Accepted Manuscript

A hybrid algorithm for speckle noise reduction of ultrasound images

Karamjeet Singh , Sukhjeet Kaur Ranade , Chandan Singh

PII: S0169-2607(17)30197-9
DOI: 10.1016/j.cmpb.2017.06.009
Reference: COMM 4438

To appear in: Computer Methods and Programs in Biomedicine

Received date: 20 February 2017

Revised date: 30 May 2017
Accepted date: 20 June 2017

Please cite this article as: Karamjeet Singh , Sukhjeet Kaur Ranade , Chandan Singh , A hybrid
algorithm for speckle noise reduction of ultrasound images , Computer Methods and Programs in
Biomedicine (2017), doi: 10.1016/j.cmpb.2017.06.009

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service
to our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please
note that during the production process errors may be discovered which could affect the content, and
all legal disclaimers that apply to the journal pertain.
 ACCEPTED MANUSCRIPT

Highlights

 A hybrid speckle noise reduction algorithm is proposed.

T
 It combines the advantages of Guided, Bilateral and rotation invariant NLM.
 Achieve an optimal balance between speckle noise reduction and feature preservation by

IP
proposed hybrid algorithm.
 The proposed hybrid algorithm is validated by both quantitative and qualitative measures.

CR
 The hybrid algorithm provides excellent denoising performance on synthetic, simulated
and real US images.
 Experimental results prove that the hybrid algorithm have better denoising capability than
recently proposed speckle noise reduction filters.

US
AN
M
ED
PT
CE
AC

1
 ACCEPTED MANUSCRIPT

Title: A hybrid algorithm for speckle noise reduction of ultrasound images

T
Author Name: Karamjeet Singh

IP
Research Scholar, Department of Computer Science, Punjabi University, Patiala-147002, India.

CR
E-mail: karampup@gmail.com, M: 91-9888009027

Author Name: Sukhjeet Kaur Ranade

US
Associate Professor, Department of Computer Science, Punjabi University, Patiala-147002, India.
AN
Author Name: Chandan Singh

Professor, Department of Computer Science, Punjabi University, Patiala-147002, India.

M

E-mail: chandan.csp@gmail.com Phone: 91-175-3046316, M: 91-9872043209, Fax: 0175-

3046313
ED

Corresponding Author Name: Karamjeet Singh

PT

Research Scholar, Department of Computer Science, Punjabi University, Patiala-147002, India.

E-mail: karampup@gmail.com, M: 91-9888009027

CE
AC

2
 ACCEPTED MANUSCRIPT

A hybrid algorithm for speckle noise reduction of ultrasound images

Karamjeet Singh, Sukhjeet Kaur Ranade, Chandan Singh

T
Department of Computer Science, Punjabi University, Patiala-147002, India.

IP
Abstract
Background and objective

CR
Medical images are contaminated by multiplicative speckle noise which significantly reduce the
contrast of ultrasound images and creates a negative effect on various image interpretation tasks.
In this paper, we proposed a hybrid denoising approach which collaborate the both local and

US
nonlocal information in an efficient manner. The proposed hybrid algorithm consist of three stages
in which at first stage the use of local statistics in the form of guided filter is used to reduce the
effect of speckle noise initially. Then, an improved speckle reducing bilateral filter (SRBF) is
AN
developed to further reduce the speckle noise from the medical images. Finally, to reconstruct the
diffused edges we have used the efficient post-processing technique which jointly considered the
advantages of both bilateral and nonlocal mean (NLM) filter for the attenuation of speckle noise
M

efficiently.
Methods
The performance of proposed hybrid algorithm is evaluated on synthetic, simulated and real
ED

ultrasound images. The experiments conducted on various test images demonstrate that our
proposed hybrid approach outperforms the various traditional speckle reduction approaches
included recently proposed NLM and optimized Bayesian-based NLM.
PT

Results
The results of various quantitative, qualitative measures and by visual inspection of denoise
CE

synthetic and real ultrasound images demonstrate that the proposed hybrid algorithm have strong
denoising capability and able to preserve the fine image details such as edge of a lesion better than
previously developed methods for speckle noise reduction.
AC

Conclusions
The denoising and edge preserving capability of hybrid algorithm is far better than existing
traditional and recently proposed SR filters. The success of proposed algorithm would help in
building the lay foundation for inventing the hybrid algorithms for denoising of ultrasound images.

Keywords: Speckle noise, Rotation invariance, Edge preservation, Nonlocal means.

3
 ACCEPTED MANUSCRIPT

1. Introduction
In today scenario, medical image modality like ultrasound imaging is one of the prominent
diagnostic techniques to view the internal anatomy (such as abdomen, liver, kidney and

T
musculoskeletal) of a human body. But, ultrasound (US) images are generally corrupted with

IP
speckle noise that appears as granular pattern due to constructive and destructive coherent
interferences of backscattered echoes from the scatters [1-2]. This phenomenon degrades the

CR
contrast resolution of US images which makes the extraction of significant information from the
speckle corrupted US images as one of the complicated tasks. In addition, it also reduces the
accuracy of various other medical image processing tasks such as feature extraction, segmentation,

preprocessing steps for the analysis and processing of ultrasound images. US

registration and classification. Thus, speckle noise reduction becomes one of the necessary

Prior to the various image processing applications, speckle noise needs to be removed but without
AN
affecting significant feature details. Speckle noise can be removed by compounding [7-10], post-
processing techniques or by both techniques [11]. In compounding approach, images are acquired
from different transducers which minimize the effect of noise. But, the cost of compounding
M

approach is very expensive. Therefore, in literature, many researchers have provided the various
post-processing methods which can further be categorized into following two classes:
ED

(a) spatial-based approaches that are applied directly to the original image and (b) transform-based
approaches which firstly transform the image into the frequency domain and then introduce the
denoising process to improve its quality. Numerous methods have been developed to reduce the
PT

effect of speckle noise in ultrasound images which mostly includes local statistics-based methods
such as median [11-12], Lee [14], Frost [15], Kaun [16-17], speckle reducing anisotropic diffusion
(SRAD) [18], oriented SRAD (OSRAD) [19], squeeze box filter (SBF) [20-21], bilateral [22] and
CE

guided [23] filter.

Among various filters, the median [12-13] filter is a useful tool to reduce the impulsive noise. It is
also used for the reduction of speckle noise from US images. This filter replaces the middle pixel
AC

value of its neighboring window with its median value by sorting it either in increasing or
decreasing order. The Lee filter [14] is based upon the minimization of mean-square error
(MMSE) by considering a weighted average of sub-regions created at each pixel locations. The
relationship between smoothing performed by Lee filter and variance of a region is inversely
proportional to each other. It means smoothing operation is not performed, when variance of a

4
 ACCEPTED MANUSCRIPT

region is quite high, especially near the edges. On the other hand, an effective smoothing operation
is executed by Lee filter to attenuate speckle noise from the low variance regions (such as
homogenous or flat regions). An adaptive Frost filtration [15] technique is also used to maintain

T
the appropriate balance between smoothing and preservation of fine edge structures of different

IP
regions. This balance is achieved by using an exponential damped kernel which adapts according
to different regions by exploiting the local statistics (i.e. mean and variance) of neighboring

CR
window created at each pixel. To preserve edges, Kaun filter [16-17] is the appropriate filter to
use,and it is based on the conversion of multiplicative noise model into signal dependent additive
noise model. The mathematical function utilized in the Kaun filter has the same form as Lee filter

US
have but with a different weighting function. But, the main disadvantage about these traditional
speckle reduction (SR) filters are, they eradicate the weak and diffused edges which make the
ultrasound images harder to interpret especially for doctors and physicians. Therefore, it is more
AN
difficult to provide appropriate diagnosis to a patient suffering from a critical disease. Another
efficient and widely used filter in the spatial domain is the speckle reducing anisotropic diffusion
(SRAD) filter [18] which removes the speckle noise efficiently and preserve all the essential
M

information. This method is based upon the minimization of MMSE similar to that of Lee and
Frost filters. The SRAD filter uses the second order partial differential equations in its iteration-
ED

based diffusion process which serves as an edge detector and able to preserve the useful
information about the edge prone areas. The oriented SRAD (OSRAD) filter [19] improves the
denoising capability of SRAD filter by exploiting the local directional variance of image pixels
PT

present in the neighboring window. Both SRAD and OSRAD filters are iterative in nature, due to
iteration-based procedures some fine details and edges are completely vanished from the images
which is one its major drawback. To overcome this drawback, some authors suggest the use of
CE

iteration-based denoising method such as squeeze box filter (SBF) [20-21], classical bilateral filter
(CBF) [22] and Guided filter (GF) [23] as a preprocessing step to reduce the effect of speckle
noise from US images. The SBF is an iterative scheme proposed by Tay et al. [20-21] which
AC

remove the outliers and reduces the local variance of image in its each iteration. The CBF [22]
have strong edge preservation capability, but it cannot be directly applied to speckle corrupt US
images, therefore it is used as a pre-processing step to minimize the effect of speckle noise. He et
al. [23] proposed the concept of guided filter which exploits the local second order statistics and
demonstrates better denoising capability as compared to traditional SR filters. Therefore, Zhang et

