Christ Cse

This article has been accepted for publication in a future issue of this journal, but has not been
fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TBME.2015.2486042, IEEE Transactions on Biomedical Engineering
> TBME-00795-2015.R1<
A New Feature-enhanced Speckle Reduction

Method based on Multiscale Analysis for
Ultrasound B-mode Imaging
Jinbum Kang, Student Member, IEEE, Jae Young Lee, and Yangmo Yoo*, Member, IEEE
Abstract Goal: Effective speckle reduction in ultrasound

B-mode imaging is important for enhancing the image quality and
improving the accuracy in image analysis and interpretation. In
this paper, a new feature-enhanced speckle reduction (FESR)
method based on multiscale analysis and feature enhancement
filtering is proposed for ultrasound B-mode imaging. In FESR,
clinical features (e.g., boundaries and borders of lesions) are
selectively emphasized by edge, coherence and contrast
enhancement filtering from fine to coarse scales while
simultaneously suppressing speckle development via robust
diffusion filtering. In the simulation study, the proposed FESR
method showed statistically significant improvements in edge
preservation, mean structure similarity, speckle signal-to-noise
ratio and contrast-to-noise ratio (CNR) compared with other
speckle reduction methods, e.g., oriented speckle reducing
anisotropic diffusion (OSRAD), nonlinear multiscale wavelet
diffusion (NMWD), the Laplacian pyramid-based nonlinear
diffusion and shock filter (LPNDSF) and the Bayesian non-local
means filter (OBNLM). Similarly, the FESR method
outperformed the OSRAD, NMWD, LPNDSF and OBNLM
methods in terms of CNR, i.e., 10.700.06 vs. 9.000.06, 9.780.06,
8.670.04 and 9.220.06 in the phantom study, respectively.
Reconstructed B-mode images that were developed using the five
speckle reduction methods were reviewed by three radiologists for
evaluation based on each radiologists diagnostic preferences. All
three radiologists showed a significant preference for the
abdominal liver images obtained using the FESR methods in
terms of conspicuity, margin sharpness, artificiality and contrast,
p<0.0001. For the kidney and thyroid images, the FESR method
showed similar improvement over other methods. However, the
FESR method did not show statistically significant improvement
Manuscript received Dec 07, 2014; revised Jun 28, 2015 and Aug 26, 2015;
accepted Sep 23, 2015. Date of publication XXX XX, 2015; date of current
version XXX XX, 2015. This research was supported by the Converging
Research Center Program through the Ministry of Science, ICT and Future
Planning, Korea (2014066284). Also, this research was supported by the MSIP
(Ministry of Science, ICT and Future Planning), Korea, under the C-ITRC
(Convergence
Information
Technology
Research
Center)
(IITP-2015-H8601-15-1004) supervised by the IITP (Institute for Information
& communications Technology Promotion).
J. Kang is with the Department of Electronic Engineering, Sogang
University, Seoul 121-742, Korea (e-mail: jbkang@sogang.ac.kr).
J. Lee is with the Department of Radiology, Seoul National University
Hospital, Seoul 110-744, Korea (e-mail: leejy4u@snu.ac.kr).
*Y. Yoo is with the Department of Electronic Engineering and the
Interdisciplinary Program of Integrated Biotechnology, Sogang University,
Seoul 121-742, Korea (e-mail: ymyoo@sogang.ac.kr).
Copyright (c) 2015 IEEE. Personal use of this material is permitted.
However, permission to use this material for any other purposes must be
obtained from the IEEE by sending an email to pubs-permissions@ieee.org.
compared with the OBNLM method in margin sharpness for the

kidney and thyroid images. These results demonstrate that the
proposed FESR method can improve the image quality of
ultrasound B-mode imaging by enhancing the visualization of
lesion features while effectively suppressing speckle noise.
Index TermsSpeckle reduction, feature
multiscale analysis, medical ultrasound imaging
enhancement,
I. INTRODUCTION
quality in ultrasound B-mode imaging has been

