Research Article The Impact of Pixel Resolution ...downloads.hindawi.com/journals/ijo/2016/1370259.pdf · several studies have used the pixel resolution masa baseline resolution in

Research ArticleThe Impact of Pixel Resolution Integration ScalePreprocessing and Feature Normalization on TextureAnalysis for Mass Classification in Mammograms

Mohamed Abdel-Nasser1 Jaime Melendez2 Antonio Moreno1 and Domenec Puig1

1Departament drsquoEnginyeria Informatica i Matematiques Universitat Rovira i Virgili Avinguda Paisos Catalans 2643007 Tarragona Spain2Department of Radiology Radboud University Medical Center 6525 GA Nijmegen Netherlands

Correspondence should be addressed to Mohamed Abdel-Nasser egnasergmailcom

Received 6 November 2015 Revised 21 March 2016 Accepted 22 March 2016

Academic Editor Chenggen Quan

Copyright copy 2016 Mohamed Abdel-Nasser et alThis is an open access article distributed under the Creative CommonsAttributionLicense which permits unrestricted use distribution and reproduction in anymedium provided the originalwork is properly cited

Texture analysis methods are widely used to characterize breast masses in mammograms Texture gives information about thespatial arrangement of the intensities in the region of interestThis information has been used inmammogram analysis applicationssuch as mass detection mass classification and breast density estimation In this paper we study the effect of factors such aspixel resolution integration scale preprocessing and feature normalization on the performance of those texture methods for massclassification The classification performance was assessed considering linear and nonlinear support vector machine classifiers Tofind the best combination among the studied factors we used three approaches greedy sequential forward selection (SFS) andexhaustive search On the basis of our study we conclude that the factors studied affect the performance of texture methods so thebest combination of these factors should be determined to achieve the best performance with each texture method SFS can be anappropriate way to approach the factor combination problem because it is less computationally intensive than the other methods

1 Introduction

Breast cancerwas responsible for the largest number of cancerdeaths among the EU females in 2014 [1] Mammography isconsidered in general the most effective method for earlydetection of breast cancer and thus has been adopted forbreast cancer screening Computer-aided detection (CAD)systems are typically used to analyze mammograms inscreening While radiologists are generally pleased with theperformance of CAD for clustered microcalcification detec-tion they have little confidence in CAD for mass detectionThe most common complaint of radiologists is that CADsystems lead to a large number of false positives [2]

A breast cancer CAD system consists of threemain stagessegmentation of a region of interest (ROI) from the mam-mogram feature extraction from the ROI and classificationAlthough mammography is a highly sensitive method forearly detection of breast cancer low specificity has beenachieved in the classification of benign andmalignantmasses

Texture analysis methods constitute one of the options forimproving the specificity of classification algorithms appliedto mammography These methods may provide additionalinformation in distinguishing benign and malignant massesAlthough several feature extraction methods have beenproposed for analyzing mammograms improving the classi-fication performance remains a challenging problem

Texture analysis methods have been widely used toanalyze mammographic images because they produce infor-mation about the spatial arrangement of intensities in themammogram Texture is one of the major mammographiccharacteristics for mass classification For instance severalstudies have used texture analysis methods to distinguishbetween normal and abnormal tissue [3ndash8] or to discriminatebetween benign and malignant masses [9ndash11] Table 1 brieflysummarizes some of this previous work In addition otherstudies have used texture analysis methods to estimate breastdensity [12] or to segment masses from mammograms [13]

Hindawi Publishing CorporationInternational Journal of OpticsVolume 2016 Article ID 1370259 12 pageshttpdxdoiorg10115520161370259

2 International Journal of Optics

Table 1 Summary of texture analysis methods that have been used to analyze mammograms

Method Extracted features Utilized classifiers Purpose[6] Local binary pattern (LBP) Support vector machines (SVMs) Classification of ROIs into massnormal

[7] Histogram of oriented gradients(HOG) SVM Classification of ROIs into massnormal

[11] Haralickrsquos features (HAR) 119896-nearest neighbour (119896-NN) Microcalcification classification[8] Gabor filters (GF) Threshold-based approach Breast cancer detection

[29]Grey levels texture and featuresrelated to independent componentanalysis

Neural network (NN) Classifying ROIs into normalabnormalClassifying ROIs into benignmalignant

[30] A set of texture features SVM Mass detection[31] Ripleyrsquos 119870 function texture measures SVM Detection of breast masses

[9] Texture features derived fromconcurrence matrix NN Microcalcification classification

[32] A set of texture features 119896-NN SVM random forests logisticmodel trees and Naive Bayes Lesion classification

[10] HAR Bayesian classifierFisher linear discriminant

Study the effect of pixel resolution on theperformance of texture methods

[3]LBP robust LBP centre symmetricLBP fuzzy LBP local grey levelappearance LDN HOG HAR and GF

119896-NN linear SVM nonlinear SVMrandom forest and Fisher lineardiscriminant analysis (FLDA)

Finding the best combination among thetexture methods to classify ROIs intomassnormal

