research article despeckle filtering for multiscale...

Research ArticleDespeckle Filtering for Multiscale Amplitude-ModulationFrequency-Modulation (AM-FM) Texture Analysis of UltrasoundImages of the Intima-Media Complex

C P Loizou1 V Murray23 M S Pattichis2 M Pantziaris4

A N Nicolaides5 and C S Pattichis6

1 Departement of Computer Science School of Sciences Intercollege 92 Ayias Phylaxeos Street P O Box 51604CY-3507 Limassol Cyprus

2 Departement of Electrical Engineering Universidad de Ingenieria y Tecnologia 2221 Lima Peru3Departement of Electrical and Computer Engineering The University of New Mexico Albuquerque NM 87131 USA4Cyprus Institute of Neurology and Genetics 1683 Nicosia Cyprus5The Vascular Screening and Diagnostic Centre 1080 Nicosia Cyprus6Departement of Computer Science University of Cyprus 1678 Nicosia Cyprus

Correspondence should be addressed to C P Loizou panloicylogosnetcynet

Received 1 December 2013 Accepted 31 January 2014 Published 9 March 2014

Academic Editor Tiange Zhuang

Copyright copy 2014 C P Loizou et al This is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

The intima-media thickness (IMT) of the common carotid artery (CCA) is widely used as an early indicator of cardiovasculardisease (CVD) Typically the IMT grows with age and this is used as a sign of increased risk of CVD Beyond thickness there isalso clinical interest in identifying how the composition and texture of the intima-media complex (IMC) changed and how thesetextural changes grow into atherosclerotic plaques that can cause stroke Clearly though texture analysis of ultrasound images canbe greatly affected by speckle noise our goal here is to develop effective despeckle noise methods that can recover image textureassociated with increased rates of atherosclerosis disease In this study we perform a comparative evaluation of several despecklefiltering methods on 100 ultrasound images of the CCA based on the extracted multiscale Amplitude-Modulation Frequency-Modulation (AM-FM) texture features and visual image quality assessment by two clinical experts Texture features were extractedfrom the automatically segmented IMC for three different age groups The despeckle filters hybrid median and the homogeneousmask area filter showed the best performance by improving the class separation between the three age groups and also yieldedsignificantly improved image quality

1 Introduction

TheWorld Health Organization ranks cardiovascular disease(CVD coronary artery disease cerebrovascular disease andperipheral artery disease) as the third leading cause of deathand adult disability in the industrial world [1] In the UnitedSates alone more than 76 million American adults have oneor more types of CVD of whom about half are estimated tobe age 65 or older It is estimated that by 2015 there will be20 million deaths due to atherosclerosis that will be associ-ated with coronary heart disease and stroke Atherosclerosiscauses enlargement of the arteries and thickening of the

artery walls It begins early in life and silently progresses untilclinical events appear

The intima-media thickness (IMT) is used as a validatedmeasure for the assessment of atherosclerosis [2 3] (seeFigure 1) We present in Figure 1 anatomical locations ofthe common carotid artery (CCA) ultrasound image foratherosclerosis indicating the location of the intima-mediacomplex (IMC) at the far wall The extracted IMC is shownin Figure 1 and has been extracted using an automatedsnake segmentation algorithm as described in [4] In [4] weshowed that automated IMT media-layer thickness (MLT)and intima-layer thickness (ILT) measurements could be

Hindawi Publishing CorporationInternational Journal of Biomedical ImagingVolume 2014 Article ID 518414 13 pageshttpdxdoiorg1011552014518414

2 International Journal of Biomedical Imaging

IntimaMediaAdventitia

Intima-media thickness (IMT) Z5-Z6

Z5Z6Z7

15

16

17

(a)

(b)

Figure 1 Anatomical locations of the common carotid artery ultrasound image components at the far wall The IMT is defined as the layer(band) which is comprised by the bands Z5 and Z6 as demonstrated in (a) The intima-media-complex (IMC) in (b) has been extractedusing automated segmentation as described in [4 24] where the IMTaver = 079mm (between the bands Z5 and Z6 middle bar) IMTmax =08367mm (left bar) IMTmin = 06356mm (right bar) and IMTmedian = 075mm)

carried out successfully It was furthermore proposed that theIMT its thickness [2 4] and its textural characteristics [4 5]may be associated with the risk of developing stroke

Speckle a form of locally correlated multiplicative noisecorrupts medical ultrasound imaging making visual obser-vation difficult [6 7] Clearly though excessive despecklingmay result in the loss of image structureUnfortunately imagestructure is hard to assess and this is especially difficult fornoisy ultrasound images

We are interested in characterizing image structure usinga multiscale Amplitude-Modulation Frequency-Modulation(AM-FM) texture analysis system as described in [8] andby the independent visual assessment of clinical experts whoare asked to judge the quality of the despeckled imagesMultiscale AM-FM models represent nonstationary imagecontent using spatially-varying amplitude and phase com-ponents We use the term instantaneous amplitude (IA) todescribe spatially-varying amplitude components Similarlywe use the term of instantaneous frequency (IF) to describespatially-varying frequency content AM-FM componentsare estimated using a multiscale filterbank that is tuned todifferent frequency bands Thus we use the term multiscaleAM-FM analysis to summarize the analysis The promise ofAM-FM methods for texture analysis can be summarized in(see [9]) as follows (i) they provide physically meaningfultexture features (eg instantaneous frequency in cycles permm) over multiple scales at pixel level resolution (ii)textures can be reconstructed from AM-FM components sothat we can visualize content (iii) we can extract AM-FMdecompositions for different frequency coverage and (iv) wehave the recent development of robust methods for AM-FMdemodulation (see examples in [9ndash11])

Early work in AM-FM image representations has beenreported in [12 13] In [12] Havlicek et al discussed the useof multicomponent AM-FMmodels where Kalman filters areused to track changes in the instantaneous amplitude andinstantaneous frequency components In [13] Havlicek et alintroduced a complex-extension of the Hilbert- transform

for images and suggested the use of quasi-eigenfunctionapproximations (QEAs) for estimating the instantaneousfrequency from discrete signals An early application of AM-FMmethods in medical imaging appeared in [14] in electronmicroscopy In [15] the authors introduced a foveated videoquality assessment method based on the use of continuous-space AM-FM transforms In [16] the authors extended thisresearch to video compression based on local instantaneousfrequency content and the characteristics of the human visualsystem An effective model for texture analysis based onmultidimensional frequency modulation was introduced in[10] The multiscale AM-FM methods that are used in thispaper were first introduced in [9] and first applied in medicalimage analysis in [17]

For one-dimensional signals the empirical mode decom-position pioneered in [18] has been applied to severalapplications The empirical mode decomposition uses aspecial case of AM-FM functions intrinsic mode functionsas basis functions for decomposing signals These intrinsicmode functions decompose fractional Gaussian noise usinga dyadic filter bank as documented in [19] The popularimplementation of the empirical mode decomposition basedon [20] requires that the instantaneous amplitude for thefirst mode be estimated using the extrema of the input signaland then using interpolation to determine the envelopeThis approach is clearly sensitive to noise since the envelopeestimation is very sensitive to additive noise artefacts We arenot aware of any robust extensions that avoid the significantnoise artefacts or robust 2D extensions required for thecurrent application On the other hand the use of multiscaleAM-FM decomposition allows us to deal with noise throughthe use of suitable designed band-pass filters [9] For thecurrent application the effectiveness of this approach isdemonstrated in [8]

While AM-FM decompositions can effectively modelnonstationary texture it is also clear that we want to avoidmeasuring noise We want to investigate the use of despecklefiltering for reducing the levels of noise in the image Yet

International Journal of Biomedical Imaging 3

excessive despeckling can destroy the nonstationary signalcontent It can result in a reduction of the discriminatorypower of the texture analysis system introduced in [8 9 11]andwe expect this to also be detected by the clinical experts asa reduction in image quality Ultimately we are interested inthe application of despeckle filteringmethods that can lead toimprovements in computerized texture analysis methods aswell as significant improvements in the image quality judgedby clinical expertsTheuse ofAM-FM features to characterizethe CCA will also provide complimentary information toclassical texture analysis features like the gray-scale mediancontrast and coarseness AM-FM texture features can beassociated with the progression of cardiovascular risk fordisease and the risk of stroke with age

Prior research on the use of despeckle filtering on CCAplaque images was reported by Loizou et al in [21 22]In this paper we investigate the use of two additionaldespeckle filtering methods namely the despeckle filterKuhawara (see Section 231 (2)) and the hybrid median fil-tering method (see Section 232) and study their effects onimage quality and multiscale AM-FM texture analysis on thethin structures of the intima media As we described in thispaper the new filters gave significantly better results thanprior research in [21 22]

Formally an input image 119891(119909 119910) is expressed as a sum ofAM-FM components using [9 11] as follows

119891 (119909 119910) =

119873

sum119899=1

119886119899(119909 119910) cos120593

119899(119909 119910) (1)

where 119886119899(119909 119910) denotes the 119899th instantaneous amplitude

(IA) function 120593119899(119909 119910) denotes the 119899th instantaneous phase

(IP) component and 119899 = 1 2 119873 indexes the dif-ferent AM-FM components For each AM-FM component119886119899(119909 119910) cos120601