5
 ACCEPTED MANUSCRIPT

al. [29] efficiently used the advantages offered by guided filter in its wavelet-based algorithm.
Since, quantity of speckle noise is quite less in low frequency sub-bands, therefore it is a suitable
condition to use guided filter for the reduction of speckle noise from the low frequency sub-band

T
coefficients. Same authors have also used BF to reduce speckle from low frequency sub-bands, but

IP
this approach does not achieve an optimal balance between speckle suppression and feature
preservation. Some of other significant wavelet-based approaches for speckle denoising can be

CR
found in [24-30].
Recently, nonlocal means (NLM)-based [31-38] approaches have also been popularly used due
their effective denoising capabilities. Baudes et al. [31] introduced revolutionary approach that

US
efficiently utilizes the patterns and self-similarities present in image in form of local patches rather
than using single image pixel information. These local patches characterize the structural
information such as image edges and texture details in better way than the information contained
AN
in a single image pixel. In spite of its advantages, the classical nonlocal means (CNLM) filter is
not designed to deal with multiplicative (speckle) noise model. Therefore, theory of CNLM is
extended by the various authors [32-38] to make it suitable for speckle noise suppression. First of
M

all, Coupe et al. [33] proposed an optimized Bayesian-based nonlocal mean (OBNLM) filter for
speckle noise reduction. In their proposed approach, they utilized the block-wise NLM approach
ED

and then apply Pearson distance measure to compute the similarity between the two patches. The
performance of OBNLM is far better than the CNLM in term of its denoising capability and edge
preservation. Later on, Guo et al. [34] proposed two stage filtering scheme in which at first stage
PT

they apply maximum likelihood-based local image denoising filter, and it is based upon Rayleigh
statistics. In the second stage, they apply the CNLM filter to reconstruct the broken edges and
remove the left behind speckle noise from synthetic and real ultrasound images. Zhan et al. [35]
CE

used the signal dependent speckle noise model to derive the novel weight refining approach based
on Bayesian framework. The newly calculated weights are improved through an iterative
procedure in a lower dimensional subspace using principal component analysis. Radlak and
AC

Smolka [36] proposed an adaptive nonlocal means filter which is able to denoised the speckle
corrupted US images. Recently, Sudeep et al. [37] proposed novel despeckling approach based on
the computing of different parameters of speckle noise statistics by using maximum likelihood
estimation approach. The disadvantages of aforementioned NLM-based approaches are that, either
they generate over smooth images or either speckle noise is still visible near the edges. Hence, it is

6
 ACCEPTED MANUSCRIPT

highly desirable to design a hybrid algorithm which is able to maintain an optimal balance
between speckle suppression, feature preservation and having less computational complexity. In
this paper, we present a hybrid algorithm which offer many advantages such as low computational

T
complexity, free from iterative procedure (because, iterative process usually removes the fine

IP
edges details), able to maintain an optimal balance between denoising process and edges
preservation, and provide less-reduced contrast in the denoised images. The proposed hybrid

CR
algorithm is consist of three stages in which at first stage, guided filter is applied to reduce the
effect of speckle noise initially, and after that square-chord-based distance measure is used in
bilateral filter to deal with multiplicative noise model and in last stage, rotation invariance is

US
attained by combining the advantages of BF and CNLM which further increases the denoising
capability of the hybrid algorithm. Extensive experiments are conducted on synthetic, B-mode,
simulated and as well on real ultrasound images which demonstrate the superiority of the proposed
AN
hybrid algorithm over the other eleven state-of-the-art despeckling method in term of speckle
noise attenuation and edges preservation.
The reminder of this paper is organized as follows: speckle noise model is introduced in a Section
M

2. Material and methods which include a brief explanation of guided filter, bilateral filter and
rotation invariant bilateral and nonlocal means filter are represented in Section 3. In Section 4,
ED

extensive experiments and comparative performance analysis of traditional SR approaches with

proposed hybrid algorithm is presented. Finally, the conclusion is presented in Section 5.

2. Mathematical model of speckle noise

PT

It is imperative to understand the mathematical model of a speckle noise before any despeckling
algorithm is invented. The extensive experimental studies done by various researchers suggests
CE

that the noise distribution in US images is signal dependent and approximately model as
multiplicative noise whose mathematical representation is defined as follows [1-6].
( ) ( ) ( ) ( ) (1)
AC

where are the pixel position, ( ) is the original noise free image, is a factor which
depends on the hardware of the ultrasound devices, ( ) is a zero-mean Gaussian noise
having variance . The behavior of the factor depends upon the type of the raw ultrasound data
used. In our study, we set equal to 0.5 for B-mode and real ultrasound images and equal to 1 for
synthetic images. Also, Loupes et al. [12] verified that when the factor is set equal to 0.5 then

7
 ACCEPTED MANUSCRIPT

the model defined by Eq. (1) is most suitable for the real ultrasound images or even better than
from the Rayleigh model defined for speckle noise. Therefore, in our all experiments, we have
considered the above mentioned model to deal with speckle noise corrupted US images.

T
3. Material and methods

IP
3.1 Guided filter (GF): The fundamental assumption of the guided filter [23] is a local linear
model between the guidance I and the filter output q. We assume that q is a linear transform of I in

CR
a window centered at the pixel k.
(2)
where ( ) are some linear coefficients assumed to be constant in . We use a square
window of radius . This local linear model ensures that
because US
has an edge only if
. To determine the linear coefficients, we seek a solution to Eq. (2) that
minimizes the difference between and the filter input
have an edge,

. Specifically, we minimize the following

AN
cost function in the window:

( ) ∑ (( ) )
(3)
M

where is a regularization parameter preventing from being too large.

∑ ̅
| | (4)
ED

(5)
̅
PT

where, and are the mean and variance of in , ∣ ∣ is the number of pixels in , and

̅ ∑ is the mean of in So, after computing ( ) for all patches in the

| |

image, we compute the filter output as follows.

CE

∑
| | (6)
AC

̅ ̅

where ̅ | |
∑ and ̅ | |
∑ .

3.2 Classical bilateral filter (CBF): The CBF filter is initially developed by Tomasi and
Manduchi [22] in which they replaced the central image pixel value with the weighted average of

8
 ACCEPTED MANUSCRIPT

its neighbors where the weight computations are dependent on the both the spatial and intensity
distances. The mathematical formulation for the standard bilateral filter is expressed as follows.
∑ ( ) ( )

T
( )
( ) (7)
∑ ( ) ( )

IP
where ( )is the neighborhood created at each image pixel ( ) and ( ) are the
weights computed on the basis of radiometric and geometric distance between pixels. Therefore,

CR
the formulation of weight computations in the CBF is expressed as follows.

( ( ) ( )) (( ) ( ) )
( ) ( ) ( ) (8)

The parameter and control the falling-off the weights in the intensity and spatial domain,
respectively. The choice of these two parameters is vital to denoise the images. The high values of
US
AN
these parameters yield over-smoothing of the image which blurs the fine and edge structures of the
image, while the low values provide poor denoising results.

(a) Speckle reducing bilateral filter (SRBF): The CBF is designed to deal with the Gaussian
M

noise. The functional capability of CBF [39-40] to attenuate speckle noise is not quite effective.
Therefore, the adaptive speckle reducing bilateral filter is highly desirable to reduce the speckle
noise from US images. To accomplish this task, consider the general speckle noise model as
ED

represented by Eq. (1). We first transform the equation of speckle noise model by dividing it
by √ ( ) on both sides of Eq. (1). The new transformed equation is represented as follows.

( ) ( ) ( ) ( )
PT

(9)
√ ( ) √ ( ) √ ( )
If ( ) ( ) and then above equation becomes:
CE

( )
√ ( ) ( ) (10)
√ ( )
Therefore, the process of the computation of weights for the speckle noise is different from that for
AC

the Gaussian noise which is expressed by Eq. (8). For the speckle noise it takes the form

( ) ( )
( ) (( ) ( ) )
√ ( ) √ ( )
( ) ( ) (11)
( )

9
 ACCEPTED MANUSCRIPT

Since is unknown, a simple estimator is used in a classical manner just by substituting

. In this way, we propose the SRBF which is defined as follows.