[33] Local ternary pattern and local phasequantization SVM Classifying tumors into benignmalignant

[34]Novel sets of texture descriptorsextracted from the cooccurrencematrix

SVM Six medical datasets were used forvalidation one of them for breast cancer

[35]Texture analysis techniques based onthe cooccurrence matrix andregion-based approaches

SVM 15 datasets were used for validation one ofthem for breast cancer

[36]HOG dense scale invariant featuretransform and local configurationpattern

SVM 119896-NN FLDA and decision tree Classifying ROIs into normalabnormalClassifying ROIs into benignmalignant

[37] Curvelet moments 119896-NN Classifying ROIs into normalabnormalClassifying ROIs into benignmalignant

CAD systems usually focus on a ROI to study breastmassesThe texture of this ROI describes the pattern of spatialvariation of gray levels in a neighbourhood that is smallcompared to the breast area but big enough to include themasses In other words texture must be analyzed in a regionand the size of this region should be tuned Thus we shouldanswer the question what is the optimal neighbourhood size(integration scale) for texture analysis In addition the sizeof a mammogram is usually in the range of thousands ofpixels Consequently several works have reduced the originalresolution of a mammogram to reduce the computationalcomplexity and the execution time of their algorithms [14] orto save resources (eg memory and storage space) Howeverimage downsampling may also affect the performance of thetexture analysis methods Therefore we should answer thequestion how far can we downsample the image while keepingthe performance of the texture methods

In breast cancer CAD systems several preprocessingoperations such as image filtering or enhancement are usuallyapplied tomammograms Pisano et al show that the contrast-limited adaptive histogram equalization (CLAHE) applied toa mammogram before it is displayed can make the indicative

structures of breast cancermore visible [15] Sharpening (SH)is used to improve the detection of clustered calcifications[16] The median filter (MF) is used to remove the noisefrom the mammograms [17] Preprocessing may affect theperformance of texture analysis methods because it effectivelychanges the gray levels of the images This effect should beassessed After extracting the texture features from a givenmammogram they are usually normalized before proceedingto the classification stage The utilized normalization methodmay also affect the final classification results

In this paper we study the effect of pixel resolutionintegration scale preprocessing and feature normalizationon the performance of texture analysis methods when usedto classify masses in mammograms For that purpose wehave chosen five widelyrecently used texture methodslocal binary pattern (LBP) local directional number (LDN)histogram of oriented gradients (HOG) Haralickrsquos features(HAR) and Gabor filters (GF) In order to evaluate theperformance of the aforementioned methods we extracteda set of regions of interest (ROIs) containing lesions fromthe mini-MIAS database [18] and we used each tex-ture analysis method to classify the ROIs into benign or

International Journal of Optics 3

malignant The performance of each texture method is eval-uated with five pixel resolutions (200120583m 400 120583m 600120583m800 120583m and 1000 120583m) six integration scales (25times25 32times3250 times 50 64 times 64 75 times 75 and 100 times 100 pixels) threepreprocessing steps (CLAHE MF and SH) and five featurenormalization methods In addition linear and nonlinearSVM classifiers are used

To the best of our knowledge only one previous studyhas conducted a similar evaluation Rangayyan et al studiedthe effect of pixel resolution on texture features of breastmasses in mammograms [10] However only pixel resolutionand Haralickrsquos features were considered In contrast thecurrent study takes into account a wider range of factorssuch as pixel resolution integration scale preprocessing andfeature normalization and it considers a larger number andmore powerful texture descriptors that have been success-fully applied in recent relevant work Moreover we includelinear and nonlinear SVMs thus both relatively simple andcomplex classification approaches can be assessed Lastly weanalyze the combination of the best options for those factorsusing three approaches greedy sequential forward selection(SFS) and exhaustive search (ExS)

The rest of this paper is organized as follows Section 2describes the database and the methods used in this studySection 3 shows our experimental results which are thendiscussed in Section 4 Finally Section 5 concludes our study

2 Materials and Methods

In this study we assess the performance of five texture analy-sis methods (LBP LDN HOG HAR and GF) while varyingthe pixel resolution integration scale image preprocessingalgorithm and data normalization method To that end weextracted a set of ROIs containing either benign or malignantmasses from themini-MIAS database Given a certain textureanalysis method a feature vector is extracted from each ROIto be fed into a linear support vector machine (LSVM) ora nonlinear support vector machine (NLSVM) The trainedmodels are used to determine if an unseen ROI contains abenign or a malignant mass

21 Materials The mini-MIAS database consisting of 322mediolateral oblique images of 161 cases is used in ourexperiments It was created from the original MIAS databaseby downsampling the images from 50 120583m to 200120583mper pixeland clippingpadding to a fixed size of 1024 times 1024 pixels Aground truth was prepared by experienced radiologists andconfirmed using a biopsy procedure The dataset is availableat httppeipaessexacukinfomiashtml In this study 109ROIs 60 containing a benign mass and 49 containing amalignant mass were used Figure 1 shows examples of theextracted ROIs Interested researchers can request the ROIsfrom the corresponding author of the paper