119899(119909 119910) we define the instantaneous frequency

(IF) by nabla120601119899(119909 119910) and the magnitude of the IF given by

nabla120601119899(119909 119910) Textural characteristics are described in terms

of the IA and the IF extracted from different frequency scalesHere frequency scales are defined based on the IFmagnitudeand are further classified into lowmedium- and high-frequency scales

In [8] AM-FM analysis on the IMC media-layer (ML)and intima-layer (IL) structures showed that there aresignificant differences in AM-FM texture features extractedfrom different age groups and different sexes In this paperwe investigate the AM-FM texture features that can showsignificant differences and also appear to be improvingin simulations involving the use of despeckle filtering onground-truth signals

The rest of the paper is organized as follows In thenext sections materials and methods experimental resultsdiscussion concluding remarks and future work are given

2 Materials and Methods

21 Ultrasound Images Acquisition A total of 100 B-modelongitudinal ultrasound images of the CCA were recordedusing the ATL HDI-3000 ultrasound scanner (AdvancedTechnology Laboratories Seattle USA) [23] as described in

[8]The images were recorded at the Cyprus Institute of Neu-rology and Genetics in Nicosia Cyprus The recordings werecarried out in agreement with the Cyprus national bioethicscommittee rules on clinical trials and after patientrsquos writtenconsent For the recordings we used a linear probe (L74) witha recording frequency of 4ndash7MHz [23] a velocity of 1550msand a1 cycle per pulse which resulted in a wavelength (spatialpulse length) of 022mm and an axial resolution of 011mmFurthermore the scanner is equipped with 64 elements finepitch high-resolution 38mm broadband array an acousticaperture of 10 times 8mm and a transmission focal range of 08ndash11 cm

The B-mode scan settings were adjusted to allow forthe maximum dynamic range with a linear postprocessingcurve In order to ensure that a linear postprocessing curveis used these settings were preselected (by selecting theappropriate start-up presets from the software) and wereincluded in the part of the start-up settings of the ultrasoundscanner The position of the probe was adjusted so that theultrasonic beam was vertical to the artery wall The time gaincompensation (TGC) curve was adjusted (gently sloping) toproduce uniform intensity of echoes on the screen but it wasvertical in the lumen of the artery where attenuation in bloodwasminimal so that echogenicity of the far wall was the sameas that of the near wall The overall gain was set so that theappearance of the carotid wall was assessed to be optimal andslight noise appeared within the lumen It was then decreasedso that at least some areas in the lumen appeared to be freeof noise (black) Thus the standardization effort follows theACSRS acquisition guidelines as detailed in [24]

Images were acquired with the subjectrsquos head rotated by45∘ away from the study side A single longitudinal imagewas captured at the distal end of the CCA during the diastolicphase of a cardiac cycle All captured images were revealingoptimal visualization of the IMC of the far wall and the nearwall of the CCA at the same time thus corresponding to amidline horizontal longitudinal representation of the CCAwalls

During image acquisition the sonographers varied spa-tial resolution to provide optimal imaging at different depths[4 5] However without standardizing image resolution theestimated AM-FM components would not be comparableTo see this note that continuous-space image frequenciesare expressed in cycles per millimeter and unless we have acommon spatial resolution the estimated digital frequencieswould correspond to different continuous-space (analogue)frequencies As a result we then had to use bicubic splineinterpolation to resize all digital images to a standard pixeldensity of 1666 pixelsmmThe use of bicubic spline interpo-lation does not add additional information to the image Inother words interpolation does not recover high-frequencycontent that was not present in the original acquisitionThuswhen comparing among images it is important to note thathigh-frequency content is comparable to the extent that it isshared among all of the resolutions Also note that most ofthe images were acquired at the target resolution In otherwords we only made small corrections to spatial resolution

The images were also intensity normalized as describedin [25] where a manual selection of blood and adventitia


performed by the user of the system is required The gray-scale intensity normalized image was obtained through alge-braic (linear) scaling of the image by linearly adjusting theimage so that the median gray level value of the blood was0ndash5 and the median gray level of the adventitia (artery wall)was 180ndash190 [25 26] The images were recorded from 42female and 58 male asymptomatic patients These subjectshad not developed any clinical symptoms such as a strokeor a transient ischemic attack The primary-care physiciansinformed the subjects of our stroke-prevention researchstudy Overall patientsrsquo ages varied between 26 and 95 yearswith amean age of 54 yearsThe images were partitioned intothree different age groups In the first group we included 27images from patients who were younger than 50 years oldIn the second group we had 36 patients who were 50ndash60years old In the third group we included 37 patients whowere older than 60 years

22 Simulated Images To understand the effects of despecklefiltering on AM-FM estimation we perform a simulationon a synthetic image (see Figure 2(a)) which was created toresemble clinical ultrasound images The synthetic IA imagewas 1024 times 1024 pixels with two strips For the simulation weset the IA and the IF as follows

119860 (119909 119910) = 158120587

75le 120601119909le

120587

45 120601

119910= minus120601119909

for 0 le 119909 le 272 (dark background)

119860 (119909 119910) = 250120587

65le 120601119909le

120587

55 120601

119910= 120601119909

for 273 le 119909 le 306 (bright upper strip)

119860 (119909 119910) = 102120587

75le 120601119909le

120587

45 120601

119910= minus120601119909


119860 (119909 119910) = 250120587

65le 120601119909le

120587

55 120601

119910= 120601119909

for 703 le 119909 le 750 (bright lower stip)

119860 (119909 119910) = 182120587

75le 120601119909le

120587

45 120601

119910= minus120601119909

for 751 le 119909 le 1023 (dark backround)

(2)

The resulting synthetic image is shown in Figure 2(b)We add multiplicative noise (see Figure 2(c)) to generate119892119894119895

= 119891119894119895+ 119899119894119895119891119894119895 where 119892

119894119895and 119891

119894119895represent the noisy

and the original images respectively and 119899119894119895is a uniformly

distributed random noise with zero mean and noise variance1205902119899= 007 We show the results after applying a low frequency

AM-FM estimation in Figures 2(d)ndash2(h) Figure 2(d) illus-trates the instantaneous amplitude (IA) estimation from thenoisy image while Figures 2(e)-2(f) show the IF estimationfor the 119909- and 119910-directions In Figure 2(f) we show the IAestimation from the denoised image after using the hybridmedian despeckle filter Finally we show the IF estimationusing this method in Figures 2(g)-2(h)

23 Despeckle Filtering

231 Linear Filtering

(1) First Order Statistics Filtering (DsFlsmv DsFwiener)Thesefilters utilize the first order statistics such as the variance andthe mean of a pixel neighbourhood and may be describedwith a multiplicative noise model [21 22 27] Hence thealgorithms in this class may be traced back to the followingequation

119891119894119895= 119892 + 119896

119894119895(119892119894119895minus 119892) (3)

where 119891119894119895

is the estimated noise-free pixel value 119892119894119895

is thenoisy pixel value in the moving window 119892 is the local meanvalue of a 5 times 5 rectangular region surrounding and includingpixel 119892

119894119895 119896119894119895is a weighting factor with 119896 isin [0 sdot sdot sdot 1] and 119894 119895

are the pixel coordinates The factor 119896119894119895

is a function of thelocal statistics in a moving window and can be found in theliterature [21 22] as

119896119894119895=

1 minus 11989221205902

1205902 (1 + 1205902119899) (4)

The values 1205902 and 1205902119899represent the variance in the moving

window and the variance of noise in the whole image respec-tively The noise variance is calculated in the logarithmicallycompressed image using the average noise variance overa number of windows with dimensions considerably largerthan the filtering window [21 22] The Wiener filter uses apixel-wise adaptive method [6 7 22 28] and is implementedas given in (3) with a different weighting factor 119896

119894119895= (1205902 minus

1205902119899)1205902 [13] For both despeckle filters which are proposed in

this subsection the moving window size was 5 times 5 and thenumber of iterations was set to two

(2) Homogeneous Mask Area Filtering (DsFkuhawaraDsFlsminsc) The Kuhawara despeckle filter is a 1D filteroperating in a 5 times 5 pixel neighbourhood by searching for themost homogenous neighbourhood area around each pixel[22 29] The middle pixel of the 1 times 5 neighbourhood is thensubstituted with the median gray level of the 1 times 5 mask Thefilter was iteratively applied 2 times on the image

The DsFlsminsc is a 2D filter operating in a 5 times 5pixel neighbourhood by searching for the most homogenousneighbourhood area around each pixel using a 3 times 3 subsetwindow [21 22]Themiddle pixel of the 5 times 5 neighbourhoodis substitutedwith the average gray level of the 3times 3maskwiththe smallest speckle index 119862 where 119862 for log-compressedimages is given by

119862 =1205902

119904

119892119904

(5)

where 1205902119904and 119892

119904represent the variance and mean of the

3 times 3 window The window with the smallest 119862 is themost homogenous semiwindow which presumably does notcontain any edge The filter is applied iteratively one time inthe image


(a) (b) (c) (d)

(e) (f) (g) (h)

(i)