rapidly improved over the last few decades through
technological advancements in transducer sensitivity, digital
beamformer and image processing techniques [1]. However,
ultrasound B-mode imaging still suffers from low contrast
resolution because the difference in acoustic impedance among
various soft tissues is marginal (~1%) [2]. In addition, speckle,
which appears as a granular pattern formed by the constructive
and destructive interferences from backscattered ultrasound
waves, substantially lowers the image contrast and obscures
image details [3]. Moreover, speckle interferes with image
analysis and interpretation, e.g., segmentation, registration, and
computer aided detection (CAD) [4].
The speckle pattern depends strongly on the structure of the
imaged tissues and various imaging parameters, e.g., the
frequency and geometry of the ultrasound transducer [5]. It is
difficult to suppress the speckle when only applying simple
averaging and low noise acoustic and electric system designs
[6]. Furthermore, the excessive suppression of speckle disturbs
the interpretation of ultrasound images that contain lesions
because a certain amount of speckle contains necessary
diagnostic information. Thus, effective speckle reduction is
important for improving the image quality in medical
ultrasound imaging and the performance of image analysis and
interpretation.
Several speckle reduction techniques, such as frequency
compounding (SR-FC), spatial compounding (SR-SC), and
post filtering (SR-PF), have been proposed [7-39]. In SR-FC,
uncorrelated sub-images, which are obtained either by varying
the center frequency on transmission or by dividing the
spectrum of radio-frequency (RF) signals on reception, are
compounded to lower speckle [7-17]. In SR-SC, the
sub-images are acquired at different beam orientations at the
expense of temporal resolution [18-21]. These compounding
techniques considerably reduce speckle and make the
MAGE
0018-9294 (c) 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
> TBME-00795-2015.R1<
ultrasound images cleaner with clearer boundaries, but the
images suffer from degraded spatial resolution and increased
system complexity. In contrast, rather than compounding
multiple sub-images, in SR-PF, a specially designed spatial
filter is applied to a single image [22-32]. The SR-PF method
can be classified as a single-scale linear or nonlinear spatial
filtering method and as a spatial filtering method based on
multiscale analysis (i.e., SR-PFSC and SR-PFMS, respectively).
In SR-PFSC, various spatial filters, e.g., smoothing [22],
Wiener [23], unsharp masking [24], region growing [25-27],
slope-facet modeling [28] and the stochastic approach [29],
have been proposed to reduce speckle. Though these spatial
filters can indeed reduce speckle, they suffer from excessive
suppression, blurred boundaries and obscured details. To more
effectively suppress speckle, diffusion-based spatial filtering
methods utilizing partial difference equations (PDE) have been
proposed.
In the speckle reducing anisotropic diffusion (SRAD)
method [30], an edge-sensitive anisotropic diffusion process
that utilizes the instantaneous coefficient of variation, which is
a function of the local gradient magnitude and Laplacian
operators, is applied. Similarly, an advanced SRAD (OSRAD)
method has been proposed, in which directional filtering of the
image along structures is performed by a diffusion matrix to
preserve boundaries [31]. Similarly, in nonlinear coherent
diffusion (NCD), three different diffusion filters are
dynamically applied, depending on the various speckle regions
[32]. The OSRAD and NCD methods show enhanced speckle
reduction and edge perseverance, but they still have limitations
in retaining fine features such as small lesions or distinguishing
small structures from tiny cysts.
Such nonlinear iterative diffusion methods based on gradient
operators have difficulty in accurately differentiating signal and
noise. Thus, a multi-scale analysis approach, which has the
potential to provide effective signal and noise separation, can
be utilized for speckle reduction, i.e., SR-PFMS. The
wavelet-based speckle reduction method, in which soft
thresholding of the logarithmically transformed image is used,
has been proposed [33]. However, it is difficult to determine a
suitable threshold level for the wavelet coefficients. The
recently proposed nonlinear diffusion filtering and edge
enhancement method based on dyadic wavelet transform
(NMWD) shows favorable denoising properties. These
properties are supported by the integration of nonlinear
diffusion and wavelet shrinkage, but the approach is lacking in
terms of its ability to preserve edges and structural similarities
[34].
In the same manner, the Laplacian pyramid transform
method combined with nonlinear diffusion (LPND) has been
proposed. In the pyramid domain, a signal is decomposed into
decimated lowpass and bandpass signals on a finer scale and is
then reconstructed into a single interpolated signal from a
coarse scale [35]. The pyramid decomposition and
reconstruction of a two dimensional (2D) image has the
advantage of not requiring a quadrature mirror filter (QMF)
pair, as the wavelet-transform methods do [33, 34]. However,
although the LPND method successfully suppresses speckle, it
2
cannot preserve the edges and boundaries of lesions. To
compensate for the blurring caused by the LPND method, a
shock filter was adopted, i.e., the Laplacian pyramid-based
nonlinear diffusion and shock filter (LPNDSF) [36]. The
LPNDSF method utilizes nonlinear diffusion and a shock filter
in a coupled partial differential equation (PDE) process in the
Laplacian pyramid domain. However, the edges and structures
of lesions in the processed image tend to appear too artificial or
unnatural for use in ultrasound imaging.
In addition, diverse non-local patch based algorithms with a
Bayesian framework have recently shown competitive
performance [37-39]. Among these methods, the optimized
Bayesian non-local means with blockwise approach (OBNLM)
method [37] is performed to reduce the complexity burden
while suppressing speckle noise and preserving structural
details. Nevertheless, its patch comparison procedure for
selecting a relevant criterion still leads to high computational
complexity.
In this paper, a new feature-enhanced speckle reduction
(FESR) method based on multiscale analysis and feature
enhancement filtering is proposed to more effectively reduce
speckle noise while preserving tissue structures and enhancing
low echogenicity with improved border definition and
continuity. The performance of the proposed FESR method is
quantitatively compared with other speckle reduction methods
(i.e., OSRAD, NMWD, LPNDSF and OBNLM) in terms of
edge preservation, mean structure similarity, speckle
signal-to-noise ratio (SSNR) and contrast-noise ratio (CNR).
Moreover, a qualitative analysis conducted by three
experienced radiologists that demonstrates their diagnostic
preferences is presented in this paper.
This paper is structured as follows. In Section II, a speckle
model for medical ultrasound imaging is presented. In Section
III, the nonlinear anisotropic diffusion and pyramid transform
that are the foundation of the proposed FESR method are
reviewed. In Section IV, the proposed FESR method is detailed.
In Section V, the experimental setup and evaluation metrics are
presented. In Section VI, the experimental results are discussed
and compared with other speckle reduction methods in
simulation, phantom, and in vivo studies with both quantitative
and qualitative evaluation. Section VII concludes the paper.
II. SPECKLE MODELING
Variations in tissue attenuation, propagation and scattering
properties cause ultrasound echoes to interfere in complex
ways. Echoes from reflectors spaced closer than the resolution
limit will interfere, thus creating artificially large (i.e.,
constructive interference) or small (destructive interference)
signals. Therefore, constructive and destructive coherent
interferences from backscatterer echoes that are typically much
smaller than the wavelength of an ultrasound incident wave
create phase summations or phase cancellations. Thus, the
speckle pattern can be defined as a multiplicative noise, rather
than a type of normal additive noise. The speckle noise model
for a medical ultrasound image can be approximated as
> TBME-00795-2015.R1<
( x, y) f ( x, y)wm ( x, y) wa ( x, y)
(1)
where ( x, y) is the received signal at a spatial location of

( x, y) , f ( x, y) is the noise-free original image, and wm ( x, y)
value k is a crucial factor as a diffusion coefficient

determinant. In addition, if the diffusivity function c( x) is
linear, the diffusion filter can be an isotropic diffusion function
such as the Gaussian filter.
and wa ( x, y) represent the multiplicative noise and additive

noise, respectively. The effect of the additive noise wa ( x, y)
can be disregarded because it is insignificantly small compared
with the multiplicative noise ( wm ( x, y)2 wa ( x, y)2 ). Then,
Eq. (1) can be rewritten as
( x, y) f ( x, y)wm ( x, y)
(2)
Generally, ultrasound imaging systems have a

log-compression stage for adequately displaying the envelope
signal on a monitor. As such, multiplicative noise can be
converted into additive noise as follows:
log ( x, y) log f ( x, y) log wm ( x, y)
(3)
Fig. 1. Graph of the nonlinear diffusivity coefficient c1 ( x) as the gradient

threshold k is varied.
B. Hierarchical Laplacian pyramid model
Thus, a speckle pattern can be modeled as an additive white

Gaussian noise distribution, which is the logarithmically
compressed signal developed after envelope detection [40].
This Gaussian noise distribution is utilized in this study.
III. NONLINEAR DIFFUSION AND PYRAMID TRANSFORM
A. Nonlinear anisotropic diffusion
Diffusion is a physical process of balancing concentration
changes, such as those found in heat transfer. The diffusion
equation (Ficks second law) is a partial differential equation
(PDE), which describes density fluctuations in a material
undergoing diffusion. Based on this physical phenomenon,
Perona and Malik [39] proposed the nonlinear PDE to remove
noise from an image on a continuous domain:
I ( x, y, t )
div c I I
t
I (t 0) I 0
Fig. 2. Hierarchical Laplacian pyramid transform.
(4)
where is the gradient operator, div is the divergence

operator, I is the gradient magnitude of an image I , c( x) is
the diffusion coefficient or edge-stopping function, and I 0 is
the initial image.
In the anisotropic diffusion method, the gradient magnitude
I and the diffusion coefficient c( x) , which is a
monotonically decreasing function, have a reciprocal
correlation. If the gradient magnitude I , then c( x) ;
if I 0 , then c( x) 1 . That is, the diffusion filter assumes
that large gradient magnitudes are the edges or boundaries of an
image and that small gradient magnitudes are areas of noise
from the image. Figure 1 shows that the gradient threshold
To extract features or structures from noise, the input image

I (n) (i.e., an image of the nth frame) is decomposed into a
decimated signal by 2D Gaussian lowpass filtering G( )
followed by a sub-sampling by a factor of two in both directions,
as shown in Fig. 2. Each decimated signal (i.e., Gaussian
pyramid) can be subtracted from the up-sampling signal
(insertion of zeros) such that it contains only residual
information (i.e., the Laplacian pyramid [42]). The residual
signal consists of the difference between the signal on a finer
scale and the interpolated signal on a coarser scale. For an input
image I (n) , let its Gaussian pyramid at layer l be Gl (n) and its
Laplacian pyramid at layer l be Ll (n) , where l = 0, 1, 2 . . . b-1
and b is the total decomposition layer. Then, the Gaussian and
Laplacian pyramid transform can be represented by
> TBME-00795-2015.R1<
G0 (n) I (n)
Gl (n) R Gl 1 (n)
(5)
Ll (n) Gl (n) E Gl 1 (n)
where R and E indicate the decimation and interpolation of the

signal, respectively. The sub-band images can be reconstructed
by reversing the decomposition steps.
if too many scales are decomposed, highpass information (e.g.,

edge or boundaries) may be missed at a coarse scale. In addition,
it is shown in Section V. E. that four scales are sufficient to
produce an effective result. Note that the FESR method
performs different enhancement filtering (i.e., edge, continuity
and contrast) from the characterization of each level, unlike
other multiscale approaches [33-36].
IV. FEATURE-ENHANCED SPECKLE REDUCTION ALGORITHM