The authors of themini-MIASdatabase reported that theyreduced the pixel resolution of the original MIAS database(digitized at 50120583m) to 200120583m by popular request Moreoverseveral studies have used the pixel resolution 200 120583m as a

baseline resolution in their applications [14 19] We do thesame in this work

22 Texture Analysis Methods This section explains theutilized texture analysis methods including the parametersselected for each of them

221 Local Binary Pattern The LBP labels the pixels of animage by comparing a 3 times 3-pixel neighbourhood with thevalue of the central pixel [20] Pixels in this neighbourhoodwith a value greater than the central pixel are labelled as1 and the rest as 0 thus each pixel is represented by 8bits The size of the neighbourhood may vary on differentapplications (eg 3 times 3 and 5 times 5) A uniform LBP is anextension of the original LBP in which only patterns thatcontain at most two transitions from 0 to 1 (or vice versa)are considered In uniform LBP mapping there is a separateoutput label for each uniform pattern and all the nonuniformpatterns are assigned to a single label In this study a3 times 3 neighbourhood is used to generate the histogram ofuniform LBPs for each ROI The uniform mapping produces59 output labels (59 dimensions) for neighbourhoods of 8pixels The implementation of LBP descriptor is available athttpwwwcseoulufiCMVDownloadsLBPMatlab

222 Local Directional Number In the LDN [21] the edgeresponses are computed in eight different directions byconvoluting the Kirsch compass masks [22] with the ROIsThe locations of the top positive and negative edge responsesare used to generate a 6-bit code for each pixel Finally thehistogram of the LDN codes is calculated in the given ROI(64 dimensions) The implementation of LDN descriptor isavailable at httpsgitlabcommy-researchlocal-directional-number-patterngit

223 Histogram of Oriented Gradients In the HOG method[23] the occurrences of edge orientations in a ROI arecounted The image is divided into blocks (small groupsof cells) and then a weighted histogram is computed foreach of them The combination of the histograms of allblocks represents the final HOG descriptors In order toget the best performance of HOG its parameters havebeen empirically tuned In this study we used a 3 times 3cell size 8 times 8 cells for the block size and a 9-bit his-togram The implementation of HOG descriptor is availableat httpwwwvlfeatorgoverviewhoghtml

224 Haralickrsquos Features The HAR features are computedfrom the gray level cooccurrence matrix (GLCM) In theGLCM the distribution of cooccurring gray level values ata given offset (direction and distance) is computed [24] AGLCM is computed from each ROI and then 14 texturefeatures are calculated angular second moment contrastcorrelation variance inverse difference moment sum averagesum variance sum entropy entropy difference variance differ-ence entropy informationmeasure of correlation 1 informationmeasure of correlation 2 and maximal correlation coefficient[10] The mathematical expression of each feature can be


(a) (b)

Figure 1 ROIs extracted from the mini-MIAS breast cancer database A ROI containing (a) a benign mass and (b) a malignant mass

found in the relevant previous work [25 26] The implemen-tation of HAR descriptors is available at httpsgithubcomnutsiepullyspiffblobmastersrcharalickm

225 Gabor Filters A two-dimensional Gabor filter 119892(119909 119910)can be expressed as a sinusoidwith a particular frequency andorientation modulated by a Gaussian envelope

119892 (119909 119910) = exp(minus12)(11990921205902119909+119910

21205902119910)expminus1198952120587(1199060119909+V0119910) (1)

where (1199060 V0) is the centre of a sinusoidal function and 120590119909and 120590119910 are the standard deviations along two orthogonaldirections (which determine the width of the Gaussianenvelope along the 119909- and 119910-axes in the spatial domain)Given a ROI 119868(119909 119910) the filtered ROI 119891(119909 119910) is the resultof convoluting 119868(119909 119910) and 119892(119909 119910) Tuning GF to specificfrequencies and directions can lead them to detect both localorientation and frequency information from an image [27]In this study we used 4 scales and 6 orientations to obtainthese filtered ROIs This design produces 24 responses Foreach ROI the energies of the 24 responses are calculatedand then they are aggregated in order to form the featurevector The implementation of Gabor filters is available athttpsgithubcommhaghighatgabor

23 Preprocessing The performance of the texture analysismethods is evaluated with three preprocessing algorithmsCLAHE median filter (MF) and sharpening (SH)

(i) CLAHE it works on small regions of the input ROI(known as tiles)The contrast of each tile is enhancedconsequently the histogram of the output regionapproximatelymatches a predefined distribution [28]In this study the Rayleigh distribution is used [15]

(ii) MF each pixel in the filtered ROI contains themedian value of the 119898 times 119899 neighbourhood aroundthe corresponding pixel in the input ROI [17] In thisstudy a 3 times 3 neighbourhood is used

(iii) SH in order to sharpen a ROI it is first blurred edgesare detected in the blurred ROI and added to it toproduce a sharper image [16]

The preprocessing operations can be carried out usingthe following MATLAB functions CLAHE (adapthisteqm)median filter (medfilt2m) and sharpening (imsharpenm)Figure 2 shows examples for MF SH and CLAHE when theyare applied to benign and malignant masses