Figure 2 Results using a low frequency scale for the AM-FM methods (a) Noise-free synthetic IA image (b) Noise-free synthetic AM-FMimage with low frequency information (c) AM-FM image corrupted with speckle noise (d) IA estimation from the noisy image of (c) (e)IFx estimation from the noisy image (f) IFy estimation from the noisy image (g) IA estimation from the denoised image using the hybridmedian filter (h) IFx estimation from the denoised image using the hybridmedian filter (i) IFy estimation from the denoised image using thehybrid median filter Under each image we show the same AM-FM results but with a focus (zoom) on the top strip for better visual analysispurposes

232 Nonlinear Filtering (DsFmedian DsFhybridmedian)The first filter proposed in this subsection [22] is a medianfilter applied over windows of size 5 times 5 This is extendedin the hybrid median despeckle filter [30] which producesthe average of the outputs generated by median filteringwith three different windows (cross shape window times-shapewindow and normal window)

233 Diffusion Filtering (DsFsrad DsFnldif)

(1) Speckle Reducing Anisotropic Diffusion Filtering Specklereducing anisotropic diffusion is described in [7] It is basedon setting the conduction coefficient in the diffusion equationusing the local image gradient and the image LaplacianThe speckle reducing anisotropic diffusion filter [7] uses twoseemingly different methods namely the Lee [27] and theFrost diffusion filters [28] A more general updated functionfor the output image by extending the PDE versions of thedespeckle filter is [7 22]

119891119894119895= 119892119894119895+1

120578119904

div (119888srad (1003816100381610038161003816nabla119892

1003816100381610038161003816) nabla119892119894119895) (6)

The diffusion coefficient for the speckle anisotropic diffusion119888srad(|nabla119892|) is derived [7] as

1198882

srad (1003816100381610038161003816nabla119892

1003816100381610038161003816) =(12)

10038161003816100381610038161003816nabla119892119894119895

10038161003816100381610038161003816

2

minus (116) (nabla2119892119894119895)2

(119892119894119895+ (14) nabla2119892119894119895)

2 (7)

It is required that 119888srad(|nabla119892|) ge 0The above instantaneouscoefficient of variation combines a normalized gradientmagnitude operator and a normalized Laplacian operator toact like an edge detector for speckle images High-relativegradient magnitude and low-relative Laplacian indicate anedge The filter proposed in this subsection utilizes specklereducing anisotropic diffusion after (5) with the diffusioncoefficient 119888srad(|nabla119892|) in (7) [7]

(2) Coherent Nonlinear Anisotropic Diffusion Filtering Thisfilter extends the conduction coefficient using a symmetricpositive semidefinite diffusion tensor [31] with the param-eters as given in [22] Therefore the filter will take thefollowing form

119889119892119894119895119905

119889119905= div [119863nabla119892] (8)

where 119863 isin R21199092 is a symmetric positive semidefinitediffusion tensor representing the required diffusion in both


gradient and contour directions and hence enhancing coher-ent structures as well as edges The design of 119863 as well asthe derivation of the coherent nonlinear anisotropic diffusionmodel may be found in [31] and is given as

119863 = (1205961 1205962) (

1205821

0

0 1205822

)(1205961198791

1205961198792

) (9)

with

1205821=

120572(1 minus(1205831minus 1205832)2

1199042) if (120582

1minus 1205822)2le 1199042

0 otherwise(10)

1205822= 120572 (11)

where the eigenvectors 1205961 1205962and the eigenvalues 120582

1 1205822

correspond to the directions of maximum and minimumvariations and the strength of these variations respectivelyThe flow at each point is affected by the local coherencewhich is measured by (120583

1minus 1205832) in (10) The parameters used

in this work for the coherent nonlinear anisotropic diffusionfilter were 1199042 = 2 and 120572 = 09 which were used for thecalculation of the diffusion tensor 119863 and the parameter stepsize 119898 = 02 which defined the number of diffusion stepsperformed The local coherence is close to zero in very noisyregions and diffusion becomes isotropic (120583

1= 1205832= 120572 =

09) whereas in regions with lower speckle noise the localcoherence corresponds to (120583

1minus 1205832)2gt 1199042 [31]

24 IMC Snakes Segmentation All images were automat-ically segmented to identify the IMC regions Automaticsegmentation was carried out after image normalization anddespeckle filtering using the snakes segmentation systemproposed and evaluated on ultrasound images of the CCAin [4] The segmentation system is based on the Greedyactive contour algorithm [32] Using the definitions given inFigure 1 we first segment the IMC [33] by extracting theI5 (lumen-intima interface) and the I7 boundaries (media-adventitia interface) In order to achieve standardization inextracting the thickness from the IMC segments with similardimensions the following procedure was carried out Aregion of interest of 96mm (160 pixels) in length was firstextracted This was done by estimating the center of the IMCarea and then selecting 48mm (80 pixels) left and 48mm (80pixels) right of the center of the segmented IMC Selection ofthe same IMC length fromeach image is important in order tobe able to extract comparable measurements between imagesand subject groups

We note that there was no significant difference betweenthe manual and automated segmentation measurements forthe IMC [4]

25 Texture Analysis Using Multiscale Amplitude-ModulationFrequency-Modulation (AM-FM) Methods Multiscale AM-FM texture features were extracted over different channelfilters We refer to [8] for full details on the approach Herewe provide a brief summary for completeness

First a complex valued image is obtained using anextended 2D Hilbert operator The operator is implementedby taking the 2D FFT of the input image zeroing out theupper two frequency quadrants multiplying the remainingfrequency components by 2 and taking the inverse 2D FFT

The complex-valued output image is processed througha collection of 2D channel filters with passbands restrictedover the (nonzeroed) lower two quadrants We refer to [8]for a clear description of the filterbank Here we simplynote that we have lowmedium- and high-frequency scalesbased on the passband frequency magnitudes Based on thedyadic frequency decomposition we have (1) low-frequencycomponents from 104 to 295 cyclesmm that correspondto instantaneous wavelengths (IWs) from 566 to 16 pixels(034ndash096mm) (2) medium-frequency components from208 to 589 cyclesmm that correspond to IW from 283 to8 pixels (017ndash048mm) and (3) high-frequency componentsfrom 417 to 1179 cyclesmm that correspond to IW from 141to 4 pixels (0085ndash024mm) [8]

AM-FM demodulation is carried out separately for thelow- medium- and high-frequency scales Adaptively foreach frequency-scale at each image pixel we estimate IA bytaking the absolute value of the channel response Then ateach pixel among the channel responses of each scale weselect the channel that gives the maximum IA The phase foreach scale is then estimated by taking the phase response ofthe dominant channel

An adaptive method is used for estimating IF compo-nents The IF components are estimated using

119889120601 (119909 119910)

119889119909cong1

119899arccos(

119891 (119909 + 119899 119910) + 119891 (119909 minus 119899 119910)

2119891 (119909 119910)) (12)

and similarly for 119889120593119889119910 where 119891 denotes the estimatedFM image cos120593(119909 119910) cos120593(119909 119910) In (12) we consider 119899 =

1 2 3 4 for the low frequencies 119899 = 1 2 for the mediumfrequencies and 119899 = 1 for the high frequencies Among theIF estimates we select the one that generates the minimumargument to the arccos function This is expected to be themost accurate [11] The AM-FM texture features are thenformed by taking the 32-bin histograms of the resulting IAand IF estimates from each one of the three frequency scales

The Mann-Whitney rank sum test (for independentsamples of different sizes) [34] was used in order to identifyif there were significant differences (S) or not (NS) betweenthe extracted AM-FM texture features at119875 lt 005The resultswill be explained in Section 32 and summarized in Table 3

26 Visual Evaluation by Experts The visual evaluation wascarried out according to the ITU-R recommendations withthe Double Stimulus Continuous Quality Scale (DSCQS)procedure [21 22] The 100 segmented IMC structures ofthe CCA were evaluated visually by two vascular experts acardiovascular surgeon and a neurovascular specialist beforeand after despeckle filtering For each case the original andthe despeckled images were presented at random andwithoutlabeling to the two experts The experts were asked to assigna score in the one to five scale corresponding to low and


high subjective visual perception criteria Five was given to animage with the best visual quality Therefore the maximumscore for a filter is 500 if the expert assigned the score of fivefor all the 100 images For each filter the score was divided byfive to be expressed in percentage format The experts wereallowed to give equal scores to more than one image in eachcase For each class and for each filter the average score wascomputed

We have furthermore used the recently proposed NIQEindex assessment tool [35] for objective evaluation of thequality of the images The tool is based on the constructionof a quality aware collection of statistical features based ona simple and successful space domain natural scene statisticmodelThese features are derived from a collection of naturalundistorted images The quality of the despeckled image isexpressed as a simple distance metric between the modelstatistics and those of the original image A software releaseof the NIQE index is available at httpliveeceutexaseduresearchQualityindexhtm

