You are on page 1of 6

JOURNAL OF COMPUTING, VOLUME 3, ISSUE 9, SEPTEMBER 2011, ISSN 2151-9617

HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING
WWW.JOURNALOFCOMPUTING.ORG 24
A Hybrid Despeckling Model for Medical
Ultrasound Images
R.Vanithamani, G.Umamaheswari, A.AjayKrishnan, C.Ilaiyarasan,
K. Iswariya and C.G.Kritika
AbstractA hybrid despecklingmodelbased on wavelet shrinkage and bilateral filter isdesigned and tested for ultrasound im-
ages. The different wavelet thresholding techniques along with bilateral filter before decomposition and after reconstruction using
wavelets are compared against NeighShrinkSURE technique with and without bilateral filter.The performance of the proposed method
is assessed using standard performance metrics like Signal to Noise Ratio (SNR), Edge Preservation Index (EPI) and Mean Square
Error(MSE). The results demonstrate that the use of bilateral filter in combination with NeighShrinkSUREthresholding technique yields
an increase in the Signal to Noise Ratio(SNR), Edge Preservation Index (EPI) anda decrease in the Mean Square Error (MSE) when
compared to other methods.
Index TermsBilateral filter, Despeckling,NeighShrinkSURE.


1 INTRODUCTION
EDICAL Ultrasonography is one of the most
widely used techniques for imaging organs and
soft tissue structures in the human body. Unfortu
nately,theultrasoundimagesareinherentlycorruptedby
speckle noise which makes it difficult for the observer to
diagnose important details. In addition, speckle reduces
thecontrastoftheimageandcomplicatestheimageproc
essingtaskssuchassegmentationandclassification.Thus,
denoising is indeednecessary andit is considered as the
firststepbeforetheimagedataisanalyzed.Itisnecessary
to apply an efficient denoising technique to compensate
foranydatacorruption[1].Thegoalofthedenoisingisto
remove the noise while preserving the important image
informationasmuchaspossible.
Many denoising methods have been proposed over
the years, such as the wiener filter, wavelet thresholding
[2], and bilateral filtering [3]. Among these, wavelet
thresholding has been reported as a highly successful
method. In wavelet thresholding, a signal is decomposed
into approximation and detail subbands and the coeffi
cients in the detail subbands are processed via hard or
soft thresholding. The hard thresholding eliminates coef
ficients that are smaller than a threshold; the soft thresh
olding shrinks the coefficients that are larger than the
threshold as well. The main task of wavelet thresholding
istheselectionofthresholdvalueandtheeffectofdenois
ingdependsontheselectedthreshold.Abiggerthreshold
will throw off the useful information and the noise com
ponents at the same time while a smaller threshold can
noteliminatethenoiseeffectively.Donoho[2]gaveagen
eralestimationmethodforthreshold,butthebestthresh
oldcannotbefoundbythismethod.Changetal.[4]have
used predictive models to estimate the threshold. It is a
spatially adaptive threshold based on context modeling.
They also presented a datadriven threshold for image
denoising in a Bayesian framework. In the SUREShrink
approach [5], the optimal threshold value based on the
Steins Unbiased Estimator for Risk (SURE) is estimated.
NeighShrink[6]thresholdsthewaveletcoefficientsaccord
ingtothemagnitudeofthesquaresumofallthewavelet
coefficientswithintheneighbourhoodwindowusingDo
noho universal threshold. NeighShrinkSURE[7]
adoptsSteins Unbiased Risk Estimator instead of Dono
hos universal threshold to obtain the optimal threshold
withminimumriskforeachsubband.Amajorstrengthof
wavelet thresholding is the ability to treat different fre
quencycomponentsofanimageseparately.Thisisimpor
tantbecausenoiseinrealscenariosmaybefrequencyde
pendent but in wavelet thresholding the problem experi
encedisgenerallysmoothingofedges.
The bilateral filter proposed in [3] is an alternative to
wavelet thresholding. It applies spatially weighted aver
agingwithoutsmoothingedges.Thisisachievedbycom
bining two Gaussian filters. One filter works in spatial
domainandtheotherintheintensitydomain.Therefore,
not only the spatial distance but also the intensity dis
tance is important for the determination of weights [3].
Hence,thesetypesoffilterscanremovenoiseinanimage
while retaining edges. The main objective of this work is
to design a filterfor effectivedespecklingof medical ultra
soundimageswithoutsmoothingedges.
M