T
(√ ( ) √ ( )) (( ) ( ) )
( ) ( ) ( ) (12)

IP
∑ ( ) ( ) ( )
( ) (13)
∑ ( )

CR
( )

The above formulation of SRBF provides better results than the CBF by considering the various
qualitative measures used for denoising of US images.

3.3 Rotation invariant bilateral-nonlocal means filter: The classical nonlocal means (CNLM)
[31] filter effectively eliminates Gaussian noise by exploiting the pattern redundancy and self-
similarity among the structures present in an image. The filter denoises each pixel of the image by
US
AN
computing a weighted average of neighborhood pixels using a similarity measure that takes into
account all the neighboring pixels surrounding the pixel to be compared. Mathematically,
∑ ( ) ( )
M

(14)
() ,
∑ ( )

‖ ‖ (15)
( ) ( )
ED

where represents the search window which in this paper, is of the size pixels,
( ) are weights satisfying the conditions ( ) and ∑ ( ) ,
PT

and are image patches located at pixels and , respectively. Here, the parameter , known
as smoothing or filtering parameter, controls the decay of the filtering process. The value of the
CE

parameter h is set to α|SW|  2 [ 41, 42], where |SW| is the size of the window which, in the
present case, is 11× 11 ,  is the estimate of the Gaussian noise and α is a parameter, normally
taken between 0.75 and 1.0. The NLM approach have shown good results on smooth areas and
AC

repetitive textures for which the redundancy is high but on singular structures such as edges and
corners the number of similar patches is small which turns out to be insufficient for proper
denoising. This is referred to as the rare patch effect and has been analyzed by [43–45]. The
presence of noise may cause fake patch detection in NLM, referred to as the patch jittering blurr
effect [46]. Moreover, two similar but rotated patches produce low similarity value which

10
 ACCEPTED MANUSCRIPT

otherwise should be high. It is observed that although the NLM filter provides better denoising
performance in the case of Gaussian noise as compared to the bilateral filter, the latter is more
useful in retaining the singular structures of the image.

T
Recently, Coupe et al. [33] proposed an optimized Bayesian nonlocal means (OBNLM) filter for

IP
speckle noise reduction of US images based on the block-wise implementation of CNLM using
Pearson distance measure. However, OBNLM filter is unable to remove the noise effectively from

CR
the images containing fine structural details and results in the over smoothening of the edges and
textures present in the image. In order to overcome the above limitations, we propose the use of
rotation invariant bilateral nonlocal means filter (RIBNLM) filter proposed by Manjon et al. [41]

US
as a post processing step which optimizes the tradeoff between the smoothing required in singular
areas and preserves fine details of the image. The proposed filter uses an effective rotation
invariant similarity measure based on the intensity values and the corresponding neighborhood
AN
patch means and ,as follows:

∑ ( ) ( ) (16)
() ,
∑ ( )
M

( ) ( ) (17)
( ) ( )
ED

where are the means of the reference patch and the patch being processed in the search

window ( ) and parameter is the smoothing parameter. The local means and are
PT

computed for the whole image only once using 5×5 neighborhoods and saved in an array which is
of the size of the image. A small neighborhood is used to compute and to prevent the over
smoothing of the singular structures of the image. The RIBNLM filter has low computational
CE

complexity and gives appropriate weightage to the patches that are structurally similar but have
different orientations with respect to the reference patch.
AC

4. Proposed hybrid algorithm

In this section, we propose a hybrid algorithm for speckle noise reduction of US images that
maintain an appropriate balance between the speckle suppression and preservation of fine feature
details. The proposed hybrid algorithm is a non-iterative three stage process as shown in Fig. 1
with procedural steps as follows.

11
 ACCEPTED MANUSCRIPT

Procedural Steps:

Step 1: Apply guided filter (GF) to suppress the effect of speckle noise using Eq. (6).

T
Step 2: Apply speckle reducing bilateral filtering (GSRBF) on image pre-filtered by GF using the

IP
weights based on square chord distance based on Eq. (7).

Step 3: Finally, in order to retain the edges and fine structure details, apply bilateral nonlocal

CR
means filter with weights computed using the rotation invariant similarity measure as given
in Eq. (17).

US
AN

Fig. 1 Stepwise procedure of hybrid algorithm

M

In the first stage, the GF is applied to suppress the effect of speckle noise. The use of local
statistics in the GF makes it more suitable to reduce the effect of speckle noise from the crucial
areas of medical images such as sharp details present in a lesion. At the second stage, SRBF filter
ED

is applied for a better restoration of fine structural details of the pre-filtered obtained after the first
stage and hence improve the quality of the denoised image. At this stage a square chord-based
distance measure is used due to its inherent capability in handling the multiplicative nature of the
PT

speckle noise. In addition, we performed exhaustive experimentation using various other distance
measures such as Euclidean distance, Pearson distance measure and L2-squared norm and
CE

observed that the square chord-based distance provided better results and hence is more suitable
for the reduction of speckle noise. Further, in order to enhance the denoising performance and
effective utilization of differently oriented self-similar structures present in the image we propose
AC

the use of RIBNLM at the last stage. The RIBNLM filter was introduced by Manjon et al. [41] for
denoising of MR images. They used ODCT3D filter to attenuate Rician noise from MR images
and then applied rotation invariant NLM to enhance the denoising performance. The main
advantage of using RIBNLM lies in its ability to detect the singular areas (such as sharp edges,
texture areas including sharp corners) better than the CNLM by adding the mean of the reference
patch and patch to be search in a search window in the similarity measure expressed by Eq. (17).

12
 ACCEPTED MANUSCRIPT

The proposed hybrid algorithm is free from any iterative process that tends to diffuse the weak
edges and fine sharp details. The use of second order statistics at the second stage helps to restore
the essential fine edge information while the weight computation using a rotation invariant

T
similarity measure significantly enhance the performance of NLM based denoising filter. The

IP
proposed algorithm has lower computational requirement and the use of different stages in single
hybrid algorithm make it more suitable to deal with speckle noise corrupted US images. The

CR
overall denoising performance of the proposed hybrid algorithm is significantly higher than that of
the individual filter. Further, an exhaustive empirical investigation suggests that the proposed that
the given sequence of these three stages provides the best results and any change in the order
results in suboptimal denoising performance.

5. Experimental analysis
US
AN
In the section, the performance of the proposed hybrid method is analyzed and compared with
classical spatial filters such as Lee, Frost, Kaun, and SRAD filters and recently proposed SR filters
such as SBF, GF, CNLM and OBNLM filters. The proposed hybrid algorithm have been
M

implemented in Visual C++ 6.0 under Microsoft Windows environment on a PC with 2.13 GHz
CPU and 4GB RAM. Four different set of experiments are conducted to test the efficacy of the
proposed algorithm. Experiment I is carried out on the synthetic images which are corrupted with
ED

various level of speckle noise. Experiments II and III are performed on B-mode ultrasound images
and ultrasound simulated images created through the Field II [47-48] simulation, respectively.
Experiment IV is performed on the real ultrasound images downloaded from
PT

http://ultrasoundcases.info/ [51]. To add speckle noise in synthetic images, we set and for
real ultrasound images, we set by using Eq. (1). To evaluate the performance of hybrid
CE

filter on different medical images, the following quantitative and qualitative measures have been
used.

(a) Mean-square error (MSE): The quality of denoised image signal ̂ with respect to original
AC

image can be measured quantitatively using mean-square error which is defined as follows.

∑( () ̂( )) (18)

(b) Signal-to-noise ratio (SNR): The efficacy of denoising methods is judged by SNR [49]
values, and described as follows.

13
 ACCEPTED MANUSCRIPT

∑ ( () ̂( ) )
(19)
∑ ( () ̂( ))
(c) Mean structural similarity index measure (MSSIM): The structure similarity ( ) is a

T
method for measuring the similarity between two images using their structural information.

IP
The error parameter calculates the similarity in a local neighborhood or window by
combining differences in average, variation, and correlation. The value of measure

CR
between patches and from original and denoised images, respectively, is given by [50].

( ̂ )( ̂ )
( ) (20)
( )( )
where and ̂

and , respectively.
are the average gray values while

̂ is the covariance of
̂

and ̂ and
̂
and ̂

(
US
are the variance of patches
) , ( ) are two
AN
parameters which stabilize the division with weak denominator. The parameter is the
dynamic range of the pixel values (255 for 8-bit gray scale images) and and
are default values derived from exhaustive experiments [50]. Using ( ), the
M

index is computed as follows.