3 Experimental Results

31 Artificial Carotid Image Despeckle filtering was evalu-ated on an artificial carotid artery image corrupted by specklenoise (see Figure 2(a)) as described in the materials andmethods section Figure 2 presents the results using a lowfrequency scale for the AM-FM methods In Figure 2(a) wepresent the original synthetic image while in Figures 2(b)and 2(c) we show the original synthetic image with lowfrequency information and the image from Figure 2(b) withspeckle noise respectively In Figure 2(d) we show the IAestimation from the noisy image while in Figure 2(e) wepresent the IFx estimation from the noisy synthetic image InFigure 2(f) the IFy estimation from the noisy synthetic imageis illustrated while in Figure 2(g) the IA estimation from thedespeckled image using the hybrid median filter is shownFinally in Figures 2(h) and 2(i) we present the IFx estimationfrom the despeckled artificial image using the hybrid medianfilter and the IFy estimation from the despeckled image usingthe hybrid median filter respectively Below each figure wepresent the zoom of the top part of the synthetic image AM-FM results including the top strip for visualization purposes

Table 1 presents the results of despeckle filtering demon-strating its advantages applied to a synthetic AM-FM exam-ple We note significant noise estimation improvements forthe narrow strip for both the IF component for both the119909- and 119910-direction Here we were not interested in the IAerror since it was piecewise-constant and estimation could besignificantly improved by simply using median-filtering onthe estimated values The results are reported over the low-frequency scales where most of the image energy is usuallyconcentrated

32 Real Carotid Ultrasound Images We show in Figure 3an example of the original IMC ultrasound image in thefirst column and the corresponding despeckled images withhybrid median and Kuhawara filters in the second and thirdcolumns respectively The figure also shows the logarithmicviews of the IA components LIA MIA and HIA the IF

components LIF MIF HIF and the reconstructed FM com-ponent The last row shows the FM demodulation (integralof the IF) of the images in the low frequencies For bettervisualization the images have been interpolated to be 300 times20 pixels In this Figure image regions where the estimatedinstantaneous frequency is outside the low-scale frequencyrange are depicted as dark (black) By comparing the figures(in the three different columns of Figure 3) it is clear thatthe hybrid median approach has improved the estimationsignificantly In other words there are fewer dark regionsin the results of the second column than there are in thethird column of Figure 3 (see Log LIA column) For theKuhawara filter (see Figure 3 third column) segmentationgave a slightly expanded version of the original segmentationresults Furthermore the area of the dark regions appearsto be greater than that for the hybrid median filter Alsoin this case the Kuhawara filter does not show significantimprovements over the results on the original image

The first part of Table 2 tabulates the results of the visualevaluation of the original and despeckled IMC images madeby two experts a cardiovascular surgeon and a neurovascularspecialist It is clearly shown in Table 1 that the best despecklefilter is the hybrid median with a score of 73 followedby Kuhawara with a score of 71 It is interesting to notethat these two filters were scored with the highest evaluationmarkings by both experts The other filters gave poorerperformance like the DsFnldif DsFlsminsc and DsFsradthat gave an evaluation score of 62 58 and 56 respec-tively The third row of Table 2 presents the overall averagepercentage () score assigned by both experts for each filterThe second part of Table 2 illustrates the objective evaluationperformed for all despeckled filters investigated betweenthe original and the despeckled images by using the NIQEindex It is shown that the best results were obtained by thehybrid median despeckle filter (NIQE = 0987) followed bythe Kuwahara (NIQE = 0981) despeckle filterThe last row ofTable 2 presents the final filter ranking

Table 3 presents the statistical analysis between the LowMedium and High AM-FM features extracted from theIMC for the three different age groups below 50 (lt50)between 50 and 60 (50ndash60) and above 60 (gt60) years oldbased on the Mann-Whitney rank sum test showing onlythe features that exhibited statistically significant differenceat 119875 lt 005 It is shown in Table 3 that all the despeckle filtersinvestigated increased the number of AM-FM features thatexhibited significant differences between the different ages(compare the first row for the Original images versus the restof the columns that represent the despeckled images) Morespecifically using the hybrid median filter we can use thefollowingAM-FMcomponents that demonstrated significantdifferences for differentiating between the different IMC agegroups

(a) For the lt50 and 50ndash60 years old use the LIA compo-nent

(b) For the lt50 and gt60 years old use the MIA andorthe LIF and the HIF components

(c) For the 50ndash60 and gt60 years old use the LIA andorthe LIF and the MIF components


Table 1 Despeckle filtering demonstrating its advantages applied to a synthetic AM-FM example (see text for details) Note significant noiseestimation improvements for the narrow strips

Frequency componentx-component of low instantaneous frequency

(LIFx)y-component of low instantaneous frequency

(LIFy)Backgrounds Strips Combined Backgrounds Strips Combined

Noise-free low-scale AM-FM(upper bound of what can beachieved)

39119864 minus 06 76119864 minus 02 27119864 minus 03 29119864 minus 02 55119864 minus 01 48119864 minus 02

Speckled image low-scale AM-FMestimation (no despeckling) 55119864 minus 04 12119864 minus 01 49119864 minus 03 33119864 minus 02 49119864 minus 01 49119864 minus 02

Despeckling using DsFlsmv 73119864 minus 04 51119864 minus 02 25119864 minus 03 68119864 minus 02 26119864 minus 01 75119864 minus 02

Despeckling usingDsFhybridmedian 16119864 minus 03 46119864 minus 02 31119864 minus 03 62119864 minus 02 41119864 minus 01 74119864 minus 02

Despeckling using DsFKuhawara 69119864 minus 03 73119864 minus 02 93119864 minus 03 64119864 minus 02 48119864 minus 01 79119864 minus 02

Table 2 Percentage scoring of visual and objective evaluation of the original and despeckled images by the experts and the natural imagequality evaluation (NIQE) index Bolded values show best performance

Experts original First order statistics Homogeneous mask area Non-linear filtering DiffusionDsFlsmv DsFwiener DsFkuhawara DsFlsminsc DsFmedian DsFhybridmedian DsFnldif DsFsrad

Visual EvaluationExpert 1 33 26 27 65 51 43 71 59 61Expert 2 40 30 23 77 65 47 75 65 51Average 37 28 25 71 58 45 73 62 56

Objective EvaluationNIQE 0861 0834 0810 0981 0956 0901 0987 0962 0923Ranking 7th 8th 9th 2nd 4th 6th 1st 3rd 5thNIQE Naturalness image quality evaluation

Table 3 Statistical analysis between the low medium and high AM-FM features extracted from the IMC for the automated segmentationmeasurements for the three different age groups below 50 (lt50) between 50 and 60 (50ndash60) and above 60 (gt60) years old based on theMann-Whitney rank sum test for all despeckle filtering techniques Only the features that exhibited statistical significant differences at P lt

005 are shownFilter name Age groups 50ndash60 gt60 Score Table 2 ranking

Original (see also) [8]lt50 MIA 3 7th50ndash60 LIAHIF

DsFlsmv lt50 MIAHIF MIALIF 5 8th50ndash60 MIF (04)

DsFwiener lt50 MIAHIA MIAHIAMIFHIFLIA 7 9th50ndash60

DsFKuhawara lt50 LIA MIALIFHIF 6 2nd50ndash60 LIAMIF

DsFlsminsc lt50 LIAHIF HIF 5 4th50ndash60 LIAMIA

DsFmedian lt50 MIALIFHIF 4 6th50ndash60 LIA

DsFhybridmedian lt50 LIA MIALIFHIF 7 1st50ndash60 LIALIFMIF

DsFnldif lt50 LIAMIALIF MIAHIA 9 3rd50ndash60 LIAMIAHIAHIF

DsFsrad lt50 LIAHIF MIALIF 7 5th50ndash60 LIAMIAHIALIA MIA HIA Low Medium High instantaneous amplitude LIF MIF HIF Low medium high instantaneous frequency Score Illustrates the numbers ofsignificantly different features


Table 4 Comparison of the mean standard deviation (STD) median and different quartile ranges between the high medium and low AM-FM features extracted from the IMC for the three different age groups below 50 (lt50) between 50 and 60 (50ndash60) and above 60 (gt60) yearsold for the original the DsFhybrimedian and the DsFkuhawara filters Here the IA and IF values have been pre-multiplied by 100 for bettervisualization Recall that the original images were normalized to a maximum brightness value of 1Thus the IA values represent a percentageof the maximum input image intensity The instantaneous frequency magnitude IF is measured in cyclesmm (100x Magnified by 100)

Mean STD Median P5 P10 P25 P75 P90 P95

Original

LIA lt50 247 045 229 181 19 22 28 30 34LIA 50ndash60 288 066 262 231 25 25 27 40 43LIA gt60 26 042 248 191 226 233 281 333 335

LIF lt50 145 41 145 140 141 142 146 147 154LIF gt60 144 52 144 135 137 139 145 150 153LIF 50ndash60 143 38 145 136 137 139 146 147 147

MIF 50ndash60 285 80 284 275 276 280 289 296 303MIF gt60 284 87 282 273 275 278 289 295 303

HIF lt50 574 28 574 545 546 556 578 599 641HIF gt60 566 16 564 541 545 557 574 585 597

DsFhybrimedian



MIF 50ndash60 284 822 283 273 274 279 287 296 301MIF gt60 283 835 281 274 275 276 290 296 302

HIF lt50 564 1776 568 539 541 550 570 586 599HIF gt60 556 126 553 540 543 545 566 575 580