In general, speckle reduction techniques in ultrasound
B-mode imaging involve a tradeoff between noise suppression
and structure preservation because of the images low
resolution, high noise and high texture. High-frequency
components of B-mode images are characterized by large
changes in gray values over small distances, such as edges and
noise. In contrast, low-frequency components of B-mode
images can be described by small changes in gray values,
including textures and backgrounds. Therefore, the proposed
method performs multiscale analysis to emphasize the features
or structures of a specific frequency by improving speckle noise
suppression.
Fig. 4. The overall block diagram of the proposed FESR method.
In FESR, all of the decomposed layers (i.e., L0 (n) L3 (n) in

Fig. 4) from the Laplacian pyramid model first fulfill effective
noise smoothing by robust noise reduction filtering, as speckle
noise or artifact is established for all frequency components
with local statistical properties. Then, the noise-suppressed
sub-band signals (i.e., L1 (n) L3 (n) in Fig. 4) are selectively or
respectively enhanced in accordance with their textural or
structural properties, as shown in Fig. 4. For the feature
enhancement processing, at the layer L1 (n) (e.g., Fig. 3(b)) of
Fig. 3. The two-dimensional input image and examples decomposed by
Laplacian pyramid transform from fine to coarse scales. (a) (d) noise, line,
texture and object detection, respectively.
Figure 3 illustrates an example of the decomposed

two-dimensional sub-signals of the Laplacian pyramid. As
shown in Fig. 3, the extracted true structures and signals of a
specific layer have different information from the identified
noise to the defined textures. Thus, appropriate noise reduction
and feature enhancement for each scale characteristic must be
considered.
Figure 4 shows the overall block diagram of the proposed
feature-enhanced speckle reduction (FESR) method, which is
decomposed to four levels using the hierarchical Laplacian
pyramid model. The FESR method has two steps: the
decomposition and analysis step by Laplacian pyramid
transform and the filtering and synthesis step by noise reduction
and feature enhancement processing, as illustrated in Fig. 4.
The main reason that the FESR method performs in limited
level is scale limitation, which is studied by [43]. For example,
Laplacian pyramid in Fig. 4, edge enhancement is performed by

a morphological filter combined with nonlinear anisotropic
diffusion for strong discontinuity. For layer L2 (n) (e.g., Fig.
3(c)), coherence enhancement for continuity is processed based
on directional tensor diffusion filtering. The coarsest layer
L3 (n) (e.g., Fig. 3(d)) of the Laplacian pyramid has object
components with gray level values such that histogram-based
local contrast enhancement is implemented in a modified
homomorphic system combined with nonlinear anisotropic
diffusion to improve contrast resolution. Thereafter, all
enhanced sub-band images are synthesized by Laplacian
pyramid reconstruction. Details of the noise reduction (Section
A) and enhancement process (Section B to D) are described in
the following sections.
A. Robust noise reduction filtering
As shown in Fig. 3, whereas constructive or destructive
speckle noise is dominant over all tissues, pepper or black hole
artifacts are diffused. Thus, effective nonlinear anisotropic
diffusion filtering is crucial for multiscale speckle noise
suppression because this step controls the overall performance
> TBME-00795-2015.R1<
of the proposed algorithm. In the FESR method, edge-sensitive

anisotropic diffusion filtering based on robust statistical
estimation is performed for all levels, as shown in Fig. 4. The
relationship between anisotropic diffusion and robust statistics
can be closely defined in the robust estimation framework [44].
Thus, in FESR, Tukeys biweight error norm estimator, which
increases robustness and reject outliers more effectively than
the Least-square and Lorentzian error norms, was selected as
follows:
x 1 x / 2 2 ,
x

x,
0,
otherwise
(6)
where is a scale factor. To analyze the performance of a

given -function, the relationship between a robust error norm
-function and a diffusivity function c x is shown in Fig. 5.
ut sign u u , u x, y, t | t 0
(8)
where the subscripts represent partial derivatives, and

u (ux , u y )T is the gradient of u. For some pixel in the
influence zone of a maximum where its Laplacian operator
u uxx u yy is negative, Eq. (8) becomes ut u . The
evolution of this equation over time t is equal to the dilation
process with a disk-shaped structuring element of radius t [45].
In the influence zone of a local minimum, the Laplacian
operator is positive, and Eq. (8) can be reduced to an erosion
equation ut u with the structuring element.
Thus, the edge enhancement method combining the
morphological filter and nonlinear anisotropic diffusion in (7),
which is applied to the finer scale (i.e., edges and boundaries),
can be expressed by
u
div D u (1 D) sign(G * u ) | u |
t
(9)
where D is the structure tensor based on PDE and is the

determined parameter between edge enhancing and speckle
suppression. The term sign(G * u ) | u | is the advanced
Fig. 5. Tukeys biweight function. ( x)
version of a morphological filter where the second directional

derivative u is convolved with a Gaussian low pass filter G
' ( x) c( x) x .
to increase robustness.
Based on the optimization criterion, we can obtain the robust

noise reduction filter equation by derivation from Eqs. (4) and
(6) as follows:
u
u
div | u | | G u |
t
| u |
(7)
where | G u | is the gradient of a Gaussian lowpass filtered

image with a standard deviation to increase the robustness
from sensitive noise rather than the gradient u of Eq. (4).
B. Sharp edge enhancement
Generally, a simple edge enhancement method shows a
crucial effect in enhancing edges but also tends to amplify noise.
Thus, the FESR utilizes a multiscale morphological filter [45],
which has certain advantages over linear multiscale approaches.
In previous work, the multiscale morphological filter based on
Laplacian pyramid model showed effective improvement of the
detectability of edges and small structures [36], but it showed
unsuitable results for medical imaging due to an artificiality
resulting from excessive edge enhancement. Therefore, the
FESR method solely emphasizes the identified edges or
boundaries (i.e., L1 (n) in Fig. 4) from multiscale analysis, in
contrast to the enhancement of all decomposed scales in [36].
The original morphological filter [45] on the continuous
image f : 2 can be represented by
C. Directional coherence enhancement

The continuity of specular reflectors is important for
visualizing linear striation patterns observed in fat and muscle.
Additionally, the improved linearity and continuity between
structures can enhance the delineation of margins and
connective tissues.
To compensate for collapsed structures or broken lines of
tissue, coherence-enhancing anisotropic diffusion [46] based
on a diffusion tensor rather than the scalar diffusivity is applied
to specific sub-band images (i.e., L2 (n) in Fig. 4). The phased
angle of a structure tensor determines the diffusion direction,
and the diffusivity in this direction is increased by the local
coherence of the signal. Hence, the diffusion tensor D R 2 x 2 is
a symmetric positive semi-definite diffusion tensor
representing the required diffusion in both gradient and contour
directions. To lower the noise sensitivity, the diffusion tensor
matrix D (i.e., second-order derivatives) and a smoothing
kernel K are convoluted as follows:
K * u x2
K * (u xu y )
J u
( 0)
K * (u u
K * u y2
x y)

j
j
11 12
j21 j22
where K is a convolution kernel defined by
(10)
> TBME-00795-2015.R1<
K x, y (2 2 ) 1 exp(
x2 y 2
)
2 2
(11)
According to the eigenvalue decomposition, the structure

matrix in Eq. (10) can then be rewritten as
J (u ) (1 | 2 ) 1
0
0 1T

2 2T
(12)
where the eigenvectors 1 , 2 and the eigenvalues 1 , 2

correspond to the tensor directions of the maximum and
minimum variations and the tensor strengths of these variations,
respectively. The diffusion tensor reflects the local image
structure by using the same set of eigenvectors as the structure
tensor. The constant areas are specified by 1 2 0 , and the
the filter is transiting between and , ( ) is the distance

from the square point to the center point of the two-dimensional
Fourier spectrum, and c is the cutoff frequency.
Figure 6 shows the proposed homomorphic filtering system
for estimating the illumination coefficient and the
reflectance coefficient while changing the local contrast of
the image. As shown in Fig. 6, the coefficients and
emphasize or reduce the illumination and reflectance
components, respectively. The homomorphic system
automatically derives the coefficients and using
histogram analysis, unlike other existing homomorphic systems
[48].
edges are given by 1 2 0 , and corners yield 1 2 0 .