∑∑ ( ̂) (21)
ED

The value of lies in the interval , where higher values of indicate better
retention of structural information.
PT

(d) Method noise: Most of the denoising algorithms alter the original structural information of an
image resulting in the loss of useful image contents. Method noise is used to judge the amount
of alteration occurred in the contents of the original image and evaluate the loss of structural
CE

or edge information after denoising. The method noise ̂ is defined as follows.

̂ ̂ (22)
is the input noisy image and ̂ is the denoised image. A denoising algorithm is
AC

where
expected to produce the difference or residual image that looks like a noisy image which
should contain little anatomical structures [31].

14
 ACCEPTED MANUSCRIPT

5.1 Experiment I: Synthetic images

To test the objectivity of different quantitative and qualitative measures, we consider the synthetic

T
image which includes rhombus, concentric circles, lines and rectangles as shown in Fig. 2. The
size of synthetic image is We conduct experiments on the synthetic images corrupted

IP
with different level of speckle noise. The noise-free and noisy image corrupted with speckle noise
having variance is presented in Fig. 2(a)-(b) and the parameters used by the different

CR
speckle reducing filters are listed in Table 1. The corresponding denoised results obtained by the
various SR filters and the proposed hybrid algorithm are depicted in Fig. 3(c)-(n).

Method
Lee
Reference
[14]
Category
Second order statistics
and local adaptive.
Second order statistics
Window size
US
Table 1 The different parameters required for experiments.
Smoothing parameter
-

-
AN
Frost [15]
and local adaptive
Kaun [16, 17]
Second order statistics -
and local adaptive
SRAD [18] Iterative - dt=0.25, iteration=200
SBF [20, 21] Iterative iteration=54
M

GF [23] Local adaptive eps=650*36

CBF [22] Local adaptive
CNLM [31] Nonlocal means h=139.0
ED

Block-wise nonlocal
OBNLM [33] h=251.0
means
eps=650*36
GSRBF Proposed Local adaptive
eps=650*4
Hybrid
PT

Proposed Hybrid ,
algorithm
h=14.3

It is clearly observed from the denoised images and Table 2 that the performance of traditional SR
CE

filters and recently proposed methods are quite ineffective because significant amount of noise is
still observed in the denoised images. It is worth to mention here that, the balance between speckle
suppression and edge preservation is not properly maintained by well-known speckle noise
AC

reduction approaches. In case of SRAD, CNLM and OBNLM filters over smoothing are clearly
visible and weak edges are vanished from the denoised images. The performance of GF is better
than the CBF in term of edge preservation, but the overall performance of this filter is not
satisfactory. Also, we observed that at all noise level the denoising performance of SRBF is better
than the CBF. It means square chord-based distance measure is more effective to attenuate the
speckle noise from synthetic images.

15
 ACCEPTED MANUSCRIPT

T
IP
(a) (b)

CR
Fig 2. (a) Noise free Phantom image (b) Noisy Phantom image corrupted with speckle noise with variance.

The average PSNR achieved by SRBF and CBF is 19.95 dB and 19.51 dB. Thus, a hike in SNR
values of 0.45 dB in the case of SRBF over CBF has been observed, and a similar kind of trend is
also observed in the case of MSSIM values which are reflected from Table 3.

US
Table 2 MSE and SNR values of the denoised synthetic images corrupted with different level of speckle noise with

MSE SNR
AN
Methods Average
0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8
SNR
Noisy 307.6 1190.2 2435.9 3655.1 17.07 11.44 8.66 7.21 11.09
Lee 124.9 161.3 211.1 266.7 20.86 19.76 18.62 17.66 19.23
Frost 93.3 129.6 189.8 273.8 22.13 20.73 19.11 17.57 19.88
M

Kaun 116.6 145.4 184.3 265.3 21.15 20.21 19.22 17.71 19.57
SRAD 94.1 176.4 329.2 275.4 22.15 19.38 16.73 17.52 18.94
SBF 124.4 207.8 547.3 983.6 20.81 18.58 14.26 10.35 16.01
GF 69.9 143.8 163.8 192.6 23.37 20.28 19.69 19.02 20.59
ED

CBF 75.2 193.4 187.4 243.4 23.11 19.01 18.39 17.49 19.51
SRBF 66.46 157.7 209.1 264.6 23.64 19.86 18.66 17.67 19.95
CNLM 50.2 153.2 196.1 213.7 24.83 19.94 18.87 18.52 20.54
OBNLM 43.5 124.3 186.6 198.2 25.46 20.96 19.96 19.18 21.39
GSRBF 49.9 133.4 201.3 259.7 24.91 20.57 18.84 17.77 20.52
PT

Proposed 33.3 103.9 136.2 164.4 26.64 21.68 20.51 19.71 22.14

The denoising performance of recently proposed GF is better than the CBF, and it restores edge
information very well which is depicted from high MSSIM values obtained by GF. The main
CE

reason behind the improved denoising performance and structural preserving capability of GF is
the use of second order statistics. However, the performance of GSRBF is better than GF at the
lower noise level, but at the higher noise level, the performance of GSRBF degrades drastically.
AC

But, it is worth to mention here that, the performance of GSRBF is also better than the CNLM at
the lower noise level, as the noise level increases, the performance of GSRBF is also decreases. A
similar kind of trend is also seen in the case of MSSIM values obtained by GSRBF at lower and
higher noise levels which signifies that at lower noise level the GSRBF restore more edge
information than GF and CNLM filters, whereas, at higher noise level GF and CNLM restore more

16
 ACCEPTED MANUSCRIPT

delicate edge information. This fact is also strengthened by the method noise images obtained by
GSRBF, GF and CNLM filters as shown in Fig 4. It is clearly visible from the method noise
images that at lower noise level the denoising performance and delicate edge preserving capability

T
Comment [G1]: Inserted: d
Comment [G2]: Inserted: cat
of GSRBF is far better than the CNLM filter. A few details are left behind in a method noise

IP
Comment [G3]: Inserted: del
image is seen in the case of GSRBF as compared to GF and CNLM filter. Thus, to enhance the Comment [G4]: Inserted: io
performance of GSRBF, we suggest the use of rotation invariant bilateral-nonlocal means filter at Comment [G5]: Inserted: forma

CR
Comment [G6]: Inserted: i
the third stage. This is the one of the main reason that our proposed hybrid algorithm consists of
Comment [G7]: Inserted: edge
three stages. The performance of hybrid algorithm in term of denoising capability and preservation Comment [G8]: Inserted: the
of useful structural information is far better than the recently proposed OBNLM, CNLM, GF, and Comment [G9]: Inserted: the

traditional SR filters.
US
Table 3 MSSIM values of the denoised synthetic images corrupted with different speckle noise level.
Comment [G10]: Inserted: ,
Comment [G11]: Inserted: the
Comment [G12]: Inserted: s
AN
MSSIM Comment [G13]: Inserted: ,
Methods
0.2 0.4 0.6 0.8 Comment [G14]: Inserted: the
Noisy 0.4338 0.2813 0.2177 0.1842 Comment [G15]: Inserted: ,
Lee 0.7494 0.6357 0.5473 0.4806 Comment [G16]: Inserted: the
Frost 0.7873 0.6704 0.5631 0.4761
Comment [G21]: Deleted:also
M

Kaun 0.7625 0.6554 0.5771 0.4844

SRAD 0.6264 0.6016 0.5472 0.4977 Comment [G22]: Deleted:structural
SBF 0.6648 0.5833 0.4384 0.2497 Comment [G23]: Deleted:co
GF 0.8003 0.6957 0.6622 0.6434 Comment [G24]: Deleted:e
ED

CBF 0.7121 0.5793 0.5596 0.5093

Comment [G25]: Deleted:t
SRBF 0.7715 0.6119 0.5717 0.5172
CNLM 0.7948 0.6527 0.6454 0.6248 Comment [G26]: Deleted:f
OBNLM 0.8503 0.7141 0.7071 0.6694 Comment [G27]: Deleted:n
GSRBF 0.8067 0.7168 0.5905 0.5284 Comment [G28]: Deleted: is
PT

Proposed 0.8764 0.7905 0.7539 0.7177

Comment [G29]: Deleted:fin
Comment [G17]: Inserted: the
However, the denoising and structure preserving capability of OBNLM is far better than the Comment [G31]: Deleted:ur
CNLM. But, in the case of our proposed hybrid algorithm, the denoising capability and structure
CE

Comment [G32]: Deleted:main

Comment [G18]: Inserted: e edg
preserving capability is depicted from high SNR and MSSIM values seen at the bottom row of the
Comment [G19]: Inserted: a
Table 2 and 3. The average SNR obtained by proposed hybrid algorithm is 22.14 dB followed by
Comment [G20]: Inserted: li
AC

OBNLM (21.59 dB), CNLM (20.54 dB) and least in the case of GF (20.59 dB) is observed. The Comment [G30]: Deleted: stru

increase in the average SNR values of 0.55 dB, 1.6 dB, and 1.55 dB has been reflected by the
proposed hybrid filter over the OBNLM, CNLM and GF approaches, respectively.