DsFKuhawara



MIF 50ndash60 284 822 283 273 274 279 287 296 301MIF gt60 283 83 281 274 275 276 290 296 303

HIF lt50 564 178 568 539 541 550 570 586 599HIF gt60 556 125 553 540 543 546 566 575 580

IMC Intima-media-complex

Also the Kuhawara despeckle filter that can be used showed asimilar performance as above except for the LIF componentin (c)

Table 4 presents a comparison of the mean standarddeviation (STD) median 5 10 25 75 90 and95 quartiles between the high medium and low AM-FMfeatures extracted from the IMC for the original and thedespeckled filters DsFhybrimedian and DsFkuhawara for thethree different age groups below 50 (lt50) between 50 and60 (50ndash60) and above 60 (gt60) years old Only those featuresthat showed significant differences in almost all different agegroups according to Table 3 are presentedThe results indicatethat for the high instantaneous frequency (HIF) magnitude

median for the IMC the 75th percentile value of the gt60 agegroup remains lower than the median value of the lt50 agegroup (cyclesmm) Furthermore we note the original HIFstandard deviation of 0028 (lt50) and 0016 (gt60) cyclesmmversus the hybrid median filter with 01776 (lt50) and 0126(gt60) cyclesmm and the Kuhawara filter with 0178 (lt50)and 0125 (gt60) cyclesmm It is clear that image despeck-ling produces more than a 5-fold increase in the spreadof the high instantaneous frequency range This suggeststhat high-frequency texture information does benefit fromdespeckling

This is a positive result since speckle noise can havedetrimental effects on high frequencies Another significant


Scale Original DsFhybridmedian DsFkuhawara

Log LIA

Log MIA

Log HIA

LIF

MIF

HIF

FM

Figure 3 AM-FM analysis of the IMC original (1st column) and despeckled images with the DsFhybrimedian (2nd column) andDsFkuhawara (3rd column) from a male asymptomatic subject aged 49 In the 1st row the IMT measurements of the original (IMTaver =066mm IMTmax = 0827mm IMTmin = 0526mm IMTmedian = 068) DsFhybrimedian (IMTaver = 069mm IMTmax = 091mm IMTmin =053mm IMTmedian = 069) andDsFkuhawara (IMTaver = 063mm IMTmax = 077mm IMTmin = 049mm IMTmedian = 063mm) are shownIn the following rows we present the AM-FM components of the instantaneous amplitude of Log of LIAMIA and HIA and of instantaneousfrequency of LIF MIF and HIF The last row shows the FM demodulation (integral of the IF) of the images in the low frequencies For bettervisualization the images have been interpolated to be 300 times 20 pixels

difference is observed in the standard deviation for the LowInstantaneous Frequency (LIF) In this case it is interestingto compare the LIF for 50ndash60 that gives 0038 cyclesmmfor the despeckled images versus 0058 cyclesmm for thehybrid median filter This shows a significant increase inthe low-frequency magnitude spread in the results for thehybrid median filter As discussed in Figure 2 successfulAM-FM estimation over larger regions of the image alsocontribute to this larger spread On the other hand note asomewhat smaller spread for the gt60 group (0052 versus0047 cyclesmm) Overall it is clear that the IF spreadsfor the despeckled images tend to either have a significantincrease or remain essentially the same as the original(speckled) images

4 Discussion and Concluding Remarks

Clinically no significant changes are anticipated in the IMTbefore the age of 50 [36] It was shown in [4] (based on asimilar group of subjects with the one used in this study aswell) that between the ages of 50 and 60 the age borderlinefor the young (lt50 years) and the adult (gt60 years) ages

and a subtle increase in IMT can be demonstrated and IMCtextural changes can be initially observed Above the ageof 60 IMT increases and changes in the IMC are moreevident Moreover most of the stroke incidences in this agegroup are associated with the carotid atherosclerosis diseaseSignificant texture changes between the different age groupswere reported in [5] for age and sex More specifically(a) some of the texture features can be associated with theincrease (difference variance entropy) or decrease (grey scalemedian (GSM)) of patientrsquos age (b) the GSM of the medialayer (ML) falls linearly with increasing ML thickness (MLT)and with increasing age (c) the GSM of male subjects islarger than that of female subjects (see Figure 4) and (d)maleand female subjects may be better distinguished using texturefeatures extracted from the IMC

Despeckle filtering improved the class separationbetween the three age groups as measured by the numberof significantly different AM-FM texture features Theimprovements were also reflected in better instantaneousfrequency estimation and also the significantly improvedimage quality as evaluated by two clinical experts In termsof performance the nonlinear hybrid median despecklefilter (DsFhybridmedian) gave the best results followed by


the homogeneous mask area filter (DsFkuhawara) Morespecifically using the hybrid median filter we can use thefollowingAM-FMcomponents that demonstrated significantdifferences for differentiating between the different IMCage groups (a) for the lt50 and 50ndash60 years old use the LIAcomponent (b) for the lt50 and gt60 years old use the MIAandor the LIF and the HIF components (c) for the 50ndash60and gt60 years old use the LIA andor the LIF and the MIFcomponents Also the Kuhawara filter that can be usedshowed a similar performance as above except for the LIFcomponent in (c)

These filters combined with multiscale AM-FM analysiscan be used to differentiate between the three age groupsinvestigated (that could loosely correspond to low mediumand high-risk) It should be noted that this is not thecase when the nondespeckled AM-FM analysis was usedas documented in [8] (and shown in Table 3) In factalmost all despeckling filters improved class separation overthe nondespeckled filters In Table 3 this is reflected inthe increased number of significant (despeckled) AM-FMfeatures that can be used to differentiate between classes

The intensity normalization method used in this studywas found to be helpful in the manual contour extraction[22 26] as well as the snakes segmentation of the IMC[4 33] and the extraction and evaluation of texture featuresfrom ultrasound images of the CCA [5] The method usesprior knowledge of the high- and low-intensity values of theadventitia and blood so that the new intensity histogramof the lesion has its maximum peak close to its averagegray-scale value [25] Moreover this method increased theclassification accuracy of different plaque types as assessedby the experts [37] Ultrasound image normalization wascarried out prior to segmentation of the IMTon carotid arteryultrasound images for increasing the image contrast in [38]Using the above intensity normalisationmethod theAM-FMtexture analysis proposed in this study may be also applieddirectly to the logarithmic compressed images

The proposed despeckle filtering methods have beenevaluated on 550 ultrasound images of the CCA togetherwith other despeckle filters in [21 22 26] using texturefeatures image quality metrics observers evaluation andkNN classification More specifically it was shown thatthese filters can be used to improve the class separationbetween asymptomatic and symptomatic subjects based onthe statistics of the extracted texture features and improve theclassification success rate and the visual evaluation by expertsA number of other despeckle filtering methods have beenproposed by other researches in the last 20 years for increas-ing the accuracy of edge detection in images [39] improvethe image visual perception evaluation [7 21 22 27ndash31] andaid the segmentation of the IMC and atherosclerotic carotidplaque in ultrasound images [4 33] or videos of theCCA [40]Recently a despeckle filtering toolbox for ultrasound videoshave been proposed [41] which can also be downloaded inexecutable code from httpwwwmedinfocsucyaccy

Future work will investigate whether it is possible toidentify a group of patients at risk of atherosclerosis basedon their texture features extracted from the IL ML andthe IMC of high-resolution ultrasound images of the CCA

It may also be possible to identify and differentiate thoseindividuals into high and low risk groups according to theircardiovascular risk before the development of plaques Theproposed methodology may also be applied to a group ofpeople which already developed plaques in order to studythe contribution of the ML texture features to cardiovascularrisk Both groups of patients may benefit by prognosingandmanaging future cardiovascular events Another possiblefuture application of the proposed methodology is that it canbe used to investigate possible effects of statins or other drugsin texture feature changes of the ML of the CCA

The results will need to be validated on larger datasetsbefore they can be transitioned to clinical use Furthermorethe effect of despeckling on automated segmentation textureanalysis and classification of atherosclerotic plaques needs tobe further researched

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

References

[1] American Heart Association Heart Disease and Stroke Statis-tics Update Dallas Tex USA 2011

[2] P Pignoli E Tremoli and A Poli ldquoIntimal plus medial thick-ness of the arterial wall a direct measurement with ultrasoundimagingrdquo Circulation vol 74 no 6 pp 1399ndash1406 1986

[3] L E Chambless A R Folsom L X Clegg et al ldquoCarotidwall thickness is predictive of incident clinical stroke theAtherosclerosis Risk in Communities (ARIC) studyrdquo AmericanJournal of Epidemiology vol 151 no 5 pp 478ndash487 2000

[4] C P Loizou C S Pattichis A N Nicolaides andM PantziarisldquoManual and automated media and intima thickness measure-ments of the common carotid arteryrdquo IEEE Transactions onUltrasonics Ferroelectrics and Frequency Control vol 56 no 5pp 983ndash994 2009

[5] C P LoizouM PantziarisM S Pattichis E Kyriacou andC SPattichis ldquoUltrasound image texture analysis of the intima andmedia layers of the common carotid artery and its correlationwith age and genderrdquo Computerized Medical Imaging andGraphics vol 33 no 4 pp 317ndash324 2009