- R.Vanithamani is with the Department of Biomedical Instrumentation
Engineering, Faculty of Engineering, Avinashilingam University for
Women, Coimbatore, India.
- Dr.G.Umamaheswari, A.Ajay Krishnan, K.Iswariya, C.G.kritikaand
C.Ilaiyarasanare with the Department of Electronics and Communication
Engineering, PSGCollege of Technology, Coimbatore, India.

- 2011 Journal of Computing Press, NY, USA, ISSN 2151-9617
- http://sites.google.com/site/journalofcomputing/
JOURNAL OF COMPUTING, VOLUME 3, ISSUE 9, SEPTEMBER 2011, ISSN 2151-9617
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING
WWW.JOURNALOFCOMPUTING.ORG 25
Section2ofthepaperintroducestheconceptofwave
let thresholding and works on it. Section 3 explains con
cepts of bilateral filtering. Section 4 describes the pro
posed hybrid denoising model. Results are discussed in
section5.Finallytheconclusionsaredrawninsection6.
2 WAVELET DECOMPOSITION
The wavelet decomposition process involves three basic
steps:
i)AlinearforwardDiscreteWaveletTransform(DWT)
ii)Nonlinearthresholdingstepand
iii)AlinearInverseDiscreteWaveletTransform(IDWT).
Forimagedenoising,2Dwavelettransformhastobeper
formed.Ateverydecompositionlevelfourfrequencysub
bands are obtained namely LL, LH, HL, and HH. The
next level should be applied to the low frequency sub
band LL. This process is continued until a prespecified
levelisreached.Sincethenoiseisaveragedoutinthelow
frequency wavelet coefficients, the coefficients in these
frequenciesareunalteredandonlythewaveletcoefficients
in the high frequency levelsLH, HL andHH arethre
sholded.
2.1 Wavelet Thresholding
It has been observed that in many signals, energy is
mostlyconcentratedinasmallnumberofdimensionsand
the coefficients of these dimensions are relatively large
compared to other dimensions or to any other signal
(noise) that has its energy spread over a large number of
coefficients. Hence, in wavelet thresholding, each coeffi
cient is thresholded by comparing against a threshold to
eliminate noise, while preserving important information
of the original signal. Usually two types of thresholding
techniquesareused.Theyareasfollows:

2.2 Hard Thresholding


Inthehardthresholdingscheme,theinputiskept,ifitis
greaterthanthethreshold,otherwiseitissettozeroas
in(1).
( )
, X
Y Thard ,
0 ,
X if
X
otherwise

>=
= =

`
)
(1)
2.3 Soft Thresholding
The soft thresholding scheme given in (2) is an extension
ofthehardthresholding.Iftheabsolutevalueoftheinput
X is less than or equal to , then the output is forced to
zero.IftheabsolutevalueofXisgreaterthan,thenthe
outputisobtainedbysubtractingfromX.
( )
,
Y Tsoft ,
0,
X if X
X
otherwise

>=
= =

`
)
(2)
2.4 Bayes thresholding
BayesShrink is an adaptive datadriven threshold for im
age denoising via wavelet softthresholding. The thre
shold is driven in a Bayesian framework [4], and the
waveletcoefficientsineachdetailsubbandareconsidered
as random variables with Generalized Gaussian distribu
tion(GGD).ThethresholdTisobtained,whichminimizes
the Bayesian Risk. The reconstruction using BayesShrink
is found to be smoother and produces more visually ap
pealingimages.Thenoisevariance(

)inthenoisyimage
isrepresentedby(3)
( )
2 2
[ / 0.6745 ] median Y o = (3)
The signal variance (

) is estimated for each wavelet


coefficientinHL,LH,HHusing(4)

| |
2 1
2 2
, max , , 0
,
0 0
2 1
0, 0
i j y k l
k l
x
i j
L
o o
|
(
(

=
e
+
| |
(
|

\ .
(4)