17
 ACCEPTED MANUSCRIPT

T
IP
(a) Noise-free (b) Noisy (c) Lee (d) Frost

CR
(e) Kaun (f) SRAD (g) SBF US (h) GF
AN
M

(i) CBF (j) SRBF (k) GSRBF (l) CNLM

ED
PT

(m) OBNLM (n) Proposed

Fig. 3. Results of experiment I: (a) Original image, (b) Image corrupted with speckle noise with variance ;
denoised results of (c) Lee filter (d) Frost filter (e) Kaun filter (f) SRAD filter (g) SBF (h) GF (i) CBF (j) SRBF
CE

(k) GSRBF (l) CNLM filter (m) OBNLM filter (n) Proposed filter. Comment [G33]: Inserted: better
Comment [G34]: Inserted: ies
A similar kind of trend is also observed in the case of high MSSIM values shown by the proposed
Comment [G35]: Inserted: the
hybrid algorithm, which signifies that the proposed hybrid approach restore more essential edge Comment [G36]: Inserted: in
AC

information better than the recently proposed (CBF, SBF, GF, CNLM, and OBNLM) and Comment [G37]: Inserted: are
Comment [G38]: Inserted: the
traditional SR (Lee, Frost, Kaun, and SRAD) approaches.
Comment [G39]: Inserted: are
Comment [G40]: Inserted: enc
Comment [G41]: Inserted: c
Comment [G42]: Inserted: rof
Comment [G43]: Inserted: the
Comment [G44]: Deleted:ca

18
 ACCEPTED MANUSCRIPT

T
IP
(a) Noisy residual (b) Lee (c) Frost (d) Kaun

CR
(e) SRAD (f) SBF (g) GF
US (h) CBF
AN
M

(i) SRBF (j) GSRBF (k) CNLM (l) OBNLM

ED
PT

(m) Proposed

Fig. 4. Residual results of phantom image: (a) Noisy residual, (b) Lee filter (c) Frost filter (d) Kaun filter (e) SRAD
filter (f) SBF (g) GF (h) CBF (i) SRBF (j) GSRBF (k) NLM filter (l) OBNLM filter (m) Proposed filter.
CE

Also, one of the key features of proposed hybrid algorithm is the achievement of optimal balance
maintained between speckle suppression and feature preservation is clearly observed through high Comment [G45]: Deleted:ab
AC

MSSIM values as depicted in Table 3 as well from the visual inspection of a denoised image as Comment [G46]: Deleted:l
Comment [G47]: Deleted:t
shown in Fig. 3(n). The above fact is also strengthened by the method noise images obtained by
Comment [G48]: Deleted:is
various speckle denoising filters as shown in Fig. 4. It is clearly observed from the method noise Comment [G49]: Deleted:is
image of the proposed hybrid algorithm that, a less amount of details is left behind by a hybrid Comment [G50]: Deleted:at
Comment [G51]: Deleted:y
filter. Moreover, noise and over smoothing in edge prone areas are clearly visible in the case of
Comment [G52]: Deleted:more fine
well-known SR filters as depicted from their method noise images, respectively. Comment [G53]: Inserted: -

19
 ACCEPTED MANUSCRIPT

Comment [G54]: Deleted:also

5.2 Experiment II: B-mode ultrasound simulation image

The efficacy of the various SR filters and proposed method is discerned by conducting the

T
experiments on B-mode ultrasound images. The B-mode simulated image is different from the
synthetic image because of the background, and it exhibits different kind of features which are

IP
Comment [G55]: Inserted: ,
Comment [G56]: Inserted: nducting
clearly visible from Fig. 5. The noisy B-mode ultrasound image is generated in MATLAB by
Comment [G57]: Inserted: ed by c
setting the different parameters which are defined as follows. The center frequency of the

CR
Comment [G58]: Inserted: discer
ultrasonic wave is 10e6. The velocity of sound in media is 1540 m/s. The variance of speckle noise Comment [G59]: Inserted: is
Comment [G60]: Deleted:o test t
in the B-mode ultrasound image is taken as 0.01. The pulse-width of the transmitting ultrasonic
wave is 2mm, and the beam width of the transmitting ultrasonic wave is 1.5mm. The denoising

US
Comment [G61]: Inserted: ,
Comment [G62]: Inserted: the
performance of different denoising methods in case of noisy B-mode ultrasound image, and their
Comment [G63]: Inserted: the
corresponding denoised images are depicted in Fig. 5. It is clearly observed from the denoising
AN
results obtained by traditional SR filters, SBF, GF, CBF and GSRBF filters that the performance
shown by these filters are quite unsatisfactory. Moreover, the SRAD and SBF methods have
shown the negative influence on contrast because of its iterative nature.
M

Table 4 Comparison of MSE, SNR, and MSSIM values of the denoised B-mode image corrupted with speckle noise
having variance .
Methods MSE SNR MSSIM
Noisy 9227.3 5.57 0.1284
ED

Lee 9104.1 5.66 0.2744

Frost 9105.1 5.67 0.2701
Kaun 9117.5 5.65 0.2744
SRAD 9103.6 5.66 0.3119
SBF 9290.5 5.61 0.2094
PT

GF 9125.9 5.65 0.2319

CBF 9314.2 5.59 0.1374
SRBF 9115.1 5.65 0.2691
CNLM 9237.5 5.62 0.3589
CE

OBNLM 9159.1 5.64 0.3655

GSRBF 9121.7 5.65 0.2814
Proposed 9071.1 5.68 0.3672

The performance of CNLM and OBNLM in term of speckle noise attenuation is quite ineffective
AC

but, both these filters are able to restore the useful edge information which is also depicted from
their high MSSIM values as compared to other SR filters as shown in Table 4. But, the time
complexities of associated with both methods are quite high and therefore, do not able to meet the
real-time requirements of medical ultrasound systems. However, the performance of GF in term of
MSE and SNR measures is quite better than the CNLM, but not able to restore the structural

20
 ACCEPTED MANUSCRIPT

information in the denoised B-mode ultrasound image which is clearly depicted from the low
MSSIM values as shown in Table 4.

T
IP
CR
(a) Noise-free (b) Noisy (c) Lee (d) Frost

US
AN
(e) Kaun (f) SRAD (g) SBF (h) GF
M
ED

(i) CBF (j) SRBF (k) GSRBF (l) CNLM

PT
CE

(m) OBNLM (n) Proposed

Fig. 5. Results of experiment II: (a) Original image, (b) Image corrupted with speckle noise with variance ;
denoised results of (c) Lee filter (d) Frost filter (e) Kaun filter (f) SRAD filter (g) SBF (h) GF (i) CBF (j) SRBF
AC

(k) GSRBF (l) CNLM filter (m) OBNLM filter (n) Proposed filter.

A similar kind of trend is also observed in the case of synthetic images which we discussed in
Section 5.1. The performance of SRBF is far better than the CBF in term of low MSE, high SNR
and MSSIM values which are depicted from Table 4. whereas, the denoising performance of Comment [G64]: Deleted:,

GSRBF have shown better results than the SRBF and GF filters but when GSRBF is compared

21
 ACCEPTED MANUSCRIPT

with CNLM filter, although it provides low MSE and higher SNR values than CNLM, but it is not
able to restore essential structural information in denoised images.