When the region of interest corresponds to the speckle pattern
to create a result similar to fully formed speckle, diffusion must
become isotropic, 1 2 , and inversely proportional to the
information content, with a large difference 1 2 (i.e., local
coherence). Furthermore, when the anisotropic nature of the
image increases (i.e., when the image is rich with information
about the imaged tissue or structure), a diffusion filter must be
very selective in both direction and strength. In the case that a
fully structured region corresponds to 1 2 2 s 2 , diffusion
is performed in a contour direction.
D. Local contrast enhancement
Speckle noise is negligible on the coarsest scale because
low-frequency components are presented on the higher level
with gray level defined objects, as illustrated in Fig. 3(d).
Therefore, poor contrasts or weak signals (i.e., L3 (n) in Fig. 4)
are enhanced by applying a modified homomorphic filtering
based on a gray nonlinear transform. The homomorphic filter is
a convolutional filter based on an illumination-reflectance
image model [47] such that its filter may compress the results of
low-frequency components and expand the results of
high-frequency components.
Because the illumination function can be characterized by
slow-moving spatial variations, the frequency characteristics of
the illumination function concentrate on the low-frequency
components. Additionally, the reflectance function represents
the spatial variation of the object and is thus generally of high
frequency, according to the characteristics of detailed features
and edges. Thus, the transfer function of the typical
homomorphic filter is modified from the Gaussian highpass
filter and defined as
( )2
H 1 exp c 2
Fig. 6. A modified homomorphic filtering system for contrast enhancement.
In detail, the gray level image that is despeckled by nonlinear

anisotropic diffusion (NAD) (i.e., input in Fig. 6) is analyzed
through its histogram shape, and the global threshold to divide
the object and background of image is determined. For an
automatic illumination threshold value, the general Otsus
method [49], which is widely used for bimodal and multimodal
histograms, is applied. The coefficients and can then be
determined using linear approximation.
V. EXPERIMENTAL SETUP
To evaluate the performance of the proposed FESR method,
simulation, tissue-mimicking phantom and in vivo experiments
were conducted. The performance of the proposed FESR
method was compared with that of other speckle reduction
methods, i.e., OSRAD [31], NMWD [34], LPNDSF [36], and
OBNLM [37]. To derive optimal results from these five speckle
reduction approaches, algorithms were downloaded or fully
consulted for these methods [31-37], and filtering parameters
were obtained from [50]. For all studies, the number of
iterations was changed solely depending on the input image as
shown in Table I.
Table I
Filter parameters and computational time for simulation, phantom and in
vivo studies
(13)
where a constant c manipulates the steepness of the slope while
> TBME-00795-2015.R1<
The FESR method utilizes (6) to suppress speckle noise, and
the scale factor k was set to 1.95. The edge preservation
parameter e in (9) and coherence enhancing factors c in (11)
and c in (12) were set as described in Table I. The illumination
coefficient and reflectance coefficient for contrast
enhancement were automatically estimated as shown in Fig. 6.
All of the methods were implemented in the Matlab
(Mathworks Inc., Natick, MA, USA) environment, and the
computational time in Table I was measured in simulation
studies for a frame of the echogenicity map of 460 447 pixels.
A. Simulation study
To generate B-mode ultrasound RF data, we conducted
pseudo B-mode pulse-echo image generation as reported in [30,
51]. Based on the convolution of the point spread function and
tissue scattering function, we observed the generated speckle
image changes in relation to the axial sound pulse width x and
the lateral beam width y , as illustrated in Fig. 7.
7
B. Phantom study
To apply the five speckle reduction methods to the phantom
experiment, we scanned a commercial gray scale phantom (i.e.,
040 GSE, CIRS INC., Norfolk, VA, USA) with -5 dB/MHz
attenuation and obtained RF beamformed data containing 30
consecutive frames using a commercial ultrasound system (i.e.,
Accuvix V10, Samsung Medison., Korea) with a L5-12 linear
array probe. The RF data of the 30 frames were enveloped and
logarithmically compressed in the post-processing stage to
simulate the B-mode images.
C. In vivo study
The original images of 28 consecutive frames were captured
by a commercial ultrasound system (iU22 xMATRIX, Philips
Healthcare, Bothell, WA, USA) using a C5-1 convex probe and
a L9-3 linear array probe for the abdominal and thyroid areas of
a volunteer. To acquire the original data, we deactivated the
real-time compounding, time-averaging (persistence) and
speckle reduction filtering modes at that time. Then, the images
were transferred onto an external PC and processed offline.
Quantitative and qualitative assessments were performed to
evaluate the improvement in image quality and clinical
feasibility of the five speckle reduction methods.
D. Evaluation metrics
In the simulation study, the visibility errors between a
distorted image and a reference image, as shown in Fig. 10,
were quantified by measuring the speckle regions
signal-to-noise ratio (SSNR), contrast-to-noise ratio (CNR),
edge preservation factor , and mean structural similarity
quality measurement (MSSIM). The SSNR was measured in a
fully formed speckle region (in a homogeneous region) without
any relative structures and can be defined by
Fig. 7. Examples of the simulated envelope-detected image for a single point

target. (a) Single point object. (b)-(f) Computed image for x 1 ~ 5 ,
y 0.5 ~ 4.5 .
SSNR
(14)
where is the mean intensity value and is the standard

deviation in the speckle region of interest. The SSNR was also
measured for the phantom and in vivo studies. In partially
structured regions, to quantify contrast resolution, the CNR was
measured by
CNR
Fig. 8. (a) Echogenicity map and (b) speckled echo image.
Figure 8 shows the echogenicity map and the simulated

pseudo B-mode image in the simulation study. The
echogenicity map consists of six elliptical objects of different
sizes. The mean values of their intensities are 42, 54, 66, 89, 33
and 216, with a background intensity of 66 in gray scale. For
simulation, a center frequency of 10 MHz , sound speed of
1540 m / s , axial pulse width of 0.20 mm , lateral beam width
of 0.15 mm and speckle variance of 0.5 were selected.
| 1 2 |
(15)
12 2 2
where 1 and 12 are the mean and variance of the intensities

of the pixels in a given region of interest (ROI), and 2 and 22
are the mean and variance of the intensities of pixels in the
background region of the same ROI. The CNR was also
measured for the phantom and in vivo studies.
To compare the performance in edge preservation of the
different speckle reduction schemes, ranging from 0 to 1
was computed by [35]
> TBME-00795-2015.R1<
(I I , I I)
(16)
(I I , I I ) (I I, I I)
where I and I are the highpass filtered versions of I and I ,