24 Feature NormalizationMethods Feature vectors are nor-malized in order to prevent attributes with higher numericranges from dominating those with lower numeric rangesGiven a feature vector 119909 = [1199091 1199092 1199093 119909119873] the normal-ized feature vector 119909new is calculated using five normalizationmethods as follows [38 39]

(i) The zeromean unit variance (119911119904) method 119909new = (119909minus120583)120590 where 120583 and 120590 are the mean and the variance of119909

(ii) The maximum-minimum (119898119899) method 119909new = (119909 minus119909min)(119909max minus 119909min) where 119909max and 119909min are themaximum and minimum of 119909

(iii) The ℓ1 method scales 119909 to unit length using the ℓ1-norm 119909new = 119909sum119873119899=1 |119909119899|

(iv) The ℓ2 method scales 119909 to unit length using the ℓ2-norm 119909new = 119909radicsum119873119899=1 |119909119899|2

(v) The 119899ℎ method scales 119909 to unit length as follows119909new = 119909sum119873119899=1 119909

The normalization methods can be easily implemented inMATLAB ℓ1- and ℓ2-norm can be carried out using theMATLAB function normm

25 Classification Given a labelled training set of the form(119909119894 119910119894) 119894 = 1 2 119896 where 119909119894 isin R119899 are the featurevalues 119910119894 isin 1 minus1 is the class of 119909119894 119899 is the number of


Benign ROI Malignant ROI

MFCLAHE SHMFCLAHE SH

Figure 2 Examples of ROI preprocessing

features and 119896 is the number of samples an SVM attempts todiscriminate between positive and negative classes by findinga hyperplane that separates them [40] The SVM classifiersolves the following optimization problem

1205962120596120585 + 119862119896

sum119894=1

120585119894

st 119910119894 (120596119879120601 (119909119894) + 119887) ge 1 minus 120585119894 120585119894 ge 0(2)

where the soft margin parameter 119862 controls the trade-offbetween the training error and the complexity of the SVMrsquosmodel in order to fit the training data and to avoid overfittingThe weight vector 120596 is normal to the separating hyperplaneThe parameter 120585 is used to give a degree of flexibility for thealgorithm when fitting the data and 119887 represents the bias

The SVM uses a kernel function to make the data linearlyseparable It projects the training data 119909119894 to a higher dimen-sional space as follows119870(119909119894 119909119895) = (120601119879(119909119894) sdot 120601(119909119895)) The SVMalgorithm attempts to find the hyperplane with maximummargin of separation between the classes in the new higherdimensional space In the case of a LSVMclassifier 120601 refers toa dot product In the case of a NLSVM the classifier functionis formed by nonlinearly projecting the training data in theinput space to a feature space of higher dimension by using akernel function In this study we use a radial basis function(RBF) as a mapping kernel which is defined as follows

119870(119909119894 119909119895) = exp (minus120574 10038171003817100381710038171003817119909119894 minus 119909119895100381710038171003817100381710038172

2) (3)

where 120574 = 121205902 119909119894 minus 11990911989522 is the squared Euclideandistance between the two feature vectors 119909119894 and 119909119895 and120590 is a free parameter In this work we use LIBSVM[41] to implement SVM classifiers LIBSVM is availableat httpswwwcsientuedutwsimcjlinlibsvm A grid searchalgorithm is performed to find the optimal parameter of the

RBF kernel 120574 and the regularization parameter 119862 For eachtraining set we estimated the parameters used by SVM in theclassification as done in [42]

26 Evaluation The performance of each texture analysismethod is measured in terms of the area under the curve(AUC) of the receiver operating characteristics (ROC) curve[43] The SVM classifier provides decision values related tothe membership of each class To generate a ROC curve wevary a threshold over the decision values We also use the 119896-fold cross validation technique to generate the training andtesting data In this procedure the data are partitioned into119896 folds thus 1119896 of ROIs are used for testing and the rest ofROIs are used for training In this study 119896 = 10 The meanAUC value is calculated over the cross validation process

3 Experiments

In this section we present the effect of pixel resolutionintegration scale preprocessing steps and normalizationmethods on the performance of the texture analysis methodswhen they are applied to benignmalignantmass classificationin mammograms Moreover we study the effect of differentcombinations of the aforementioned factors

31 Effect of Pixel Resolution and Integration Scale As wecommented in Section 21 the pixel resolution 200 120583m hasbeen widely used in several studies [14 19] So in thisexperiment we start with this pixel resolution and then themammograms are downsampled to generate different pixelresolutions The downsampling step includes antialiasingfiltering and a bicubic interpolation Five pixel resolutions aregenerated (200120583m 400 120583m 600120583m 800120583m and 1000 120583m)and then we use six integration scales (25 times 25 32 times 3250 times 50 64 times 64 75 times 75 and 100 times 100 pixels) to analyze


Table 2 Summary of the ANOVA results of pixel resolution andintegration scale with the LSVM (the value in each cell is a 119901 value)