T
The performance of the proposed hybrid algorithm outperforms the all other eleven methods

IP
concerning MSE, SNR and MSSIM values. Moreover, denoised image obtained by the proposed Comment [G65]: Deleted:a

hybrid algorithm indicates the preservation of sharp boundaries in edge prone areas. The

CR
comparison between all SR filters and proposed hybrid algorithm exhibits that the performance of
a hybrid algorithm is far better than all the other SR methods and also able to restore various kinds
of useful features including small objects and fine edges in the denoised image. Moreover, the Comment [G66]: Deleted:r

US
Comment [G67]: Deleted:is conducted
proposed hybrid algorithm improves the image quality and enhances the contrast level in the
denoised image which is very useful to provide better diagnosis for malignant lesions as well for Comment [G68]: Inserted: a
Comment [G69]: Inserted: ning
tissue characterization in the US images.
AN
Comment [G70]: Inserted: c

5.3 Experiment III: Field II kidney simulation image Comment [G71]: Inserted: co
Comment [G72]: Inserted: re
The proposed hybrid algorithm is also tested on simulated kidney ultrasound image generated by Comment [G73]: Inserted: ,

using linear acoustic and Field II program provided by Jensen [47-48]. The Field II program Comment [G74]: Inserted: -
M

Comment [G75]: Deleted:are able to

evaluates the pulsed ultrasound fields by using Tuphole-Stepanishen method. The comparisons
Comment [G76]: Deleted:is
among all evaluated methods on simulated kidney image are shown in Fig. 6. Comment [G77]: Deleted:lso
ED

Table 5 Comparison of MSE, SNR, and MSSIM values of the denoised kidney simulated image generated by Field II Comment [G78]: Deleted:but
simulation.
Comment [G79]: Deleted:i
Methods MSE SNR MSSIM
Comment [G80]: Deleted: t
Noisy 2543.5 8.48 0.1352
Lee 2001.1 9.17 0.3786 Comment [G81]: Deleted:ms of
PT

Frost 2005.4 9.16 0.3707

Kaun 2000.4 9.17 0.3786
SRAD 1951.6 9.26 0.4387
SBF 2104.9 8.98 0.3176
CE

GF 1971.1 9.23 0.4103

CBF 2101.1 9.14 0.3079
SRBF 2001.1 9.16 0.3784
CNLM 1941.6 9.29 0.4556
OBNLM 1917.4 9.33 0.4904
AC

GSRBF 1967.1 9.17 0.3801

Proposed 1880.1 9.39 0.4962

A comparative performance analysis between the traditional SR filters, recently proposed filters
like SBF, GF, CBF, CNLM, and OBNLM with the proposed hybrid algorithm on the simulated
kidney image is demonstrated in this section.

22
 ACCEPTED MANUSCRIPT

Comment [G82]: Inserted: -

The denoising performance of traditional SR (Lee, Frost, Kaun, and SRAD) filters are comparable,
but among all traditional SR filters, SRAD provides best denoising performance on the simulated
Comment [G83]: Inserted: ,
kidney image as depicted from Table 5. Moreover, SRAD filter is able to restore the structural and

T
edge information better than the all traditional SR filters. However, the denoising performance of

IP
SBF is quite low among all the speckle denoising filters. The performance of CBF is better than
the SBF, but does not perform better than SRBF in term of MSE, SNR and MSSIM values. The

CR
higher MSSIM value obtained by SRBF indicates it’s superiority over CBF in term of denoising
and structure preserving capability.

US
AN
(a) Noise-free (b) Noisy
M

(c) Lee (d) Frost (e) Kaun (f) SRAD

ED
PT

(g) SBF (h) GF (i) CBF (j) SRBF

CE

(k) GSRBF (l) CNLM (m) OBNLM (n) Proposed

Fig. 6. Results of experiment III: (a) original image, (b) image corrupted with speckle noise with variance ;
AC

denoised results of (c) Lee filter (d) Frost filter (e) Kaun filter (f) SRAD filter (g) SBF (h) GF (i) CBF (j) SRBF
(k) GSRBF (l) CNLM filter (m) OBNLM filter (n) Proposed filter.

Also, the denoising performance of GF is superior to that of traditional SR filters, CBF, and
SRBF, respectively. The better denoising performance and structure preserving capability of GF is
clearly observed from low MSE, high SNR and MSSIM values as shown in Table 5. However, the
denoising performance of OBNLM is quite more effective than CNLM and aforementioned
23
 ACCEPTED MANUSCRIPT

approaches, respectively. But the main drawback about the OBNLM approach is that it is not able
to bridge the gap between appropriate amount of denoising required and preservation of fine edge
details in the singular areas which is depicted very well from Fig. 6(m). The denoising

T
performance of proposed hybrid algorithm in the singular areas is far better than all eleven speckle

IP
noise reduction filters. The reason behind this excellent performance offered by a hybrid algorithm
is the use of second order statistics and efficient use of rotation invariance property applied on

CR
both local and nonlocal information in a combined way. The higher SNR and low MSE values
indicates that the adequate amount of speckle noise is removed from simulated kidney image,
while preservation of fine edges is visually seen from the denoised image and as well from the
higher MSSIM values.

5.4 Experiment IV: Real ultrasound images

US Comment [G84]: Inserted: is
AN
In this section, we analyze the denoising performance and structure preservation capability of
proposed hybrid algorithm on real ultrasound images such as liver and urinary tract images which
are download from [51]. We add speckle noise according to the speckle noise model described by
M

Eq. (1) by setting the factor . The performance obtained by proposed hybrid algorithm Comment [G85]: Inserted: Th

Comment [G86]: Inserted: rithm. Also,

describes its denoising capability and structure preserving behavior on real ultrasound images.
Comment [G87]: Inserted: hybrid alg
Therefore, one of the popular qualitative measures (method noise) is used to tell about the
ED

Comment [G88]: Inserted: d

denoising and structure preserving ability of the proposed hybrid algorithm. Also, one of the key Comment [G89]: Inserted: s
Comment [G90]: Inserted: prop
objective of all denoising filters is to maintain the contrast level in the denoised images, because
Comment [G91]: Inserted: f the
speckle forms a random granular pattern which adversely effects on the contrast level of an image.
PT

Comment [G92]: Inserted: he

The denoising performance of various despeckling filters and by proposed hybrid algorithm with Comment [G93]: Inserted: ithm. Th

their corresponding denoised and method noise images have shown in Fig. (7)-(10). The denoising Comment [G94]: Inserted: ability of an
algo
CE

performance of Lee, Frost and Kaun filter are not satisfactory because they leave more edge details Comment [G95]: Inserted: to judge the
ca
in the method noise image. Moreover, the denoising capability of Frost filter in the case of urinary Comment [G96]: Inserted: tanc
tract image is erroneous as compared to Lee and Kaun filter. The artifacts related to structures is Comment [G97]: Inserted: p
AC

Comment [G98]: Inserted: have its i

appeared in both denoised [Fig. 9] and method noise image [Fig. 10]. However, the performance
Comment [G99]: Inserted: s
of SRAD filter in term of its denoising and structural preservation capability as compared to Lee,
Comment [G100]: Inserted: s
Frost, and Kaun filter is satisfactory. On the other hand, the iterative SBF method does not Comment [G101]: Inserted: ess

perform well according to the method noise image generated after the denoising filtering Comment [G102]: Inserted: the
Comment [G103]: Inserted: the
operation. A lot of edge details and speckle noise is still visible in the method noise image which
Comment [G104]: Inserted: l
tells its imperfection. Comment [G105]: Inserted: A

24
 ACCEPTED MANUSCRIPT

T
IP
(a) Noisy (b) Lee (c) Frost (d) Kaun

CR
(e) SRAD (f) SBF (g) GF
US (h) CBF
AN
M

(i) SRBF (j) GSRBF (k) CNLM (l) OBNLM

ED
PT

(m) Proposed
Fig. 7. Results of experiment IV: (a) Noisy liver image, denoised results of (b) Lee filter (c) Frost filter (d) Kaun filter
(e) SRAD filter (f) SBF (g) GF (h) CBF (i) SRBF (j) GSRBF (k) CNLM filter (l) OBNLM filter (m) Proposed filter.
CE

The denoising performance of CBF is better than SBF and SRBF also provides good denoising
and structure preserving performance than SBF and CBF but, it suffers from a distortion of edges
AC

as reflected from Fig. 8(h) and 10(h). The GF belong to the category of second order statistics Comment [G106]: Inserted: a
Comment [G107]: Inserted: s
have vanished the speckle noise more than the aforementioned filters, but small objects and edge
Comment [G108]: Inserted: ,
related structures are still visible in the method noise image. It is observed from the Figs. 8(i) and
10(i), that the GSRBF has a good despeckling performance than the GF, but at the cost of a blurred
image. The denoising performance of CNLM in a case of both real ultrasound images are quite Comment [G109]: Inserted: a

unsatisfactory, because of inefficient smoothing of sharp boundaries which are clearly visible in

25
 ACCEPTED MANUSCRIPT

the denoised [Fig. 7(k) and 9(k)] and method noise image [Fig. 8(j) and 10(j)]. Although, the Comment [G110]: Inserted: i
Comment [G111]: Deleted:ed
OBNLM filter provides satisfactory denoising performance, but due to its high complexity it’s
Comment [G112]: Deleted:In order to
tough to apply on real-time ultrasound systems.

T
mak
Comment [G113]: Deleted:e

IP
Comment [G114]: Deleted:s
Comment [G115]: Deleted:a
Comment [G116]: Deleted:ive

CR
Comment [G117]: Deleted:i
Comment [G118]: Deleted:s
Comment [G119]: Deleted:cap
Comment [G120]: Deleted:.