respectively, that were obtained using a 3 3 standard
approximation of the Laplacian operator. I and I are the
mean intensities of the pixels in the regions I and I ,
respectively, and
I1 , I 2
M N
(( I (i, j ) I (i, j ))
i , j 1
(17)
For quality evaluation based on the degradation of structural

information, the MSSIM from the perspective of image
formation was computed by [52]
MSSIM
1
M
(2 x y C1 )(2 xy C2 )
i 1
2
x
y2 C1 )( x2 y2 C2 )
(18)
for efficient iterative diffusion filtering. In FESR, the mean

square error (MSE) between two adjacent diffusion steps is
adopted as the stopping criterion because the MSE eliminates
the issue of positive and negative values by squaring the errors.
In addition, it tends to place a greater emphasis on larger errors.
This MSE can be defined by
MSE u t
S,D
1
(u i, j, t u (i, j , t 1)) 2
S D i , j 1
(19)
where u(i, j, t ) and u(i, j, t 1) are the filtered pixel (i, j )

values at time t and t 1 , respectively, and S and D are the
numbers of samples and scanlines in the ultrasound images,
respectively.
Figure 9 shows the results of the diffusion experiment, where
the convergence of the proposed FESR method is demonstrated.
As shown in Figs. 9(a)-9(d), the time step ( 1/ 2D , where
D is the number of dimensions) must be smaller than 0.25 to
avoid the divergence of stability in the 2D image [53].
where i and i are the mean and standard deviation of the

pixel intensities at the i th local window, respectively, and Ci is
the constant used to avoid instability. The value of MSSIM
varies from 0 to 1, with unity representing ideal structure
similarity.
In addition to quantitative analysis, a qualitative analysis
performed by three experienced radiologists was gathered and
assessed. In this qualitative analysis, 28 frames of in vivo liver,
kidney and thyroid images from the five speckle reduction
methods were saved in a bitmap format, and all technical
information was removed from the images. Original and the
reconstructed images were organized by pairs, but randomly
presented. Each image was independently reviewed and scored
based on the radiologists subjective preferences with respect to
tissue conspicuity, margin sharpness, artificiality and tissue
contrast. Tissue conspicuity was defined as the general
visibility or clarity of the tissue compared to the adjacent
structures. Margin sharpness was the capacity to properly
visualize the tissue border. Artificiality was the degree of
unnatural depiction. Contrast was defined as the subjective
difference in the echo intensities between the tissue and the
background. All three radiologists were asked to select the
image that showed higher quality and to note the performance
of each image. A score on a 5-point scale was assigned to each
paired dataset from 1 to 5, where a score of 1 indicated the
image to be much worse, a score of 2 indicated worse, a
score of 3 indicated normal, a score of 4 indicated better,
and a score of 5 indicated that the image was much better.
For all statistical analysis, a Wilcoxon signed-rank test was
used and a two-tailed P<.05 was considered to indicate a
statistically significant difference.
E. Stopping time in iterative diffusion filtering in FESR
To integrate the proposed FESR method into ultrasound
imaging systems, it is important to determine the stopping time
Fig. 9. Convergence of MSE in the proposed method according to time step

changes. (a) 0.10 , (b) 0.20 , (c) 0.25 and (d) 0.30 .
Fig. 10. Relative intensity profiles of a nonlinear filter from noisy input signal
for the result of (a) 1, 20, 100 iterations on single scale and (b) 2, 3, 4 and 5
scales with an iteration.
In accordance with these results, most of the MSE values

converge exponentially, with the number of iterations
depending on the different time steps and decomposed scales.
Therefore, the growth in iterations can be stopped by setting a
suitable MSE value and time step in the FESR method, whereas
> TBME-00795-2015.R1<
a diffusion filtering threshold is selected manually, as in [34].
In contrast, iteration and multiscale processing are
performed in the same direction in terms of local operations,
leading to global effects [54]. Figure 10 demonstrates that the
four scales were sufficient to produce an effective result while
100 iterations of a single scale were needed. Thus, multiscale
iterated filtering maintains the high degree of nonlinearity and
adaptability found in the iterated method while simultaneously
using a smaller number of iterations on a single scale. As such,
the decreased computational complexity and cost of the
reduced number of iterations offers substantial advantages in
real-time systems.
VI. EXPERIMENTAL RESULTS AND DISCUSSION
A. Simulation study
9
Table I, its computational time shows its complexity burden
compared with the other methods.
The SSNR, CNR, , and MSSIM values were measured for
30 simulated images with the same speckle level, with the mean
and standard deviation of the measurement values summarized
in Table II. As shown in Table II, all of the speckle reduction
methods are observed to improve the results in terms of all
parameters. Consistent with the visual assessments, all of the
parameter values were higher in the proposed FESR method
than in other speckle reduction methods, i.e., 23.362.55,
1.700.15, 0.160.01 and 0.570.01. Furthermore, the SSNR
value from the FESR method was considerably higher than for
all other methods, i.e., 23.362.55 vs. 18.311.76, 15.521.62,
17.481.55 and 21.423.86, whereas the obtained values of
CNR, and MSSIM were comparable. It must be noted that
the OBNLM method outperforms the NCD, NMWD, and
LPND methods.
SSNR, CNR,
TABLE II
and MSSIM values of 30 simulated images with the same
speckle level
Method
Fig. 11. Simulated pseudo B-mode image and filtered results. (a) Speckled
noisy image. (b)-(f) Images filtered by OSRAD, NMWD, LPNDSF, OBNLM
and FESR, respectively.
Figure 11(a) shows the envelope-detected image from the

echogenicity map used in Fig. 8(a). The example filtered
images obtained from the OSRAD, NMWD, LPNDSF,
OBNLM and FESR methods are shown in Figs. 11(b)-11(f),
respectively. Under visual assessment, the proposed FESR
method in Fig. 11(f) depicts the borders and boundaries of
objects with greater clarity while substantially reducing the
speckle noise in the homogeneous region compared with the
original image in Fig. 11(a) and the other four speckle reduction
methods. Additionally, it is observed to improve the contrast
such that it becomes easier to distinguish the overlapped objects.
The OSRAD image in Fig. 11(b) preserves the edges and
boundaries of objects, but some speckle components are still
present even after processing. The NMWD image in Fig. 11(c)
reduces speckle noise effectively, but it seems to exhibit low
visibility in edge preservation. The LPNDSF image in Fig. 11(d)
shows comparable edge preservation and speckle noise
suppression, but the remaining noise was also enhanced with
artificiality. The OBNLM image in Fig. 11(e) significantly
reduced the speckle noise with no apparent loss of detail, but it
seems to exhibit unnatural patterns because large speckles were
considered structural information. Additionally, as shown in
Parameters
SSNR
CNR
MSSIM
Noisy
8.190.79
1.020.08
0.040.00
0.270.01
OSRAD
18.311.76
1.610.13
0.140.01
0.530.01
NMWD
15.521.62
1.570.13
0.120.00
0.490.01
LPNDSF
17.481.55
1.590.14
0.140.01
0.550.01
OBNLM
21.423.86
1.630.15
0.150.01
0.560.01
FESR
23.362.55
1.700.15
0.160.01
0.570.01
Table III
CNR values for phantom images containing 30 consecutive frames
Method
CNR
ROI 1
ROI 2
ROI 3
ROI 4
Noisy
6.310.03
0.430.00
2.300.01
3.160.01
OSRAD
9.000.06
0.690.01
3.480.01
4.510.02
NMWD
9.780.06
0.680.01
3.610.01
4.580.02
LPNDSF
8.670.04
0.860.01
3.540.02
4.530.01
OBNLM
9.220.06
0.880.01
3.900.01
5.170.01
FESR
10.700.06
0.990.01
4.250.02
5.490.03
B. Phantom study
Fig. 12. Phantom image and its filtered results. (a) Original noisy image. (b)-(f)
Image filtered by OSRAD, NMWD, LPNDSF, OBNLM and FESR methods.
> TBME-00795-2015.R1<
10
Fig. 13. Ultrasound in vivo abdominal liver image and its filtered results. (a) Conventional noisy image. (b)-(f) Image filtered by OSRAD, NMWD, LPNDSF,
OBNLM and FESR, respectively.
Figure 12(a) shows the original phantom image, and the