Method Res IS ReslowastISLBP 00024 0001 0908LDN 01174 04035 08037HOG 03905 06515 04636HAR 07846 00962 02895GF 0083 08259 09864

Table 3 Summary of the ANOVA results of pixel resolution andintegration scalewith theNLSVM(the value in each cell is a119901 value)


the texture of each ROI In this experiment no preprocessingis applied and the standard 119911119904 normalization method is usedto normalize the extracted feature vectors The effect of pixelresolution and integration scale in the performance of LBPLDN HOG HAR and GF with the LSVM and the NLSVMis shown in Figure 3

As shown in Figure 3 each texture method achieves itsbest AUC value at a certain pixel resolution and integrationscale Among all texturemethods LBP achieves the best AUCvalue (078) at pixel resolution 800 120583m integration scale 75 times75

The analysis of variance (ANOVA) test [44] has beenused to examine the interaction between pixel resolutionsand integration scales The experimental design of ANOVAincludes two factors pixel resolution (Res) and integrationscale (IS) Res includes five levels (200 120583m 400120583m 600120583m800 120583m and 1000120583m) whereas IS includes six levels (25times2532 times 32 50 times 50 64 times 64 75 times 75 and 100 times 100 pixels)Each combination of the levels of Res and IS produces anAUC value (response)The confidence level is set to 005Theresults are shown in Tables 2 and 3

As shown in Table 2 with LBP and the LSVM the meanresponses for the levels of pixel resolution are significantlydifferent (119901 = 00024) Similarly the mean responses for thelevels of integration scale are significantly different In thecase of LDN HOGHAR andGF themean responses for thelevels of pixel resolution and integration scale are not signif-icantly different The 119901 values indicate that the interactionsbetween the levels of pixel resolution and integration scale(Res lowast IS) are not significant

As shown in Table 3 the mean responses for the levels ofpixel resolution are significantly different in the case of HOGwith the NLSVM In the case of LBP LDN HAR and GFthe mean responses for the levels of pixel resolution are notsignificantly different The mean responses for the levels ofintegration scale are significantly different in the case of LBPHOG and HAR With LBP and the NLSVM the interaction

between pixel resolution and integration scale (Res lowast IS) issignificant

32 Effect of Preprocessing In this experiment the integrationscale that obtained the highest AUC value with each textureanalysis method at the baseline pixel resolution of 200120583mand the standard 119911119904 normalization method are used Theeffect of no preprocessing (NP) CLAHE MF and SH onthe performance of each texture analysis method is shown inFigure 4 As can be seen each texture method produces thehighest AUC value with a certain preprocessing algorithmIn this experiment LBP achieves the highest AUC value withSH and theNLSVMwhile LDNandHAR achieve the highestAUCvaluewithNPand the LSVMHOGachieves the highestAUC value with CLAHE and the LSVM In turn GF achievesthe highest AUC value with CLAHE and the NLSVM

33 Effect of Feature Normalization Methods In this exper-iment we study the effect of five normalization methods(119911119904 119898119899 ℓ1 ℓ2 and 119899ℎ) on the performance of each textureanalysis method For each texture analysis method we usethe integration scale that produces the highest AUC valueat pixel resolution 200120583m No preprocessing method isused The effect of the normalization methods is shown inFigure 5 With the LSVM 119911119904 normalization has led LBPand LDN to AUC values better than other normalizationmethods while GF achieves its highest AUC value with ℓ1normalization and the NLSVM As shown in the figure eachtexture analysis method achieves its highest AUC value witha certain normalization method

34 Summary of the Results Thebest AUC values of each tex-ture analysismethod considering the experiments in Sections31 32 and 33 are summarized in Table 4 LBP produces thebest AUC value (078) at pixel resolution 800120583m integrationscale 75 times 75 no preprocessing 119911119904 normalization methodand the LSVM In turn HAR produces the lowest AUC value(061) LBP LDN HOG and HAR achieve their best valueswith the LSVM whereas GF achieves its best AUC value withthe NLSVM

35 Combining the Levels of All Factors To find the bestcombination among the levels of all factors we use threeapproaches greedy sequential forward selection (SFS) andexhaustive search (ExS) In the greedy approach we tryto combine the best options of the aforementioned factorsFor each texture analysis method we summarize the bestlevels of pixel resolution integration scale and normalizationmethods in Table 5

Table 6 shows that combining the best levels of pixelresolution integration scale preprocessing and feature nor-malization does not yield improvement on the AUC values ofthe texture analysis methods reported in Table 4 In fact LBPHOG and GF produced substantially lower AUC valuesTheLSVM yields higher AUC values than the NLSVM

Secondly we use a SFS approach to find the best com-bination It consists of two sequential steps finding the nor-malization method that improves the current performance


02

04

06

08

1

AUC

100 times 10075 times 7564 times 6450 times 5032 times 3225 times 25

1000 120583m800120583m600120583m

400120583m200120583m

(a)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(b)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(c)AU

C

02

04

06

08

1


1000 120583m800120583m600120583m

400120583m200120583m

(d)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(e)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(f)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(g)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(h)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(i)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(j)