[6] R FWagner SW Smith JM Sandrik andH Lopez ldquoStatisticsof speckle in ultrasound B-scansrdquo IEEE Transactions on Sonicsand Ultrasonics vol 30 no 3 pp 156ndash163 1983

[7] Y Yu and S T Acton ldquoSpeckle reducing anisotropic diffusionrdquoIEEE Transactions on Image Processing vol 11 no 11 pp 1260ndash1270 2002

[8] C P Loizou V Murray M S Pattichis M Pantziaris andC S Pattichis ldquoMultiscale amplitude-modulation frequency-modulation (AM-FM) texture analysis of ultrasound imagesof the intima and media layers of the carotid arteryrdquo IEEETransactions on Information Technology in Biomedicine vol 15no 2 pp 178ndash188 2011

[9] M S Pattichis ldquoMultidimensional AM-FM models and meth-ods for biomedical image computingrdquo in Proceedings of the34th IEEE Annual International Conference of the Engineering inMedicine and Biology Society pp 5641ndash5644 September 2009


[10] M S Pattichis and A C Bovik ldquoAnalyzing image structurebymultidimensional frequencymodulationrdquo IEEETransactionson Pattern Analysis and Machine Intelligence vol 29 no 5 pp753ndash766 2007

[11] V Murray P Rodrıguez and M S Pattichis ldquoMultiscale AM-FM demodulation and image reconstruction methods withimproved accuracyrdquo IEEE Transactions on Image Processing vol19 no 5 pp 1138ndash1152 2010

[12] J P Havlicek D S Harding and A C Bovik ldquoThe multicom-ponent AM-FM image representationrdquo IEEE Transactions onImage Processing vol 5 no 6 pp 1094ndash1100 1996

[13] J P Havlicek D S Harding and A C Bovik ldquoMultidimen-sional quasi-eigenfunction approximations and multicompo-nent AM-FM modelsrdquo IEEE Transactions on Image Processingvol 9 no 2 pp 227ndash242 2000

[14] M S Pattichis C S Pattichis M Avraam A Bovik and K Kyr-iacou ldquoAM-FM texture segmentation in electron microscopicmuscle imagingrdquo IEEE Transactions onMedical Imaging vol 19no 12 pp 1253ndash1258 2000

[15] S Lee M S Pattichis and A C Bovik ldquoFoveated video qualityassessmentrdquo IEEE Transactions on Multimedia vol 4 no 1 pp129ndash132 2002

[16] S Lee M S Pattichis and A C Bovik ldquoFoveated videocompression with optimal rate controlrdquo IEEE Transactions onImage Processing vol 10 no 7 pp 977ndash992 2001

[17] C Agurto V Murray E Barriga et al ldquoMultiscale AM-FM methods for diabetic retinopathy lesion detectionrdquo IEEETransactions on Medical Imaging vol 29 no 2 pp 502ndash5122010

[18] N E Huang Z Shen S R Long et al ldquoThe empirical modedecomposition and the Hubert spectrum for nonlinear andnon-stationary time series analysisrdquo Proceedings of the RoyalSociety A Mathematical Physical and Engineering Sciences vol454 no 1971 pp 903ndash995 1998

[19] P Flandrin G Rilling and P Goncalves ldquoEmpirical modedecomposition as a filter bankrdquo IEEE Signal Processing Lettersvol 11 no 2 pp 112ndash114 2004

[20] G Rilling P Flandrin and P Goncalves ldquoOn empirical modedecomposition and its algorithmsrdquo in Proceedings of the IEEE-EURASIP Workshop on Nonlinear Signal and Image Processing(NSIP rsquo03) vol 3 pp 8ndash11 2003

[21] C P Loizou C S Pattichis C I Christodoulou R S HIstepanian M Pantziaris and A Nicolaides ldquoComparativeevaluation of despeckle filtering in ultrasound imaging of thecarotid arteryrdquo IEEE Transactions on Ultrasonics Ferroelectricsand Frequency Control vol 52 no 10 pp 1653ndash1669 2005

[22] C P Loizou CS Pattichis Despeckle Filtering Algorithms andSoftware for Ultrasound Imaging Synthesis Lectures on Algo-rithms and Software for Engineering Morgan amp Claypool SanRafael Calif USA 2008

[23] A Philips Medical System Company ldquoComparison of imageclarity SonoCT real-time compound imaging versus conven-tional 2D ultrasound imagingrdquo Tech Rep ATL Ultrasound2001

[24] A Nicolaides M Sabetai S K Kakkos et al ldquoThe Asymp-tomatic Carotid Stenosis and Risk of Stroke (ACSRS) studyAims and result of quality controlrdquo International Angiology vol22 no 3 pp 263ndash272 2003

[25] T Elatrozy A Nicolaides T Tegos A Z Zarka M Griffin andM Sabetai ldquoThe effect of B-mode ultrasonic image standard-isation on the echodensity of symptomatic and asymptomatic

carotid bifurcation plaquesrdquo International Angiology vol 17 no3 pp 179ndash186 1998

[26] C P Loizou C S Pattichis M Pantziaris T Tyllis and ANicolaides ldquoQuality evaluation of ultrasound imaging in thecarotid artery based on normalization and speckle reductionfilteringrdquo Medical and Biological Engineering and Computingvol 44 no 5 pp 414ndash426 2006

[27] J S Lee ldquoDigital image enhancement and noise filtering byusing local statisticsrdquo IEEE Transactions on Pattern Analysis andMachine Intelligence vol 2 no 2 pp 165ndash168 1980

[28] V S Frost J A Stiles K S Shanmugan and J C Holtzman ldquoAmodel for radar images and its application for adaptive digitalfiltering of multiplicative noiserdquo IEEE Transactions on PatternAnalysis andMachine Intelligence vol 4 no 2 pp 157ndash166 1982

[29] M Kuwahara K Hachimura S Eiho andM KinoshitaDigitalProcessing of Biomedical Images Edited by K Preston and MOnoe Plenum Publishing Corporation 1976

[30] A Nieminen P Heinonen and Y Neuvo ldquoA new class of detail-preserving filters for image processingrdquo IEEE Transactions onPattern Analysis and Machine Intelligence vol 9 no 1 pp 74ndash90 1987

[31] K Z Abd-Elmoniem A-B M Youssef and Y M KadahldquoReal-time speckle reduction and coherence enhancement inultrasound imaging via nonlinear anisotropic diffusionrdquo IEEETransactions on Biomedical Engineering vol 49 no 9 pp 997ndash1014 2002

[32] D J Williams and M Shah ldquoA Fast algorithm for active con-tours and curvature estimationrdquo CVGIP Image Understandingvol 55 no 1 pp 14ndash26 1992

[33] C P Loizou C S Pattichis M Pantziaris T Tyllis and ANicolaides ldquoSnakes based segmentation of the common carotidartery intima mediardquo Medical and Biological Engineering andComputing vol 45 no 1 pp 35ndash49 2007

[34] W J Conover Practical Nonparametric Statistics John Wiley ampSons New York NY USA 3rd edition 1999

[35] A Mittal R Soundararajan and A C Bovik ldquoMaking acompletely blind image quality analyserrdquo IEEE Signal ProcessingLetters vol 22 no 3 pp 209ndash212 2013

[36] A Schmidt-Trucksass D Grathwohl A Schmid et al ldquoStruc-tural functional and hemodynamic changes of the commoncarotid artery with age in male subjectsrdquo ArteriosclerosisThrombosis and Vascular Biology vol 19 no 4 pp 1091ndash10971999

[37] A N Nicolaides S K Kakkos M Griffin et al ldquoEffect ofimage normalization on carotid plaque classification and therisk of ipsilateral hemispheric ischemic events results fromthe Asymptomatic Carotid Stenosis and Risk of Stroke studyrdquoVascular vol 13 no 4 pp 211ndash221 2005

[38] AMojsilovic M Popovic N Amodaj R Basic andM OstojicldquoAutomatic segmentation of intravascular ultrasound images atexture-based approachrdquo Annals of Biomedical Engineering vol25 no 6 pp 1059ndash1071 1997

[39] A C Bovik ldquoOn detecting edges in speckle imageryrdquo IEEETransactions on Acoustics Speech and Signal Processing vol 36no 10 pp 1618ndash1627 1988

[40] C P Loizou S Petroudi C S Pattichis M Pantziaris and AN Nicolaides ldquoAn integrated system for the segmentation ofatherosclerotic carotid plaque in ultrasound videordquo IEEE Trans-actions onUltrasonics Ferroelectrics and FrequencyControl vol61 no 1 pp 86ndash101 2014


[41] C P Loizou C Theofanous M Pantziaris et al ldquoDespecklefiltering toolbox for medical ultrasound videordquo InternationalJournal of Monitoring and Surveillance Technologies Researchvol 4 no 1 pp 61ndash79 2013

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014


Active and Passive Electronic Components

Control Scienceand Engineering

Journal of



RotatingMachinery


Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design



Shock and Vibration


Civil EngineeringAdvances in

Acoustics and VibrationAdvances in



Electrical and Computer Engineering

Journal of

Advances inOptoElectronics


Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of


Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014


Chemical EngineeringInternational Journal of Antennas and

Propagation




Navigation and Observation



DistributedSensor Networks



IntimaMediaAdventitia

Intima-media thickness (IMT) Z5-Z6

Z5Z6Z7

15

16

17

(a)