example filtered images using the OSRAD, NMWD, LPNDSF,
OBNLM and FESR methods are shown in Figs. 12(b)-12(f).
The proposed FESR method in Fig. 12(f) outperforms the other
speckle reduction methods in terms of effective speckle
suppression and texture preservation with conspicuity. In
contrast, the OSRAD method in Fig. 12(b) yields insufficient
speckle reduction, whereas the NMWD method in Fig. 12(c)
shows strong speckle suppression but also blurs the boundaries
of the defined target. For the LPNDSF and OBNLM methods

used in Figs. 12(d)-12(e), both filters substantially reduce
speckle noise but produce an artificial appearance through the
unwanted enhancement of the remaining speckle pattern.
For quantitative analysis, we measured the CNR values of 4
ROIs, which are 4 contrast regions relative to an anechoic cyst
(ROI 1) and gray scale targets of +3, -3, -6 dB (ROI 2-4),
respectively, as shown in Fig. 12(a). The measured CNR values
with their mean and standard deviation are summarized in
Table III for the 30 consecutive frames. As listed in Table III,
> TBME-00795-2015.R1<
the FESR method shows the highest CNR values compared
with the other four speckle reduction methods for all four ROIs,
i.e., 10.700.06, 0.990.01, 4.250.02 and 5.490.03,
respectively. Among the remaining four methods, the OBNLM
method shows the highest CNR values for all ROIs.
C. In vivo study
Figure 13(a) shows the original ultrasound B-mode image
that was acquired from the liver of a volunteer, and the
processed images using the five speckle reduction methods are
shown in Figs. 13(b)-13(f). The second row of Fig. 13
represents zoom-in images of a solid box in Fig. 13(a). Under
visual assessment, the FESR method shows uniform reduction
in speckle in the liver parenchyma areas with improved tissue
differentiation. In addition, the borders and boundaries of the
hepatic vessels (i.e., A in Fig. 13(a)) in the FESR image are
more clearly depicted than in the images processed by the other
speckle reduction methods. Additionally, the FESR method
significantly suppresses acoustic clutter, which produces
spurious echoes within the gall bladder (i.e., region B in Fig.
13(a)) near the liver. The other cystic contents and tissue
margins in the FESR image are shown to be a clearer
visualization without a loss of underlying diagnostic details
(e.g., the region C of Fig. 13(a)).
In contrast, the NMWD method shows higher speckle
reduction but produces blurred edges and boundaries in the
hepatic vessels, similar to the results found in the simulation
and phantom experiments. The LPNDSF and OBNLM
methods also suffer from artificial speckle patterns in the liver
parenchyma areas, although they preserve local coherent edges.
The OSRAD method substantially suppresses speckle noise
and maintains the structures of the hepatic vessel and the gall
bladder. However, its image suffers from a faded appearance.
11
standard deviations of the OBNLM and FESR methods are
greater than those for the OSRAD, NMWD and LPNDSF
methods because a relatively large speckle pattern can also be
processed as texture and can thus be erroneously enhanced by
the feature enhancement filtering used in the FESR method. In
addition, the rejected outlier is shown in all five results due to
the presence of unwanted signals, which was emphasized in the
filtering process.
In the case of the abdominal kidney, as represented by Fig.
15, FESR substantially reduces the appearance of speckle and
increases the conspicuity of low-contrast tissue compared with
the other speckle reduction methods. Additionally, the FESR
image of the thyroid case shows an improved delineation of
linear striation patterns from fat and muscle with continuity, as
shown in Fig. 16. Furthermore, the internal architecture of the
benign nodule is depicted with clear visualization and thus
provides better detectability.
The subjective preference scores for five speckle reduction
methods given by the radiologists are shown in Table IV. For
the overall parameters, the three radiologists mostly indicated a
score of better for the FESR method in three different cases,
which were compared with worse or normal for other
methods. In particular, the abdominal liver obtained
significantly different scores in the FESR method compared to
other methods, as shown in Table IV.
For margin sharpness, the FESR method shows no
statistically significant difference for the abdominal kidney (i.e.,
p=0.065) and thyroid (i.e., p=1) compared with the OBNLM
method, though all three radiologists rated it with a higher than
average score. For artificiality, the radiologists scored the
LPNDSF and OBNLM methods relatively low due to the
unnatural speckle patterns, whereas they showed considerable
results for quantitative assessment. Additionally, FESR did not
show statistical significance for thyroid images compared with
the OSRAD method. For tissue contrast, all three radiologists
indicated a score of better for the FESR method in three
different cases, which were compared with worse or normal
for the other methods. It must be noted that, in the present study,
inter- or intra-rater agreement between radiologists was not
evaluated.
VII. CONCLUSION
Fig. 14. The box plots show the distribution of SSNR and CNR in the
conventional noisy image and the filtered images processed by the OSRAD,
NMWD, LPNDSF, OBNLM and FESR methods. (a) SSNR values. (b) CNR
values. The * represents outliers.
The SSNR and CNR values were measured for 30

consecutive frames and are illustrated in the box plot in Fig. 14.
With the proposed FESR method, the mean SSNR value is
significantly higher than that with the OSRAD, NMWD,
LPNDSF and OBNLM methods (i.e., 7.620.78 vs. 5.310.48,
5.320.52, 5.920.60, 6.091.02, respectively, p<0.0001).
Similar to SSNR, the mean CNR value of the proposed FESR
method is again significantly higher than for the other methods
(i.e., 7.631.38 vs. 4.830.44, 4.840.50, 5.570.66, 5.951.08,
respectively, p<0.0001). However, as shown in Fig. 14, the
Effective speckle reduction in ultrasound B-mode imaging is

important for enhancing image quality and improving accuracy
in image analysis and interpretation. In this paper, a new
feature-enhanced speckle reduction (FESR) method based on
multiscale analysis and feature enhancement filtering is
presented to improve ultrasound B-mode imaging. To separate
true clinical features from noise and artifacts, we employ
sub-band images using multiscale analysis and a synthesis
method (i.e., Laplacian pyramid). Then, the identified noise is
suppressed by robust anisotropic diffusion filtering, and the
extracted features are selectively emphasized using appropriate
edge, coherence and contrast enhancement filtering from fine to
coarse scales. For real application, the stopping time of the
iterative diffusion filtering is automatically determined from
the estimation of the MSE between two adjacent diffusion
> TBME-00795-2015.R1<
12
Fig. 15. Ultrasound image of in vivo abdominal kidney and its filtered results. (a) Conventional noisy image. (b)-(f) Image filtered by OSRAD, NMWD, LPNDSF,
OBNLM and FESR methods, respectively.
Fig. 16. Ultrasound image of in vivo abdominal kidney and its filtered results. (a) Conventional noisy image. (b)-(f) Image filtered by OSRAD, NMWD, LPNDSF,
OBNLM and FESR methods, respectively.
> TBME-00795-2015.R1<
13
steps. To quantify the performance of the proposed method,

both simulation and phantom studies were implemented, and in
vivo images of the three cases were reviewed by three
experienced radiologists to determine the potential clinical
value of this method. From the simulation, phantom and in vivo
experiments, the FESR algorithm is proven to reduce speckle or
noise while preserving contrast resolution and tissue texture
and is observed to enhance low echogenicity by improving
border definition and continuity. Additionally, the proposed

method has the potential to assist in segmentation techniques
and CAD (computer-aided detection) for image analysis.
Moreover, it can improve clinical results and diagnostic
accuracy through better visualization with clearer tissue and
lesion structures and higher contrast resolution.
Table IV
Subjective preference scores of five speckle reduction methods in three different tissues
[3]
REFERENCES
[1]
[2]
J. Powers and F. Kremkau, Medical ultrasound systems, Interface