Figure 3 The effect of pixel resolution and integration scale on the performance of the texture methods with the LSVM (a c e g i) theNLSVM (b d f h j) (a)-(b) LBP (c)-(d) LDN (e)-(f) HOG (g)-(h) HAR and (i)-(j) GF


LBP LDN HOG HAR GF0

02

04

06

08

1AU

C

NPCLAHE

MFSH

(a)

LBP LDN HOG HAR GF0

02

04

06

08

1

AUC

NPCLAHE

MFSH

(b)

Figure 4 The performance of the texture analysis methods with NP CLAHE MF and SH using (a) the LSVM and (b) the NLSVM

LBP LDN HOG HAR GF0

02

04

06

08

1

AUC

nh

l2

l1mn

zs

(a)

LBP LDN HOG HAR GF0

02

04

06

08

1

AUC

nh

l2

l1mn

zs

(b)

Figure 5 The performance of the texture analysis methods with different feature normalization methods using (a) the LSVM and (b) theNLSVM

the most and then finding the preprocessing method thatkeeps improving this performance For each texture methodin the first step we start with the best pixel resolutionand integration scale summarized in Table 5 Then withno preprocessing the extracted features are separately nor-malized by each normalization method Then the one thatimproves the performance in combination with the previoustwo factors is added In the second step we apply eachpreprocessing option to the ROIs (NP CLAHEMF and SH)Then we extract the texture features and normalize themusing the best normalizationmethod obtained in the previousstep Both LSVM and NLSVM are used to classify the ROIsTable 7 shows that the SFS does not improve the AUC valueof GF achieved in Table 4 LBP LDNHOG andHAR achieveAUC values close to the ones listed in Table 4With all texturemethods the SFS approach achieves AUC values better thanthe greedy approach

Lastly we use an ExS algorithm which is looking for thebest combination among five pixel resolutions six integra-tion scales and four preprocessing (NP CLAHE MF andSH) and five data normalization methods resulting in 600combinations In the previous experiments we found thatthe LSVM usually achieves the best results except with GFThe NLSVM has two parameters that need to be optimizedto achieve the best classification results Adding NLSVMrsquos

parameters optimization to the ExS substantially increases itscomplexity So we decided to only use the LSVM in this finaltest

As shown in Table 8 the ExS approach improves the AUCvalues of LDN HOG and HAR The GF achieves an AUCvalue lower than the one listed in Table 4 because the LSVMcan not perfectly separate the GF features

4 Discussion

Many factors affect the performance of texture analysismethods when applied to benignmalignant mass classifi-cation In this work we study the effect of factors such aspixel resolution integration scale preprocessing and featurenormalization We use the well-known mini-MIAS databasein this study We start with the original pixel resolutionof the mini-MIAS database (200 120583m) then we downsamplethe mammograms in order to generate the pixel resolutions400 120583m 600120583m 800120583m and 1000 120583m In addition sixintegration scales are used (25 times 25 32 times 32 50 times 5064 times 64 75 times 75 and 100 times 100 pixels) These integrationscales cover most of the sizes of the masses in the mini-MIAS database which range from a few pixels to tensof pixels (the mean diameter of the circle containing themasses is about 49 pixels) Several previous studies have


Table 4 Best AUC value for each texture analysis method and the configuration that yields it considering the experiments in Sections 3132 and 33

Method Best value Res (120583m) IS Classifier Preprocessing NorLBP 078 800 75 times 75 LSVM NP 119911119904LDN 068 600 64 times 64 LSVM NP 119911119904HOG 072 600 64 times 64 LSVM NP 119911119904HAR 061 200 32 times 32 LSVM NP 119899ℎGF 075 200 100 times 100 NLSVM CLAHE 119911119904

Table 5 The best option of pixel resolution integration scale preprocessing and normalization methods with each texture method

Method Res (120583m) IS (pixels) Preprocessing NormalizationLBP 800 75 times 75 SH 119911119904LDN 600 64 times 64 NP 119911119904HOG 600 64 times 64 CLAHE 119898119899HAR 200 32 times 32 NP 119899ℎGF 400 50 times 50 CLAHE ℓ1

Table 6 Results of the greedy approach (AUC)

Method LSVM NLSVMLBP 046 040LDN 068 052HOG 044 044HAR 061 048GF 058 054

Table 7 Results of the SFS approach

Method Best AUC Best parametersLBP 0780 119911119904 NP and LSVMLDN 0679 119911119904 NP and LSVMHOG 0716 119911119904 NP and LSVMHAR 0605 119899ℎ NP and LSVMGF 0720 119911119904 CLAHE and NLSVM

Table 8 Results of the ExS approach

Method Best AUC Best parametersLBP 078 800 75 times 75 NP and 119911119904LDN 070 600 75 times 75 MF and 119911119904HOG 0737 1000 50 times 50 SH and119898119899HAR 0666 800 32 times 32 CLAHE and 119899ℎGF 0691 600 32 times 32 NP and ℓ1

used one of these integration scales to analyze the textureof mammograms [3 6 7] Thus we hypothesize that theaforementioned integration scales are able to deal with all themasses appearing in the mini-MIAS database