(b)

Figure 1 Anatomical locations of the common carotid artery ultrasound image components at the far wall The IMT is defined as the layer(band) which is comprised by the bands Z5 and Z6 as demonstrated in (a) The intima-media-complex (IMC) in (b) has been extractedusing automated segmentation as described in [4 24] where the IMTaver = 079mm (between the bands Z5 and Z6 middle bar) IMTmax =08367mm (left bar) IMTmin = 06356mm (right bar) and IMTmedian = 075mm)

carried out successfully It was furthermore proposed that theIMT its thickness [2 4] and its textural characteristics [4 5]may be associated with the risk of developing stroke

Speckle a form of locally correlated multiplicative noisecorrupts medical ultrasound imaging making visual obser-vation difficult [6 7] Clearly though excessive despecklingmay result in the loss of image structureUnfortunately imagestructure is hard to assess and this is especially difficult fornoisy ultrasound images

We are interested in characterizing image structure usinga multiscale Amplitude-Modulation Frequency-Modulation(AM-FM) texture analysis system as described in [8] andby the independent visual assessment of clinical experts whoare asked to judge the quality of the despeckled imagesMultiscale AM-FM models represent nonstationary imagecontent using spatially-varying amplitude and phase com-ponents We use the term instantaneous amplitude (IA) todescribe spatially-varying amplitude components Similarlywe use the term of instantaneous frequency (IF) to describespatially-varying frequency content AM-FM componentsare estimated using a multiscale filterbank that is tuned todifferent frequency bands Thus we use the term multiscaleAM-FM analysis to summarize the analysis The promise ofAM-FM methods for texture analysis can be summarized in(see [9]) as follows (i) they provide physically meaningfultexture features (eg instantaneous frequency in cycles permm) over multiple scales at pixel level resolution (ii)textures can be reconstructed from AM-FM components sothat we can visualize content (iii) we can extract AM-FMdecompositions for different frequency coverage and (iv) wehave the recent development of robust methods for AM-FMdemodulation (see examples in [9ndash11])

Early work in AM-FM image representations has beenreported in [12 13] In [12] Havlicek et al discussed the useof multicomponent AM-FMmodels where Kalman filters areused to track changes in the instantaneous amplitude andinstantaneous frequency components In [13] Havlicek et alintroduced a complex-extension of the Hilbert- transform

for images and suggested the use of quasi-eigenfunctionapproximations (QEAs) for estimating the instantaneousfrequency from discrete signals An early application of AM-FMmethods in medical imaging appeared in [14] in electronmicroscopy In [15] the authors introduced a foveated videoquality assessment method based on the use of continuous-space AM-FM transforms In [16] the authors extended thisresearch to video compression based on local instantaneousfrequency content and the characteristics of the human visualsystem An effective model for texture analysis based onmultidimensional frequency modulation was introduced in[10] The multiscale AM-FM methods that are used in thispaper were first introduced in [9] and first applied in medicalimage analysis in [17]

For one-dimensional signals the empirical mode decom-position pioneered in [18] has been applied to severalapplications The empirical mode decomposition uses aspecial case of AM-FM functions intrinsic mode functionsas basis functions for decomposing signals These intrinsicmode functions decompose fractional Gaussian noise usinga dyadic filter bank as documented in [19] The popularimplementation of the empirical mode decomposition basedon [20] requires that the instantaneous amplitude for thefirst mode be estimated using the extrema of the input signaland then using interpolation to determine the envelopeThis approach is clearly sensitive to noise since the envelopeestimation is very sensitive to additive noise artefacts We arenot aware of any robust extensions that avoid the significantnoise artefacts or robust 2D extensions required for thecurrent application On the other hand the use of multiscaleAM-FM decomposition allows us to deal with noise throughthe use of suitable designed band-pass filters [9] For thecurrent application the effectiveness of this approach isdemonstrated in [8]

While AM-FM decompositions can effectively modelnonstationary texture it is also clear that we want to avoidmeasuring noise We want to investigate the use of despecklefiltering for reducing the levels of noise in the image Yet





119891 (119909 119910) =

119873

sum119899=1

119886119899(119909 119910) cos120593

119899(119909 119910) (1)




















119860 (119909 119910) = 158120587

75le 120601119909le

120587

45 120601

119910= minus120601119909


119860 (119909 119910) = 250120587

65le 120601119909le

120587

55 120601

119910= 120601119909


119860 (119909 119910) = 102120587

75le 120601119909le

120587

45 120601

119910= minus120601119909


119860 (119909 119910) = 250120587

65le 120601119909le

120587

55 120601

119910= 120601119909


119860 (119909 119910) = 182120587

75le 120601119909le

120587

45 120601

119910= minus120601119909


(2)


= 119891119894119895+ 119899119894119895119891119894119895 where 119892

119894119895and 119891








119891119894119895= 119892 + 119896

119894119895(119892119894119895minus 119892) (3)

where 119891119894119895






119896119894119895=

1 minus 11989221205902

1205902 (1 + 1205902119899) (4)



119894119895= (1205902 minus





119862 =1205902

119904

119892119904

(5)

where 1205902119904and 119892




(a) (b) (c) (d)

(e) (f) (g) (h)

(i)





119891119894119895= 119892119894119895+1

120578119904

div (119888srad (1003816100381610038161003816nabla119892

1003816100381610038161003816) nabla119892119894119895) (6)


1198882

srad (1003816100381610038161003816nabla119892

1003816100381610038161003816) =(12)

10038161003816100381610038161003816nabla119892119894119895

10038161003816100381610038161003816

2

minus (116) (nabla2119892119894119895)2

(119892119894119895+ (14) nabla2119892119894119895)

2 (7)



119889119892119894119895119905

119889119905= div [119863nabla119892] (8)




119863 = (1205961 1205962) (

1205821

0

0 1205822

)(1205961198791

1205961198792

) (9)

with

1205821=

120572(1 minus(1205831minus 1205832)2

1199042) if (120582

1minus 1205822)2le 1199042

0 otherwise(10)

1205822= 120572 (11)


1 1205822




1= 1205832= 120572 =


1minus 1205832)2gt 1199042 [31]








119889120601 (119909 119910)

119889119909cong1

119899arccos(

119891 (119909 + 119899 119910) + 119891 (119909 minus 119899 119910)

2119891 (119909 119910)) (12)















































Original



MIF 50ndash60 285 80 284 275 276 280 289 296 303MIF gt60 284 87 282 273 275 278 289 295 303

HIF lt50 574 28 574 545 546 556 578 599 641HIF gt60 566 16 564 541 545 557 574 585 597

DsFhybrimedian



MIF 50ndash60 284 822 283 273 274 279 287 296 301MIF gt60 283 835 281 274 275 276 290 296 302

HIF lt50 564 1776 568 539 541 550 570 586 599HIF gt60 556 126 553 540 543 545 566 575 580

DsFKuhawara



MIF 50ndash60 284 822 283 273 274 279 287 296 301MIF gt60 283 83 281 274 275 276 290 296 303

HIF lt50 564 178 568 539 541 550 570 586 599HIF gt60 556 125 553 540 543 546 566 575 580








Log LIA

Log MIA

Log HIA

LIF

MIF

HIF

FM

















References















































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation













119891 (119909 119910) =

119873

sum119899=1

119886119899(119909 119910) cos120593

119899(119909 119910) (1)




















119860 (119909 119910) = 158120587

75le 120601119909le

120587

45 120601

119910= minus120601119909


119860 (119909 119910) = 250120587

65le 120601119909le

120587

55 120601

119910= 120601119909


119860 (119909 119910) = 102120587

75le 120601119909le

120587

45 120601

119910= minus120601119909


119860 (119909 119910) = 250120587

65le 120601119909le

120587

55 120601

119910= 120601119909


119860 (119909 119910) = 182120587

75le 120601119909le

120587

45 120601

119910= minus120601119909


(2)


= 119891119894119895+ 119899119894119895119891119894119895 where 119892

119894119895and 119891








119891119894119895= 119892 + 119896

119894119895(119892119894119895minus 119892) (3)

where 119891119894119895






119896119894119895=

1 minus 11989221205902

1205902 (1 + 1205902119899) (4)



119894119895= (1205902 minus





119862 =1205902

119904

119892119904

(5)

where 1205902119904and 119892




(a) (b) (c) (d)

(e) (f) (g) (h)

(i)





119891119894119895= 119892119894119895+1

120578119904

div (119888srad (1003816100381610038161003816nabla119892

1003816100381610038161003816) nabla119892119894119895) (6)


1198882

srad (1003816100381610038161003816nabla119892

1003816100381610038161003816) =(12)

10038161003816100381610038161003816nabla119892119894119895

10038161003816100381610038161003816

2

minus (116) (nabla2119892119894119895)2

(119892119894119895+ (14) nabla2119892119894119895)

2 (7)



119889119892119894119895119905

119889119905= div [119863nabla119892] (8)