Focus, vol. 1, pp. 477-489, 2011.W.-K. Chen, Linear Networks and
Systems. Belmont, CA: Wadsworth, 1993, pp. 123135.
J. H. Chang et al., Proof of concept: in vitro measurement of correlation
between radiodensity and ultrasound echo response of ovine vertebral
bodies, Ultrasonics, vol. 51, no. 3, pp. 253-257, 2011.
[4]
[5]
[6]
C. B. Burckhardt, Speckle in ultrasound B-mode scans, IEEE Trans.

Sonics Ultrason., vol. 25, no. 1, pp. 1-6, Jan. 1987.
R. F. Wagner et al., Statistics of speckle in ultrasound B-scans, IEEE
Trans. Sonics Ultrason., vol. 30, no. 3, pp.156-163.
B. D. Fornage, Ultrasound of the breast, Ultrasound Quarterly, vol. 11,
no. 1, pp. 1-40, 1993.
D. A. Carpenter et al., A multimode real time scanner, Ultrasound Med.
Biol., vol. 6, no. 3, pp. 279-284, 1980.
> TBME-00795-2015.R1<
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
P. A. Magnin et al., O. T. von Ramm and F. L. Thurstone, Frequency

compounding for speckle contrast reduction in phased array images,
Ultrason. Imag., vol. 4, no. 3, pp. 267-281, 1982.
G. E. Trahey et al., A quantitative approach to speckle reduction via
frequency compounding, Ultrason. Imag., vol. 8, no. 3, pp. 151-164,
1986.
S. M. Gehlbach and F. G. Sommer, Frequency diversity speckle
processing, Ultrason. Imag., vol. 9, no. 2, pp. 92-105, 1987.
R. L. Galloway et al., A frequency diversity process for speckle
reduction in real-time ultrasonic images, IEEE Trans. Ultrason.,
Ferroelectr., Freq. Control, vol. 35, no. 1, pp. 45-49, 1988.
G. Cincotti et al., Frequency decomposition and compounding of
ultrasound medical images with wavelet packets, IEEE Trans. Med.
Imag., vol. 20, no. 8, pp. 764-771, 2001.
Y. Erez et al., Space variant ultrasound frequency com-pounding based
on noise characteristics, Ultrasound in Med. & Bio., vol. 34, no. 6, pp.
981-1000, 2008.
J. R. Sanchez and M. L. Oelze, An ultrasonic imaging speckle
suppression and contrast-enhancement technique by means of frequency
compounding and coded excitation, IEEE Trans. Ultrason., Ferroelectr.,
Freq. Control, vol. 56, no. 7, pp. 1327-1339, 2009.
J. H. Chang et al., Frequency compounding imaging with a
high-frequency dual element transducer, Ultrasonics, vol. 50, no. 4-5, pp.
453-457, 2010.
T. L. Szabo, Diagnostic Ultrasound Imaging: Inside Out, Academic
press. Burlington. MA. 2004.
I. S. Song et al., Adaptive frequency compounding for speckle reduction,
in Proc. IEEE Intern. Ultrason. Symp., pp. 1435-8, 2011.
C. H. Yoon et al., Frequency equalized compounding for effective
speckle reduction in medical ultrasound imaging, Biomed. Signal
Proces., vol. 8, no. 6, pp. 876-887, 2013.
M. Berson et al., Compound scanning with an electrically steered beam,
Ultrasonic Imaging, vol. 3, no. 3, pp. 303-308, 1981.
S. K. Jespersen et al., Multi-angle compound imaging, Ultrasonic
Imaging, vol. 20, no.2, pp. 81-102, 1998.
M. ODonnell and S. D. Silverstein, Optimum displacement for
compound image generation in medical ultrasound, IEEE Trans.
Ultrason., Ferroelectr., Freq. Control, vol. 35, no. 4, pp. 470-476, 1988.
J. Opretzka et al., A high-frequency ultrasound imaging sys-tem
combining limited-angle spatial compounding and model-based synthetic
aperture focusing, IEEE Trans. Ultrason., Ferroelectr., Freq. Control,
vol. 58, no.7, pp. 1355-1365, 2011.
J. C. Bamber, Adaptive filtering for reduction of speckle in ultrasound
pulse-echo images, Ultrasonics, vol. 24, no. 1, pp. 41-44, 1986.
P. Juang and M. Wu, Ultrasound speckle image process using wiener
pseudo-inverse filtering, in Proc. IEEE Conference of Ind. Elec. Society,
pp. 2446-2449, 2007.
V. Dutt and J. F. Greenleaf, Adaptive speckle reduction filter for log
compressed B-scan image, IEEE Trans. Med. Imag., vol. 15, no. 6, pp.
802-813, 1996.
J. I. Koo and S. B. Park, Speckle reduction with edge preservation in
medical ultrasonic images using a homogeneous region growing mean
filter (HRGMF) Ultrasonic Imaging, vol. 13, no. 3, pp. 211-237, 1991.
M. Karaman et al., An adaptive speckle suppression filter for medical
ultrasonic imaging, IEEE Trans. Med. Imag., vol. 14, no. 2, pp. 283-292,
1995.
Y. Chen et al., Aggressive region growing for speckle reduction in
ultrasound images, Pattern Recogn Lett., vol. 24, no. 4-5, pp. 677-691,
2003.
H. C. Huang et al., Adaptive ultrasonic speckle reduction based on the
slope-facet model, Ultrasound in Med. & Bio., vol. 29, no. 8, pp.
1161-1175, 2003.
P. C. Tay et al., Ultrasound despeckling for contrast enhancement,
IEEE Trans. Image Process., vol. 19, no. 7, pp. 1847-1860, 2010.
Y. Yu and S. T. Acton, Speckle reducing anisotropic diffusion, IEEE
Trans. Image Process., vol. 11, no. 11, pp. 1260-1270, 2002.
K. Krissian et al., Oriented speckle reducing anisotropic diffusion,
IEEE Trans. Image Process., vol. 16, no. 5, pp. 1412-1424, May. 2007.
K. Z. Abd-Elmoniem et al., Real-time speckle reduction and coherence
enhancement in ultrasound imaging via nonlinear anisotropic diffusion,
IEEE Trans. Biomed. Eng., vol. 49, no. 9, pp. 997-1014, 2002.
X. Zong et al., Speckle reduction and contrast enhancement of
echocardiograms via multiscale nonlinear processing, IEEE Trans. Med.
Imag., vol. 17, no. 4, pp. 532-540, 1998.
14
[34] Y. Yue et al., Nonlinear multiscale wavelet diffusion for speckle
suppression and edge enhancement in ultrasound images, IEEE Trans.
Med. Imag., vol. 25, no. 3, pp. 297-311, 2006.
[35] F. Zhang et al., Nonlinear diffusion in laplacian pyramid domain for
ultrasonic speckle reduction, IEEE Trans. Med. Imag., vol. 26, no. 2, pp.
200-211, 2007.
[36] F. Zhang et al., Multiscale nonlinear diffusion and shock filter for
ultrasound image enhancement, in Proc. 2006 IEEE Computer Society.
Computer Vision and Pattern Recognition, pp. 1972-1977, 2006.
[37] P. Coupe et al., Nonlocal means-based speckle filtering for ultrasound
images, IEEE Trans. Image Process., vol. 18, no. 10, pp. 2221-2229,
2009.
[38] Y. Gu et al., Ultrasound echocardiography despeckling with non-local
means time series filter, Neurocomputing, vol. 124, pp. 120-130, 2013.
[39] Y. Zhan et al., Nonlocal means method using weight refining for
despeckling of ultrasound images, Signal Processing, vol. 103, pp.
201-203, 2014.
[40] V. Dutt, Statistical analysis of ultrasound echo envelope, Ph. D.
dissertation, Mayo Graduate School, Rochester, MN, 1995.
[41] P. Perona and J. Malik, Scale-space and edge detection using anisotropic
diffusion, IEEE Trans. on Pattern Anal. Mach. Intell., vol. 12, no. 7, pp.
629-639, Jul 1990.
[42] P. J. Burt, and E. H. Adelson, The Laplacian pyramid as a compact
image code, IEEE Trans. on Commun., vol. COM-31, No. 4, Apr 1983.
[43] F. Sattar et al., Image enhancement based on a nonlinear multiscale
method, IEEE Trans. Image Process., vol. 6, no. 6, pp. 888-895, 1997.
[44] M. J. Black et al., Robust anisotropic diffusion, IEEE Trans. Image
Process., vol. 7, pp. 421-432, Mar 1998.
[45] R. W. Brockett and P. Maragos, Evolution equations for
continuous-scale morphology, in Proc. IEEE ICASSP, pp. 125-128,
1992.
[46] J. Weickert, Multiscale texture enhancement, Computer analysis of
images and patterns, Lecture notes in Computer science, vol. 970, pp.
230-237, 1995.
[47] A. Oppenheim et al., Nonlinear filtering of multiplied and convolved
signals, IEEE Trans. Audio and Electro., vol. 56, issue 8, pp. 1264-1291,
Sep. 1968.
[48] R. Fattal et al., Gradient domain high dynamic range compression,
Proceedings of the 29th annual conference on computer graphics and
interactive techniques, vol. 21, no. 3, pp. 249-256, 2002.
[49] N. Otsu, A threshold selection method from gray-level histogram, IEEE
Trans. Systems, Man and Cybem., vol. 9, issue 1, pp. 62-66, 1979.
[50] S. Finn et al., Echocardiographic speckle reduction comparison, IEEE
Trans. Ultrason. Ferroelect. Freq. Contr., vol. 58, no. 1, pp. 82-101, Jan.
2011.
[51] J. C. Bamber and R. J. Dickinson, Ultrasonic B-scanning: a computer
simulation, Phys. Med. Biol., vol. 25, no. 3, pp. 463-479, 1980.
[52] Z. Wang et al., Image quality assessment: From error visibility to
structural similarity, IEEE Trans. Image Process., vol. 13, no. 4, pp.
1-14, Apr 2004.
[53] J. Weickert et al., Efficient and reliable schemes for nonlinear diffusion
filtering, IEEE Trans. Image Process., vol. 7, no. 3, pp. 398-410, Mar
1998.
[54] P. Mrzek et al., On iterations and scales of nonlinear filters, in Proc.
Computer Vision Winter Workshop, O. Drbohlav, Ed., Valtice, Czech
Republic, pp. 61-66, Feb. 2003.