The shape of breast masses is one of the powerful featuresthat can be used to discriminate between benign and malig-nant masses The boundaries of malignant masses usuallyhave irregular shapes while the boundaries of benign masseshave regular ones In the case of breast mass analysis pixel

resolution may be a critical factor because image downsam-pling may remove some fine detail from the image Howeveras our results indicate it would be possible to decrease theresolution far beyond 200 120583m and obtain good classificationresults A notable example is LBP which actually achieved itsbest performance at 800120583m A possible explanation is thatcore information such as that contained in the boundary ofmasses may still be preserved even after downsampling andbecome more useful for methods such as LBP that operateover higher order statistics of gray intensity values Obvi-ously when the resolution is far too low the classificationperformance degrades as the shape of the boundaries ofbenign and malignant masses will be very similar Anotherimportant factor is the integration scale as it should be bigenough to cover the masses and their boundaries and smallenough to exclude other tissues The effect of pixel resolutionand integration scale on the performance of texture methodsshould be jointly studied

As summarized in Table 5 each texture method achievesits highest AUC value at a certain pixel resolution and inte-gration scale A pixel resolution of 200120583m and an integrationscale of 32 times 32 pixels have led HAR to its highest AUCvalue In turn a pixel resolution of 800 120583mand an integrationscale of 75 times 75 pixels have led LBP to its best AUC valueThe integration scale and the pixel resolution interact witheach other in a certain way In the case of LBP LDN andHOG the texture features of each method are representedin a histogram This histogram includes the repetition ofthe patterns detected by each method at a certain pixelresolution and integration scale LBP features calculated atpixel resolution 200 120583m are different from those calculatedat pixel resolution 400 120583m LDN and HOG also producedifferent patterns at different pixel resolutions The localpatterns of LBP LDN andHOG are usually calculated withina certain integration scale Different integration scales willyield different histograms for the local patterns For instancethe histograms of LBP that are calculated with the integrationscales 75 times 75 and 100 times 100 are different


ANOVA results show that the mean AUC values of thepixel resolutions are significantly different in the case of LBPwith the LSVM In addition the mean AUC values of theintegration scales are significantly different with LBP HOGand HAR and the NLSVM The performance differenceswith respect to the pixel resolutions and the integrationscales are only significantly different with the LBP and theNLSVM (119901 = 00095) These results indicate that thechoice of the pixel resolution and the integration scale hasa direct implication on the performance of a texture-basedCAD system because our choice substantially affects theperformance of the utilized texture method

Image preprocessing also affects the performance of thetexture analysis methods HOG and GF achieve the highestAUC values with CLAHE while LDN and HAR performbetter with NP Indeed CLAHE MF and SH change theintensities of the mammograms in different ways As a resulteach texture analysis method will produce a different AUCvalue with each preprocessing technique In general thepreprocessing approach thatmakes the small-scale structuresin the ROIs more visible would give the texture methodsmore discriminative power For instance CLAHE leads GFto its best AUC value (075) There is also a coherentrelation between the principle of operation of some texturemethods and the utilized preprocessing For instance thebinary patterns of the LDN are calculated based on theedge responses of each pixel in the image MF removes theoutliers before calculating the edge responses Thus the edgeresponses will be properly calculated and the discriminativepower of LDN will improve

Prior tomass classification the calculated texture featuresshould be normalized to prevent attributes with highernumeric ranges from dominating those with lower numericranges As shown in our experiments each texture methodproduces its highest AUC value with a certain normalizationmethod This is because each normalization method pro-duces numerical values with different distributions Conse-quently the arrangement of the texture features in the featurespace with a certain normalization method is different thanwith other normalization methods Thus the normalizationtechnique changes the final values of the features computedby each texture method As shown in Table 5 LBP and LDNachieve the highest AUC values with 119911119904 normalization HOGwith119898119899 HAR with 119899ℎ and GF with ℓ1

In the classification stage we utilize two widely usedclassifiers in the field of mammogram analysis the LSVMand the NLSVM The first one tries to linearly separate thetexture features in the feature space while the second one usesa kernel function (RBF) to separate the features As shown inTable 4 the LSVM has led LBP LDN HOG and HAR to thehighest AUC values Conversely GF achieves the best AUCvalue with the NLSVM indicating that GF features are notlinearly separable

Table 4 shows a summary of the levels of pixel resolutionintegration scale preprocessing and normalization methodsthat have led each texture method to its best AUC valueconsidering the experiments in Sections 31 32 and 33HAR and GF achieve the best AUC values at pixel resolution

200120583m while LDN and HOG give their best results at pixelresolution 600120583m No method achieves its best AUC valuewith the integration scales 25 times 25 and 50 times 50 pixels