2.5NeighShrinkSURETechnique

Neighshrink [6] thresholds the wavelet coefficients ac
cording to the magnitude of the square sum of all the
wavelet coefficients within theneighbourhood window
using Donohosuniversal threshold. NeighShrinkSUREa
doptsSteins Unbiased Risk Estimator (SURE) instead of
Donohos universal threshold to obtain the optimal thre
sholdwithminimumriskforeachsubband[7].
The wavelet transform can be accomplished by applying
the lowpass and highpass filters on the same set of low
frequency coefficients recursively. This means that the
wavelet coefficients are correlated in a small neighbour
hood.Alargewaveletcoefficientwillprobablyhavelarge
coefficientsatitsneighbours.Therefore,Caietal.[10]pro
posed the wavelet denoising scheme for 1D signal by in
corporating neighbouring coefficients in the thresholding
process.Suppose
, j k
d isthesetofwaveletcoefficientsof
thenoisy1Dsignalandthesquaresumofallthewavelet
coefficients is given by (5).
(5)
2 2 2 2
, , 1 , , 1
S d d d
j k j k j k j k
= + +
+

If the square sum is less than or equal to
2
then the
wavelet coefficient
, j k
d is set to zero. Otherwise, it is
shrinkedaccordingto(6)
2 2
(1 / )
, , ,
d d S
j k j k j k
= (6)
where
2
2 log n o = andnisthelengthofthesignal.
Inthecaseofimagesforeverywaveletcoefficient
,
d
j k
,a
neighbourhood window around it is consi
dered.The window is choosen by having the same num
ber of pixels above, below, and on the left or right of the
pixel to be thresholded. The neighbourhoodwindow size
,
B
j k
JOURNAL OF COMPUTING, VOLUME 3, ISSUE 9, SEPTEMBER 2011, ISSN 2151-9617
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING
WWW.JOURNALOFCOMPUTING.ORG 26
should be 3x3, 5x5, 7x7, 9x9, etc.Fig.1 illustrates a 3x3
neighbourhood window centered at the wavelet coeffi
cient to be thresholded.Different wavelet coefficient sub
bandsarethresholdedindependently.

2 2
(7)
, ,
( , )
,
S d
j k i l
i l B
j k
=
e

Iftheabovesummationhaspixelindicesoutofthewave
letsubbandrange,correspondingtermsinthesummation
are omitted. The wavelet coefficient isthresholded, ac
cordingto(8),

(8)
, , ,
d d
j k j k j k
| =

wheretheshrinkagefactor
, j k
| isasin(9),
2 2
(1 / ) (9)
, ,
S
j k j k
| =
+

here, the + sign at end of the formula means to keep the
positivevalueandissettozerowhenitisnegative,and
2 2
2 log n o = . The neighbourhood window size
around the wavelet coefficient to be thresholded has its
influence on the denoising ability. Larger the window,
relatively smaller will be the threshold. If the size of the
windowaroundthepixelistoolarge,alotofnoisewillbe
kept,soanintermediatewindowsizeof3x3or5x5should
beused.
DonohoandJohnstonedevelopedanadaptivemethodof
selecting a threshold that minimizes the Steins Unbiased
Risk Estimate, which is known as the SUREShrink wave
letthresholdingtechnique.TheadaptivityofSUREShrinkis
achieved by choosing distinct thresholds for each sub
band of each level of the wavelet tree using an efficient
recursive process.This thresholding scheme attempts to
selectthresholdsthatadapttothedataaswellasminim
izetheestimationofmeansquareerror.
NeighShrinkisimprovedbydetermininganoptimalthre
shold and neighbouring window size for every wavelet
subbandusingtheSteinsUnbiasedRiskEstimate[7].
The threshold
s
and neighbouring window size
s
L on
subband s are chosen in order tominimize
( , , ) SURE w L
s
.
(
s
,
s
L )=
,
arg
L
min ( , , ) SURE w L
s
(10)
where
s
and
s
L arederivedassumingthenoiselevel
=1.The standard deviation is estimated from the sub-
band HH by the robust median estimator[2] and is given
by (11)