US
(a) Lee (b) Frost (c) Kaun (d) SRAD
Comment [G121]: Deleted:Als
Comment [G122]: Deleted:,
Comment [G123]: Deleted:n
AN
Comment [G124]: Deleted:f
Comment [G125]: Deleted:,
Comment [G126]: Deleted:ow
Comment [G127]: Deleted:Large
(e) SBF (f) GF (g) CBF (h) SRBF Comment [G128]: Deleted:am
M

Comment [G129]: Deleted:un

Comment [G130]: Deleted:,
Comment [G131]: Deleted: but
ED

Comment [G132]: Deleted: but

Comment [G133]: Deleted:verydiffic
Comment [G134]: Deleted:lt
(i) GSRBF (j) CNLM (k) OBNLM (l) Proposed Comment [G135]: Deleted:ied
PT

Fig. 8. Residual results of Liver image (a) Lee filter (b) Frost filter (c) Kaun filter (d) SRAD filter (e) SBF (f) GF Comment [G136]: Deleted:a
(g) CBF (h) SRBF (i) GSRBF (j) CNLM filter (k) OBNLM filter (l) Proposed filter.
Comment [G137]: Deleted:is
Comment [G138]: Deleted:good
The reason behind is the size of ultrasound images which are quite large, and OBNLM filter
CE

Comment [G139]: Deleted:e

requires too much time to process the ultrasound images. Therefore, low complexity, better Comment [G140]: Deleted:t

denoising capability, and structure preserving denoising filter are highly desirable to process the Comment [G141]: Deleted:v
Comment [G142]: Inserted: ,
real medical US images. The low algorithm complexity of GF and SRBF methods are combined
AC

Comment [G143]: Inserted: gh

with rotation invariant bilateral-nonlocal means filter to achieve better speckle noise suppression Comment [G144]: Inserted: to

and edge preservation performance as compared to all other SR approaches. Comment [G145]: Inserted: y
Comment [G146]: Inserted: ,
Comment [G147]: Inserted: s
Comment [G148]: Inserted: ,
Comment [G149]: Inserted: are

26
 ACCEPTED MANUSCRIPT

T
IP
(a) Noisy (b) Lee (c) Frost (d) Kaun

CR
(e) SRAD (f) SBF (g) GF US (h) CBF
AN
M

(i) SRBF (j) GSRBF (k) CNLM (l) OBNLM

ED
PT

(m) Proposed
Fig. 9. Results of experiment IV: (a) Noisy urinary tract image, denoised results of (b) Lee filter (c) Frost filter
CE

(d) Kaun filter (e) SRAD filter (f) SBF (g) GF (h) CBF(i) SRBF(j) GSRBF (k) CNLM filter (l) OBNLM filter
(m) Proposed filter.

In addition to, low algorithm complexity, strong speckle suppression capability and excellent edge
AC

preservation proficiency of proposed hybrid algorithm would help in the lay foundation for
developing the new hybrid algorithm by combining the advantages of various speckle noise
reduction filters to attenuate the speckle noise efficiently from real ultrasound images. Comment [G150]: Inserted: excellent
Comment [G151]: Inserted: nt

27
 ACCEPTED MANUSCRIPT

T
IP
(a) Lee (b) Frost (c) Kaun (d) SRAD

CR
(e) SBF (f) GF (g) CBF
US (h) SRBF
AN
M

(i) GSRBF (j) CNLM (k) OBNLM (l) Proposed

ED

Fig. 10. Residual results of Liver image (a) Lee filter (b) Frost filter (c) Kaun filter (d) SRAD filter (e) SBF (f) GF
(g) CBF (h) SRBF (i) GSRBF (j) NLM filter (k) OBNLM filter (l) Proposed filter.

6. Time complexity Analysis

PT

The NLM filter denoises each pixel ( ) of an image after computing the radiometric and
geometric distances between two blocks over a search window. Let the size of an image be .
The size of the search window , and the size of the local window , then the time
CE

complexity of the NLM approach is ( ) [31]. The time complexity of the OBNLM
approach is also ( ) although it involves more operations while updating the denoised
AC

intensity values in a local window. Therefore, it takes more time than the NLM approach. The
proposed approach uses three stages, (1) the guided filter, (2) the SRBF, and (3) the RIBNLM. The
time complexity of the guided filter is ( ), where represents the size of the
neighborhood ( ) used in the filter. The time complexity of the SRBF is ( ) as it is a
special case of the NLM filter with . The time complexity of the RIBNLM is also ( )
because its computational steps are similar to the CBF which uses a search window of size

28
 ACCEPTED MANUSCRIPT

and local window of size . The two approaches differ in the way the weights are computed.
Therefore, the proposed method is faster by a factor of where, is the size of the local window

T
which normally set [32, 33]. Thus, for the value of , the proposed method is faster
by a factor of 121. However, the actual reduction in the total computation is less than the

IP
theoretical value because of the three stages involved in the proposed method. It is difficult to
provide the experimental values of the total time taken by these approaches because we have

CR
implemented NLM and the proposed approach in C++ while the code for OBNLM is available in
MATLAB [33].

7. Conclusion

US
In this paper, we have proposed a hybrid algorithm for denoising of ultrasound images which fetch
the both local and nonlocal information efficiently to suppress the speckle noise and able to
AN
preserve the fine edges and small structures in a better way than other SR filters. To verify this
fact, various experiments have been conducted on synthetic, simulated and real ultrasound images.
The proposed hybrid algorithm has been compared with other well known speckle noise reduction
M

approaches. The experimental results prove this fact that, the proposed hybrid approach exhibits
the best denoising performance in terms of MSE, SNR, and MSSIM values. The proposed
algorithm is able to preserve the edges and small structures present in the crucial areas of lesions
ED

while sufficiently remove the speckle noise from edge prone areas. It is also observed that while
removing the speckle noise from ultrasound images the proposed hybrid algorithm has strong
potential to maintain the contrast level in denoised ultrasound images. The visual inspection of
PT

denoised ultrasound images suggested that the performance of the proposed hybrid algorithm is far
better than the classical speckle noise reduction approaches. The success of the proposed algorithm
CE

would help in building the lay foundation for inventing the hybrid algorithm for denoising of
ultrasound images.

Conflict of interests
AC

The authors declare that there is no conflict of interests regarding the publication of this paper.

Acknowledgements

One of the authors (KS) is thankful to the Ministry of Minority Affairs (MOMA) and University
Grant Com-mission (UGC), Govt. of India, for providing Maulana Azad National Fellowship

29
 ACCEPTED MANUSCRIPT

(F1-17.1/2012-13/MANF-2012-13-SIK-PUN-13364) for carrying out the research work. Thanks

to Mr. Debdoot Sheet, who is a senior research fellow in Indian Institute of Technology
Kharagpur, India, for providing us with the MATLAB code for B-mode ultrasound image

T
simulation. The authors would also like to thank Peter C. Tay for providing the MATLAB code of

IP
the SBF filter.

References

CR
[1] J. C. Dainty, Some statistical properties of random speckle patterns in coherent and partially
coherent illumination, Optica Acta, 17(1970) 761-772.

1150. US
[2] J. W. Goodman, Some fundamental properties of speckle, J. Opt. Soc. Am. 66, 11(1976) 1145-

[3] C. B. Burkhardt, Speckling in ultrasound B-mode scans, IEEE Trans. Son. Ultrason., 25(1978)
AN
1-6.
[4] P. N. T. Wells, M. Halliwell, Speckle in ultrasonic imaging, Ultrasonics, 19(1981) 225-290.
[5] M. Tur, K. C. Chin, J. W. Goodman, When is speckle noise multiplicative?, JMAA, 1982.
M

[6] R. F. Wagner, S. W. Smith, J.M. Sandrik, H. Lopez, Statistics of speckle in ultrasound B-

scans, IEEE Trans. on Ultrason., 30(1983) 156-163.
[7] P. A. Magnin, O. T. Von Ramm, F. L. Thurstone, Frequency compounding for speckle contrast
ED

reduction in phased array images, Ultrason. Imaging, 4 (1982) 267-281.

[8] D. P. Shattuck, O. T. Von Ramm, Compound scanning with a phased array, Ultrason. Imaging,
4(1982) 93-107.
PT

[9] M. O'Donnell, S. D. Silverstein, Optimum displacement for compound image generation in

medical ultrasound, IEEE Trans. Ultrason. Ferroelect. Freq. Contr.,35(1988) 470-476.
CE

[10]V. Behar, D. Adam, Z. Friedman, A new method of spatial compounding imaging,

Ultrasonics, 41 (2003) 377-384.
[11] Y. Chen, S.L. Broschat, P. J. Flynn, Phase insensitive homomorphic image processing for
AC

speckle noise reduction, Ultrason. Imaging, 18(1996) 122-139.