Christ Cse

Hochgeladen von

Dokumentinformationen

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

Christ Cse

Hochgeladen von

Copyright:

Verfügbare Formate

This article has been accepted for publication in a future issue of this journal, but has not been

A New Feature-enhanced Speckle Reduction

Abstract Goal: Effective speckle reduction in ultrasound

compared with the OBNLM method in margin sharpness for the

quality in ultrasound B-mode imaging has been

where ( x, y) is the received signal at a spatial location of

value k is a crucial factor as a diffusion coefficient

and wa ( x, y) represent the multiplicative noise and additive

Generally, ultrasound imaging systems have a

log ( x, y) log f ( x, y) log wm ( x, y)

Fig. 1. Graph of the nonlinear diffusivity coefficient c1 ( x) as the gradient

B. Hierarchical Laplacian pyramid model

Thus, a speckle pattern can be modeled as an additive white

Fig. 2. Hierarchical Laplacian pyramid transform.

where is the gradient operator, div is the divergence

To extract features or structures from noise, the input image

Ll (n) Gl (n) E Gl 1 (n)

where R and E indicate the decimation and interpolation of the

if too many scales are decomposed, highpass information (e.g.,

IV. FEATURE-ENHANCED SPECKLE REDUCTION ALGORITHM

Fig. 4. The overall block diagram of the proposed FESR method.

In FESR, all of the decomposed layers (i.e., L0 (n) L3 (n) in

Figure 3 illustrates an example of the decomposed

Laplacian pyramid in Fig. 4, edge enhancement is performed by

of the proposed algorithm. In the FESR method, edge-sensitive

where is a scale factor. To analyze the performance of a

-function and a diffusivity function c x is shown in Fig. 5.

where the subscripts represent partial derivatives, and

where D is the structure tensor based on PDE and is the

version of a morphological filter where the second directional

Based on the optimization criterion, we can obtain the robust

where | G u | is the gradient of a Gaussian lowpass filtered

C. Directional coherence enhancement

where K is a convolution kernel defined by

According to the eigenvalue decomposition, the structure

where the eigenvectors 1 , 2 and the eigenvalues 1 , 2

the filter is transiting between and , ( ) is the distance

edges are given by 1 2 0 , and corners yield 1 2 0 .

Fig. 6. A modified homomorphic filtering system for contrast enhancement.

In detail, the gray level image that is despeckled by nonlinear

where a constant c manipulates the steepness of the slope while

Fig. 7. Examples of the simulated envelope-detected image for a single point

where is the mean intensity value and is the standard

Fig. 8. (a) Echogenicity map and (b) speckled echo image.

Figure 8 shows the echogenicity map and the simulated

where 1 and 12 are the mean and variance of the intensities

where I and I are the highpass filtered versions of I and I ,

For quality evaluation based on the degradation of structural

for efficient iterative diffusion filtering. In FESR, the mean

where u(i, j, t ) and u(i, j, t 1) are the filtered pixel (i, j )

where i and i are the mean and standard deviation of the

Fig. 9. Convergence of MSE in the proposed method according to time step

In accordance with these results, most of the MSE values

Figure 11(a) shows the envelope-detected image from the

Figure 12(a) shows the original phantom image, and the

of the defined target. For the LPNDSF and OBNLM methods

The SSNR and CNR values were measured for 30

Effective speckle reduction in ultrasound B-mode imaging is

steps. To quantify the performance of the proposed method,

border definition and continuity. Additionally, the proposed

J. Powers and F. Kremkau, Medical ultrasound systems, Interface

C. B. Burckhardt, Speckle in ultrasound B-mode scans, IEEE Trans.

P. A. Magnin et al., O. T. von Ramm and F. L. Thurstone, Frequency

Das könnte Ihnen auch gefallen