119863 = (1205961 1205962) (

1205821

0

0 1205822

)(1205961198791

1205961198792

) (9)

with

1205821=

120572(1 minus(1205831minus 1205832)2

1199042) if (120582

1minus 1205822)2le 1199042

0 otherwise(10)

1205822= 120572 (11)


1 1205822




1= 1205832= 120572 =


1minus 1205832)2gt 1199042 [31]








119889120601 (119909 119910)

119889119909cong1

119899arccos(

119891 (119909 + 119899 119910) + 119891 (119909 minus 119899 119910)

2119891 (119909 119910)) (12)















































Original



MIF 50ndash60 285 80 284 275 276 280 289 296 303MIF gt60 284 87 282 273 275 278 289 295 303

HIF lt50 574 28 574 545 546 556 578 599 641HIF gt60 566 16 564 541 545 557 574 585 597

DsFhybrimedian



MIF 50ndash60 284 822 283 273 274 279 287 296 301MIF gt60 283 835 281 274 275 276 290 296 302

HIF lt50 564 1776 568 539 541 550 570 586 599HIF gt60 556 126 553 540 543 545 566 575 580

DsFKuhawara



MIF 50ndash60 284 822 283 273 274 279 287 296 301MIF gt60 283 83 281 274 275 276 290 296 303

HIF lt50 564 178 568 539 541 550 570 586 599HIF gt60 556 125 553 540 543 546 566 575 580








Log LIA

Log MIA

Log HIA

LIF

MIF

HIF

FM

















References















































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation












119860 (119909 119910) = 158120587

75le 120601119909le

120587

45 120601

119910= minus120601119909


119860 (119909 119910) = 250120587

65le 120601119909le

120587

55 120601

119910= 120601119909


119860 (119909 119910) = 102120587

75le 120601119909le

120587

45 120601

119910= minus120601119909


119860 (119909 119910) = 250120587

65le 120601119909le

120587

55 120601

119910= 120601119909


119860 (119909 119910) = 182120587

75le 120601119909le

120587

45 120601

119910= minus120601119909


(2)


= 119891119894119895+ 119899119894119895119891119894119895 where 119892

119894119895and 119891








119891119894119895= 119892 + 119896

119894119895(119892119894119895minus 119892) (3)

where 119891119894119895






119896119894119895=

1 minus 11989221205902

1205902 (1 + 1205902119899) (4)



119894119895= (1205902 minus





119862 =1205902

119904

119892119904

(5)

where 1205902119904and 119892




(a) (b) (c) (d)

(e) (f) (g) (h)

(i)





119891119894119895= 119892119894119895+1

120578119904

div (119888srad (1003816100381610038161003816nabla119892

1003816100381610038161003816) nabla119892119894119895) (6)


1198882

srad (1003816100381610038161003816nabla119892

1003816100381610038161003816) =(12)

10038161003816100381610038161003816nabla119892119894119895

10038161003816100381610038161003816

2

minus (116) (nabla2119892119894119895)2

(119892119894119895+ (14) nabla2119892119894119895)

2 (7)



119889119892119894119895119905

119889119905= div [119863nabla119892] (8)




119863 = (1205961 1205962) (

1205821

0

0 1205822

)(1205961198791

1205961198792

) (9)

with

1205821=

120572(1 minus(1205831minus 1205832)2

1199042) if (120582

1minus 1205822)2le 1199042

0 otherwise(10)

1205822= 120572 (11)


1 1205822




1= 1205832= 120572 =


1minus 1205832)2gt 1199042 [31]








119889120601 (119909 119910)

119889119909cong1

119899arccos(

119891 (119909 + 119899 119910) + 119891 (119909 minus 119899 119910)

2119891 (119909 119910)) (12)















































Original



MIF 50ndash60 285 80 284 275 276 280 289 296 303MIF gt60 284 87 282 273 275 278 289 295 303

HIF lt50 574 28 574 545 546 556 578 599 641HIF gt60 566 16 564 541 545 557 574 585 597

DsFhybrimedian



MIF 50ndash60 284 822 283 273 274 279 287 296 301MIF gt60 283 835 281 274 275 276 290 296 302

HIF lt50 564 1776 568 539 541 550 570 586 599HIF gt60 556 126 553 540 543 545 566 575 580

DsFKuhawara



MIF 50ndash60 284 822 283 273 274 279 287 296 301MIF gt60 283 83 281 274 275 276 290 296 303

HIF lt50 564 178 568 539 541 550 570 586 599HIF gt60 556 125 553 540 543 546 566 575 580








Log LIA

Log MIA

Log HIA

LIF

MIF

HIF

FM

















References















































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










(a) (b) (c) (d)

(e) (f) (g) (h)

(i)





119891119894119895= 119892119894119895+1

120578119904

div (119888srad (1003816100381610038161003816nabla119892

1003816100381610038161003816) nabla119892119894119895) (6)


1198882

srad (1003816100381610038161003816nabla119892

1003816100381610038161003816) =(12)

10038161003816100381610038161003816nabla119892119894119895

10038161003816100381610038161003816

2

minus (116) (nabla2119892119894119895)2

(119892119894119895+ (14) nabla2119892119894119895)

2 (7)



119889119892119894119895119905

119889119905= div [119863nabla119892] (8)




119863 = (1205961 1205962) (

1205821

0

0 1205822

)(1205961198791

1205961198792

) (9)

with

1205821=

120572(1 minus(1205831minus 1205832)2

1199042) if (120582

1minus 1205822)2le 1199042

0 otherwise(10)

1205822= 120572 (11)


1 1205822




1= 1205832= 120572 =


1minus 1205832)2gt 1199042 [31]








119889120601 (119909 119910)

119889119909cong1

119899arccos(

119891 (119909 + 119899 119910) + 119891 (119909 minus 119899 119910)

2119891 (119909 119910)) (12)















































Original



MIF 50ndash60 285 80 284 275 276 280 289 296 303MIF gt60 284 87 282 273 275 278 289 295 303

HIF lt50 574 28 574 545 546 556 578 599 641HIF gt60 566 16 564 541 545 557 574 585 597

DsFhybrimedian



MIF 50ndash60 284 822 283 273 274 279 287 296 301MIF gt60 283 835 281 274 275 276 290 296 302

HIF lt50 564 1776 568 539 541 550 570 586 599HIF gt60 556 126 553 540 543 545 566 575 580

DsFKuhawara



MIF 50ndash60 284 822 283 273 274 279 287 296 301MIF gt60 283 83 281 274 275 276 290 296 303

HIF lt50 564 178 568 539 541 550 570 586 599HIF gt60 556 125 553 540 543 546 566 575 580








Log LIA

Log MIA

Log HIA

LIF

MIF

HIF

FM

















References















































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation











119863 = (1205961 1205962) (

1205821

0

0 1205822

)(1205961198791

1205961198792

) (9)

with

1205821=

120572(1 minus(1205831minus 1205832)2

1199042) if (120582

1minus 1205822)2le 1199042

0 otherwise(10)

1205822= 120572 (11)


1 1205822




1= 1205832= 120572 =


1minus 1205832)2gt 1199042 [31]








119889120601 (119909 119910)

119889119909cong1

119899arccos(

119891 (119909 + 119899 119910) + 119891 (119909 minus 119899 119910)

2119891 (119909 119910)) (12)















































Original



MIF 50ndash60 285 80 284 275 276 280 289 296 303MIF gt60 284 87 282 273 275 278 289 295 303

HIF lt50 574 28 574 545 546 556 578 599 641HIF gt60 566 16 564 541 545 557 574 585 597

DsFhybrimedian



MIF 50ndash60 284 822 283 273 274 279 287 296 301MIF gt60 283 835 281 274 275 276 290 296 302

HIF lt50 564 1776 568 539 541 550 570 586 599HIF gt60 556 126 553 540 543 545 566 575 580

DsFKuhawara



MIF 50ndash60 284 822 283 273 274 279 287 296 301MIF gt60 283 83 281 274 275 276 290 296 303

HIF lt50 564 178 568 539 541 550 570 586 599HIF gt60 556 125 553 540 543 546 566 575 580








Log LIA

Log MIA

Log HIA

LIF

MIF

HIF

FM

















References















































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation



















































Original



MIF 50ndash60 285 80 284 275 276 280 289 296 303MIF gt60 284 87 282 273 275 278 289 295 303

HIF lt50 574 28 574 545 546 556 578 599 641HIF gt60 566 16 564 541 545 557 574 585 597

DsFhybrimedian



MIF 50ndash60 284 822 283 273 274 279 287 296 301MIF gt60 283 835 281 274 275 276 290 296 302

HIF lt50 564 1776 568 539 541 550 570 586 599HIF gt60 556 126 553 540 543 545 566 575 580

DsFKuhawara



MIF 50ndash60 284 822 283 273 274 279 287 296 301MIF gt60 283 83 281 274 275 276 290 296 303

HIF lt50 564 178 568 539 541 550 570 586 599HIF gt60 556 125 553 540 543 546 566 575 580








Log LIA

Log MIA

Log HIA

LIF

MIF

HIF

FM

















References















































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation











Log LIA

Log MIA

Log HIA

LIF

MIF

HIF

FM

















References















































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation



















References















































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation














































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation









research article despeckle filtering for multiscale...

Documents