[{ }]

0.6745
W
s
median HH
o
e
= (11)
The SUREShrinkthresholding method provides an adap
tive thresholding strategy. The performance of this me
thodisdependentonestimatingstrategiesofthewavelet
coefficients of the original image.Determination of the
threshold value is crucial as larger values mayresult in
loss of information while smaller one may allow noise to
continue and retention of the edges is also a problem.
Bilateral filter may help to achieve the target of edge re
tention.
3. BILATERAL FILTER
Thebilateralfilterproposedin[3]isanonlinearfilterthat
does spatial weighted averaging without smoothing
edges.ThisfilterusestwoGaussianfiltersoneforspatial
domain and the other for the intensity domain. The
weights are determined by both spatial distance and in
tensity distance. At a pixel location x, the output of the
filtercanbeformulatedasin(12).
( , ) ( , ) ( )
[ , ] [ , ]
( ) (12)
( , ) ( , )
[ , ] [ , ]
w x y w x y I y
R S
x n n y n n
I x
w x y w x y
R S
x n n y n n

e e
=

e e

Since n is the span of the filter, the window size is of


(2n+1)x(2n+1).The spatial (domain) component ( ) , w x y
s

of the bilateral filter uses Euclidean distance between the
pixel at x and a nearby pixel at y to determine their geo
metric closeness.The absolute difference between the two
intensity values I(x) and I(y) is used to compute the ra
diometric (range) of the bilateral filter, ( ) , w x y
R
.The
domainandrangecomponentsaregivenby(13)and(14)
( )
2
2
2
,
x y
d
w x y e
s
o

= (13)

Fig.1.Anillustrationoftheneighbourhoodwindow
centeredatthewaveletcoefficienttobethresholded.
JOURNAL OF COMPUTING, VOLUME 3, ISSUE 9, SEPTEMBER 2011, ISSN 2151-9617
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING
WWW.JOURNALOFCOMPUTING.ORG 27
( )
2
2
2
,
x y
r
w x y e
R

o

= (14)

The weight function


s
w decreases as the spatial dis
tance between x and y increases, and the function
R
w
decreasesastheradiometricdistancebetweentheintensi
ties
x
and
y
increases. The spatial component of the
weight decreases the inuence of pixels far away from x
to generally reduce blurring, while the radiometric com
ponent diminishes the inuence of pixels with
signicantlydifferentintensitiestokeeptheedgesofdis
tinct image regions sharp. The parameters
d
o and
r
o
controlthebehavioroftheweights.
4. PROPOSED METHOD
In conventionally used linear filtering techniques [1],
therearisesomeproblemssuchasblurringofsharpedges
anddestroyinglinesandotherfinerimagedetails.Dueto
these facts, a hybrid model for despeckling of medical
ultrasound images is proposed, which uses the features of
bilateral filtering and wavelet thresholding and is illu-
strated in Fig.2.

Fig.2.Theproposedmodel

Thebilateralfilteringisappliedbeforewaveletdecompo
sition using DWTand after reconstruction of the image
using IDWT.The parameters of the bilateral filter used
are, 1.6, 0.5
r
d
o o = = and the window size is 3x3 with
one level wavelet decomposition.NeighShrinkSURE tech
niqueisforthresholdingofwaveletcoefficients.
5. RESULTS AND DISCUSSION
Theperformanceofdifferentthresholdingtechniquesare
assessed using standard performance metrics like Signal
to Noise Ratio (SNR), Edge Preservation Index(EPI) and
MeanSquareError(MSE).
TheSignaltoNoiseRatio(SNR)iscalculatedasin(15)

2
10log
2 10
g
SNR
e
o
o
=
| |
|
|
\ .
(15)

where
2
g
o the variance of the noise free reference image
and
2
e
o isthevarianceoferrorbetweentheoriginaland
denoisedimage.
The Mean Square Error (MSE) represents the cumulative
squared error between the denoised and the original im
age.A definition of higher MSE does not indicate thatthe
denoisedimagesuffersfrommoreerrorsinsteaditrefers
to a greater difference between the original and denoise
dimage.ThelowMSEvaluemeansthatthereissignificant
speckle reduction. The formula for the MSE is given in
(16)
1
2
( )
, ,
1 1
M N
MSE x y
i j i j
i j M N
=
= =
(16)

whereMxNissizeoftheimage.
The edge preservation ability is compared by Edge Pre
servationIndex(EPI)andiscomputedusing(17).