[12] T. Loupas, W.N. McDicken, P.L. Allan, An Adaptive weighted median filter for speckle
suppression in medical ultrasonic images, IEEE Trans. Circuits Syst., 36(1989) 129-135.
[13] T. Huang, G. Yang, G. Tang, A fast two-dimensional median filtering algorithm, IEEE Trans.
Acoust. Speech Signal Processing, 27(1979) 13-18.

30
 ACCEPTED MANUSCRIPT

[14] J. S. Lee, Digital image enhancement and noise filtering by use of local statistics, IEEE Trans.
Pattern Anal. Mach. Intell., 2(1980) 165-168.
[15] V. S. Frost, J.A. Stiles, K.S. Shanmugan, J.C. Holtzman, A model for radar images and its

T
application to adaptive digital filtering of multiplicative noise, IEEE Trans. Pattern Anal.

IP
Mach. Intell., 4(1982) 157-166.
[16] D. T. Kuan, A. Sawchuk, T. Strand, P. Chavel, Adaptive noise smoothing filter for images

CR
with signal-dependent noise, IEEE Trans. Pattern Anal. Mach. Intell., 7(1985) 165-177.
[17] D. T. Kuan, A. A. Sawchuk, T.C. Strand, P. Chavel, Adaptive restoration of images with
speckle, IEEE Trans. Acoust. Speech Signal Process., 35(1987) 373-383.

Processing, 11(2002) 1260-1270. US

[18] Y. Yu, S. T. Acton, Speckle reducing anisotropic diffusion, IEEE Trans. on Image

[19] K. Krissian, C. F. Westin, R. Kikinis, K. G. Vosburgh, Oriented speckle reducing anisotropic

AN
diffusion, IEEE Trans. on Image Processing, 16(2007) 1412-1424.
[20] P. C. Tay, S. T. Acton, J. A. Hossack, A stochastic approach to ultrasound despeckling, in:
Proc. 3rd IEEE Int. Symp. Biomedical Imaging: Nano to Macro, (2006) 221-224.
M

[21] P. C. Tay, S. T. Acton, J. A. Hossack, Ultrasound despeckling using an adaptive window

stochastic approach, in: Proc. IEEE Int. Conf. Image Processing, (2006) 2549-2552.
[22] C. Tomasi, R. Manduchi, Bilateral filtering for gray and color images, in: Proc. of 6th Inter.
ED

Conf. on Comp. Vision, (1998) 839-846.

[23] Kaiming He, Jian Sun, Xiaoou Tang, Guided image filtering, in: Proceedings of11th
PT

European Conference on Computer Vision, Berlin, 2010, 1–14.

[24] A. Pizurica,W. Philips, I. Lemahieu, M. Acheroy, A joint inter and intra-scale statistical
model for Bayesian wavelet based image denoising, IEEE Trans. on Image Processing,
CE

11(2002) 545-557.
[25] A. Pizurica, W. Philips, I. Lemahieu, M. Acheroy, A versatile wavelet domain noise filtration
technique for medical imaging, IEEE Trans. on Med. Imag., 22(2003) 323-331.
AC

[26] S. Gupta, R.C. Chauhan, S.C. Saxena, A wavelet-based statistical approach for speckle
reduction in medical ultrasound images, Med. Biol. Engg. Comput., 42(2004) 189-192.
[27] S. Gupta, R.C. Chauhan, S.C. Saxena, Robust non-homomorphic approach for speckle
reduction in medical ultrasound images, Med. Bio. Eng. Comput., 43(2005) 189-195.

31
 ACCEPTED MANUSCRIPT

[28] G. Andria, F. Attivissimo, G. Cavone, N. Giaquinto, A.M.L. Lanzolla, Linear filtering of 2-D
wavelet coefficients for denoising ultrasound medical images, Measurement, 45(2012) 1792-
1800.

T
[29] J. Zhang, G. Lin, L. Wu, Y. Cheng, Speckle filtering of medical ultrasonic images using

IP
wavelet and guided filter, Ultrasonics, 65(2016) 177-193.
[30] J. Zhang, G. Lin, L. Wu, C. Wang, Y. Cheng, Wavelet and fast bilateral filter based de-

CR
speckling method for medical ultrasound images, Biomedical Signal Processing and Control,
18(2015) 1-10.
[31] A. Buades, B. Coll, and J. M. Morel, A review of image denoising algorithms, with a new
one, Multiscale Model Simul., 4(2005) 490–530.
US
[32] C. Kervrann, J. Boulanger, and P. Coupé, Bayesian non-local means filter, image redundancy
and adaptive dictionaries for noise removal,in Proc. conf. scale-space and variational
AN
Methods, Ischia, Italy, June 2007, pp. 520–532.
[33] P. Coupe, P. Hellier, C. Kervrann, C. Barillot, Nonlocal means-based speckle filtering for
ultrasound images, IEEE Transactions on Image Processing, 18(2009) 2221-2229.
M

[34] Y. Guo, Y. Wang, T. Hou, Speckle filtering of ultrasonic images using a modified non local-
based algorithm, Biomedical Signal Processing and Control, 6(2011) 129-138.
ED

[35] Y. Zhan, M. Ding, L. Wu, X. Zhang, Nonlocal means method using weight refining for
despeckling of ultrasound images, Signal Processing, 103(2014) 201-213.
[36] K. Radlak, B. Smolka, Adaptive Non-local means filtering for speckle noise reduction,
PT

International conference on computer vision and graphics (ICCVG), LNCS 8671, pp. 518–
525, 2014.
[37] J. Yang, J. Fan, D. Ai, X. Wang, S. Tang, Y. Wang, Local statistics and non-local mean filter
CE

for speckle noise reduction in medical ultrasound image, Neurocomputing, 195(2016), 88–95.
[38] P.V. Sudeep, P.Palanisamy, J. Rajan, H.Baradaran, L. Saba, A. Gupta, J.S. Suri, Speckle
reduction in medical ultrasound images using an unbiased non-local means method,
AC

Biomedical Signal Processing and Control, 28 (2016) 1–8.

[39] S. Balocco, C. Gatta, O. Pujol, J. Mauri, P. Radeva, SRBF: Speckle reducing bilateral
filtering, Ultrasound Med. Biol., 36(2010) 1353-1363.

32
 ACCEPTED MANUSCRIPT

[40] J. Tang, S. Guo, Q. Sun, Y. Deng, D. Zhou, Speckle reducing bilateral filter for cattle follicle
segmentation, International conference on Bioinformatics & Computational Biology (Bio
Comp 2009) Las Vegas, NV, USA. 13-16 July 2009.

T
[41] José V. Manjón, P. Coupé, A. Buades, D. L. Collins, M. Robles, New methods for MRI

IP
denoising based on sparseness and self-similarity, Medical Image Analysis 16 (2012) 18–27.
[42] H. Zhong, C. Yang, X. Zhang, A new weight for nonlocal means denoising using method

CR
noise, IEEE Sig. Process. Lett., 19 (2012) 535–538.
[43] P. Coupé, P. Yger, S. Prima, P. Hellier, C. Kervrann, C. Barillot, An optimized blockwise
nonlocal means denoising filter for 3D magnetic resonance images, IEEE Trans. Med.
Imaging, 27 (2008) 425–441.
US
[44] C. Louchet, L. Moisan, Total variation as a local filter, SIAM J. Imaging Sci., 4 (2011) 651–
694.
AN
[45] C.-A. Deledalle, V. Duval, J. Salmon, Non-local methods with shape-adaptive patches (NLM-
SAP), J. Math. Imaging Vis., 43 (2012) 103–120.
[46] C. Sutour, C.-A. Deledalle, J.-F. Aujol, Adaptive regularization of the NL-means: application
M

to image and video denoising, IEEE Trans. Image Process., 23(2014) 3506–3521.
[47] J. A. Jensen, N. B. Svendsen, Calculation of pressure fields from arbitrarily shaped, apodized,
ED

and excited ultrasound transducers, IEEE Trans. Ultrason. Ferroelectr. Freq. Control, 39
(1992) 262-267.
[48] J. A. Jensen, Field: a program for simulating ultrasound systems, presented at 10th
PT

Nordicbaltic conference on biomedical imaging, 1996.

[49] D. Sakrison, On the role of the observer and a distortion measure in image transmission, IEEE
Trans. Commun., 25(1977) 1251–1267.
CE

[50] Z. Wang, A. C. Bovik, H. R. Sheikh, Image quality assessment from error visibility to
structural similarity, IEEE Trans. Image Process., 13(2004) 600-612.
[51] Ultrasound image database collected from Gelderse Vallei Hospital in Ede, the Netherlands.
AC

http://ultrasoundcases.info/Category.aspx?cat=87 (accessed 13.06.16).

33