow phantom results. (a) Linearity of the relationship between the raw data intensity and UCA concentration. (b) Measured amplitude SNR
and its compliance to the K-distribution model. (c) Spatial coefficient of variation for different processing durations (0.1% UCA). (d) Spatial coefficient of variation
for different concentrations of UCA.
1976
Fig. 3. Continuous-flow phantom results. Single frames taken from the different processed cines (t = 1 s).
BAR-ZION et al.: DENOISING OF CONTRAST-ENHANCED ULTRASOUND CINE SEQUENCES BASED ON A MULTIPLICATIVE MODEL
1977
TABLE II
MSE AND CORRELATION COEFFICIENTS OF SINGLE PIXEL TIC ESTIMATIONS
Fitting
MIP
TC
WD
MSE [a.u.]
0.707 0.004
0.641 0.011
0.719 0.009
0.742 0.003
9.70 2.00
172.34 21.44
7.46 1.73
6.93 1.48
VI. DISCUSSION
during the wash-in phase (50 s), the peak-enhancement (100 s),
and the wash-out phase (250 s) are presented on the top, middle, and bottom column. To the right, the corresponding frames
from the processed MIP, TC, and WD movies are presented,
allowing us to compare the three methods. Since MIP holds on
to the maximum intensity value over time, there is no significant
change between the MIP image at the peak-enhancement time
and wash-out time. Moreover, the high concentration of UCA
entering the liver during the first few frames of the wash-in is
retained by MIP throughout the movie. In contrast, the UCA
signal decrease can be appreciated in both TC and WD images
(for example, inside the brain). The ability of the WD to reject noise from DCEUS images is noticeable when observing
homogeneous areas such as the carotid arteries.
The shapes of single blood vessels in the core of a hind-limb
tumor depicted by WD are presented in Fig. 6. Fig. 1ex (supplementary material) compares the WD results with MIP and
TC. An interesting feature observed in Fig. 1ex is the robustness
1978
Fig. 5. Single frames from different processed cines during wash-in, peak enhancement, and wash-out phases of a mouse embryo scan. The noise levels are
considerably lower in WD images especially during the wash-in phase (see arrows), while MIP presents partial filling of blood vessels. False structures in the
MIP scans and their locations in WD images are marked by stealth arrows. The ability of WD to present the wash-out phase is easily seen in the brain tissue
(arrowheads).
Fig. 6.
It was also found that in MIP movies the Cv values are initially
high, even when the concentration of UCA is constant. The Cv
values slowly decrease only after few seconds [see Fig. 2(c)]
to a level comparable to TC and WD. This means that the MIP
cines are very noisy at first and gradually become clearer. This
delayed response is problematic when the dynamics of DCEUS
cines is studied.
The proposed method has two main theoretical advantages
over classical TC. The first is that TC relies on the implicit assumption of the Gaussian noise. However, it is well known that
log-transformed CEUS signals suffer from low-value outliers.
These outliers motivated the introduction of a preprocessing
step for their removal. The second advantage of WD is that unlike TC, it does not require the definition of a cutoff frequency
a priori. Moreover, a low-pass filter cannot separate the signal
and the noise efficiently since they have an overlapping supports in the frequency domain. Therefore, the optimal cut-off
frequency depends on the local flow characteristics and is difficult to estimate in advance. In contrast, WD relies on the sparse
BAR-ZION et al.: DENOISING OF CONTRAST-ENHANCED ULTRASOUND CINE SEQUENCES BASED ON A MULTIPLICATIVE MODEL
TABLE III
RESOLUTION RATIO GAIN OF WD IN COMPARISON TO MIP
LS174T tumors
PC3 tumors
Mouse Embryos
All scans (n = 20)
1.66 0.38
1.96 0.26
2.27 0.19
1.99 0.37
1.45 0.1
1.74 0.05
1.62 0.13
1.60 0.15
1979
1980
Charles Tremblay-Darveau was born in St-CharlesBorromee, QC, Canada, in 1987. He received the
B.Sc. degree in physics from McGill University,
Montreal, QC, in 2009, and the M.S. degree in
medical biophysics from the University of Toronto,
Toronto, ON, Canada, in 2011 and is currently working toward the Ph.D. degree in medical biophysics at
the University of Toronto.
His research interests include nanofluidics,
Doppler contrast-enhanced ultrasound imaging, and
noninvasive blood pressure measurement with
microbubbles.