( - )( - )

2 2
( - ) ( - )
x x y y
EPI
x x y y
A A A A
=
A A A A
(17)

where x and y in equation (18) are the high pass fil-
tered versions of images x and y, obtained with a 3x3 pix-
el standard approximation of the Laplacian operator. The
larger value of EPI means more ability to preserve edges.
The wavelet filters db4 and sym2 in MATLAB are used
forwaveletdecompositionofonelevel.Theparametersof
the bilateral filter used are, 1.6, 0.5
r
d
o o = = and the
windowsizeis3x3.
The SNR, MSE and EPI svalues for the proposed algo
rithm against the differentwavletthresholdingtechni
queswith bilateral filter before decomposition and after
reconstructionforultrasound image of liver with speckle
noiseofvariance0.02aregiveninTable1andTable2.

JOURNAL OF COMPUTING, VOLUME 3, ISSUE 9, SEPTEMBER 2011, ISSN 2151-9617


HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING
WWW.JOURNALOFCOMPUTING.ORG 28

Fig.3.ComparisongraphofSNRofdifferentmethodsforUltra
soundimageofLiver.

Fig.5ComparisongraphofEPIofdifferentmethodsforUltrasound
imageofLiver

Fig.6.(a)Originalimageofliver(b)NoisyimageafteraddingSpeckle
Noisewithvariance=0.02(c)Denoisedimageafteraddingbilateral
filter with 1.6, 0.5
r d
o o = = , and the window size used is
3x3,decomposition level is 1.The wavelet used is sym2 with Neigh
ShrinkSURE

TABLE 1
Comparisonbetweendifferentthresholding
techniquesfordb4wavelet

THRESHOLDING

SNR

MSE EPI
SoftThresholding(BL) 21.4487 0.0368 0.9229
HardThresholding(BL) 21.6758 0.0359 0.9237
BayesThresholding(BL) 21.7806 0.0355 0.9261
NeighShrinkSURE(BL) 33.8540 0.0285 0.9585
NeighShrinkSURE(WBL) 23.6940 0.0891 0.3693
TABLE 2
Comparisonbetweendifferentthresholding
techniquesforsym2wavelet

THRESHOLDING

SNR

MSE EPI
SoftThresholding(BL) 21.3943 0.0370 0.9226
HardThresholding(BL) 21.6470 0.0360 0.9248
BayesThresholding(BL) 21.7556 0.0355 0.9262
NeighShrinkSURE(BL) 35.1455 0.0246 0.9824
NeighShrinkSURE
(WBL)
22.4962 0.1022 0.0848

Fig.4.ComparisongraphofMSEofdifferentmethodsforUltrasound
imageofLiver.
JOURNAL OF COMPUTING, VOLUME 3, ISSUE 9, SEPTEMBER 2011, ISSN 2151-9617
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING
WWW.JOURNALOFCOMPUTING.ORG 29

A plot of SNR, MSE and EPI for different methods is


shown in Fig.3,Fig.4 and Fig.5 respectively. Fig.6 and
Fig.7 show the original ultrasound image, its noisy ver-
sion and the despeckled image obtained from the pro-
posed method using wavelet filters db4 and sym2.From
Fig.6 and Fig.7 it is observed that the proposed method
yielded better visual quality when compared to the other
methods.
6. CONCLUSION
The ultrasound image of liver with noise variance 0.02 is
used to test different techniques.As the proposed algo
rithm exploits the potential features of both optimal
waveletthresholdingandbilateralfiltering,itoffersbetter
results interms of SNR,MSE and EPI. In addition the vis
ualqualityoftheimageisalsoenhancedwhencompared
to the other methods discussed. For the proposed me
thod, sym2 wavelet is preferred since it produces better
SNR, reduces mean square error and at the same time
edgesarepreserved.
REFERENCES
[1] Thangavel.K, Laurence Aroquiaraj.I and Manavalan.R, Re
movalofSpeckleNoisefrom UltrasoundMedical Imagebased
on Special Filters: Comparative Study, ICGSTGVIP Journal,
ISSN1687398X,Vol.9,IssueIII,June,2009.
[2] D.L. Donoho, I.M. Johnstone, Ideal Spatial Adaptation via
WaveletShrinkage,Biometrika,vol.81,pp.425455,1994.
[3] C. Tomasi and R. Manduchi, Bilateral filtering for gray and
color images, in Proc. 6th Int. Conf. Computer Vision, New
Delhi,India,pp.839846,1998.
[4] S.GraceChang,BinYuandMartinVetterli,AdaptiveWavelet
Thresholding for Image Denoising and Compression, IEEE
transactionsonimageprocessing,Vol.9,No.9September/2000.
[5] D. L. Donoho and I. M. Johnstone, Adapting to unknown
smoothness via wavelet shrinkage, Journal of the American
Statistical Association, vol. 90, no. 432, pp. 12001224,
December1995.
[6] G.Y.Chen,T.D.Bui and A.Krzyzak, Image denoising using
neighbouring wavelet coefficients 2004,ICASSP, IEEE, pp.
917920.
[7] Zhou Dengwen., and Cheng Wengang, Denoising with an
optimalthresholdandneighbouringwindow,Elsevier,Pattern
recognitionletters,2008,pp.16941697.
[8] ChinnaRao.B., and MadhaviLatha.M, Selective neighbouring
wavelet coefficients approach for image denoising, Interna
tionalJournalofComputerScienceandCommunication,Vol.2,
No.1JanuaryJune,2011,pp.7377..
[9] Sudipta Roy, NidulSinha and Asoke K. Sen, A new hybrid
imagedenoisingmethod,InternationalJournalofInformation
Technology and Knowledge Management,Vol. 2, No.2July
December,2010,pp.491497.
[10] T. T. Cai and B. W. Silverman, Incorporating information on
neighbouring coefficients into wavelet estimation, Sank
hya:The Indian Journal of Statistics,Vol. 63,Series B,Pt. 2, pp.
127148,2001.

R.Vanithamani, received her B.E degree in Electronics and Com-


munication Engineering from BharathiarUniversity, Tamil Nadu, India,
in 1991 and M.E degree in Communication Systems from AnnaUni-
versity, Chennai, Tamil Nadu, India, in 2004. She is currently pur-
suing PhD at PSG College of Technology,Coimbatore.She is working
as Assistant Professor in Department of Biomedical Instrumentation
Engineering, Faculty of Engineering, Avinashilingam University for
Women, Coimbatore. She has teaching experience of more than 15
years. Her area of interest includes Medical image processing, Sig-
nal processing.

Dr .G. Umamaheswari,received her B.E degree in Electronics and
Communication Engineering from Madras University, Tamil Nadu,
India, in 1989 and M.E degree in Electronics Engineering from Anna
University, Chennai, Tamil Nadu, India, in 1992 and Doctoral degree
fromBharathiarUniversity, Tamil Nadu, India, in 2006.She is working
as Assistant Professor (SG) in Department of Electronics and Com-
munication Engineering, at PSG College of Technology, Coimba-
tore.She has teaching experience of more than 20 years. Her area of
interest includes Image Processing,Data Communication & Informa-
tion Security. She has published 40 papers at National and Interna-
tional Conferences and Journals.

A.Ajay Krishnan,C.Ilaiyarasan, K. Iswariya and C.G.Kiritikaare
doing Final B.E Electronics and Communication Engineering at PSG
College of Technology, Coimbatore.


Fig.7.(a)Originalimageofliver(b)NoisyimageafteraddingSpeckle
Noisewithvariance=0.02(c)Denoisedimageafteraddingbilateral
filter with 1.6, 0.5
r d
o o = = , and the window size used is
3x3,decomposition level is 1.The wavelet used is db4 with Neigh
ShrinkSURE.

You might also like