The greedy SFS and ExS approaches are used to find thebest combination among the levels of all factors Althoughthe greedy approach is the least complex approach it yieldedpoor AUC values In contrast the ExS achieved good resultsbut its computational complexity is the highest The SFSapproach provides a trade-off between the accuracy andthe computational complexity It is not as complex as theExS approach and it does not produce poor AUC valuesas the greedy approach In the case of LBP LDN HOGand HAR Table 7 shows that the SFS approach producesapproximately the same results as those obtainedwith the ExSapproach The GF achieved better AUC values with the SFSapproach because it used the NLSVM whereas using it withthe ExS approach presents some additional challenges in thecalculation of the optimal values of its internal parameters (120574and 119862)

Rangayyan et al extracted 111 ROIs from mammogramswhich were obtained from three different sources mammo-graphic image analysis society (MIAS) the teaching libraryof the Foothills Hospital in Calgary and a screening test (theAlberta program for the early detection of breast cancer)[10] Although using mammograms from different sourcesmay be helpful to assess the robustness of the studied texturemethods the three mammogram sets used by Rangayyanet al were digitized at different pixel resolutions Thus thecharacteristics of the textures extracted from the 111 ROIsmaybe different This changes the characteristics of the extractedfeatures so the effect of pixel resolution on the performanceof the texture methods may have not been properly studiedIn contrast in the current study the ROIs were extractedfrom a single source (the mini-MIAS database) Rangayyanet al extracted ROIs with different sizes (each ROI includeda mass) and they did not mention the effect of the integrationscale on the performance of the texture methods Converselythe current study has considered six integration scalesWith pixel resolution 800120583m integration scale 75 times 75 nopreprocessing 119911119904 normalization method and the LSVMthe LBP achieves the best AUC value (078) compared toother texture methods exceeding the best AUC value (075)achieved by Rangayyan et al [10]This is encouraging so ourfuture work will focus on improving the capabilities of anLBP-based approach by complementing it with the analysisof the fractal dimensions in multiple integration scales atdifferent pixel resolutions

As mentioned above the work of [10] has some simi-larities to our analysis however it obtained an AUC valueless than the one of our study in addition the authorsof [45] have studied the effect of ROI size and locationon texture methods when classifying the low-risk womenand the BRCA1BRCA2 gene-mutation carriers In turn ourstudy focuses on analyzing the impact of pixel resolutionintegration scale preprocessing and feature normalizationon texture methods when classifying breast tumors intobenign or malignant

In the current work we studied the impact of theabovementioned factors on the performance of texture


methods achieving the best AUC value with the LBP (078)However some methods in the literature achieved betterbenignmalignant breast cancer classification results such asthe ones of [33ndash35] For instance the authors of [35] achievedan AUC of 092 because they used ROIs of different dataset(DDSM) and extracted the GLCM features from subwindowsor regions (they added spatial information) We expect thatthe classification results of our study will be improved whenutilizing the region-based approach of [35] with each texturemethod One of our future research lines is to integrate theregion-based approach of [35] with our analysis

5 Conclusion

Texture analysis methods when applied to benignmalignantmass classification in mammograms are sensitive to thechanges of pixel resolution integration scale preprocessingand feature normalization The best combination of theaforementioned factors should be identified to achieve thebest discriminative power of each texture analysis methodWe expect that the assessment performed in this studywill help researchers to accomplish this task Due to itscomputational cost advantage sequential forward selectionwould be a suitable approach to determine a reasonable(possibly the best) factor configuration

Competing Interests

The authors declare that they have no competing interests

Acknowledgments

This work was partly supported by the Spanish Governmentthrough Project TIN2012-37171-C02-02

References

[1] M Malvezzi P Bertuccio F Levi C La Vecchia and E NegrildquoEuropean cancer mortality predictions for the year 2014rdquoAnnals of Oncology vol 25 no 8 pp 1650ndash1656 2014

[2] F J Gilbert S M Astley M G C Gillan et al ldquoSingle readingwith computer-aided detection for screening mammographyrdquoThe New England Journal of Medicine vol 359 no 16 pp 1675ndash1684 2008

[3] M Abdel-Nasser A Moreno and D Puig ldquoTowards costreduction of breast cancer diagnosis using mammographytexture analysisrdquo Journal of ExperimentalampTheoretical ArtificialIntelligence vol 28 no 1-2 pp 385ndash402 2016

[4] R Bellotti F De Carlo S Tangaro et al ldquoA completelyautomated CAD system for mass detection in a large mammo-graphic databaserdquoMedical Physics vol 33 no 8 pp 3066ndash30752006

[5] J Melendez C I Sanchez B Van Ginneken and N Karsse-meijer ldquoImproving mass candidate detection in mammogramsvia feature maxima propagation and local feature selectionrdquoMedical Physics vol 41 no 8 Article ID 081904 2014

[6] A Oliver X Llado J Freixenet and J Martı ldquoFalse positivereduction in mammographic mass detection using local binarypatternsrdquo in Medical Image Computing and Computer-Assisted

InterventionmdashMICCAI 2007 pp 286ndash293 Springer BerlinGermany 2007

[7] V Pomponiu H Hariharan B Zheng and D Gur ldquoImprovingbreast mass detection using histogram of oriented gradientsrdquoin Medical Imaging Computer-Aided Diagnosis vol 9035 ofProceedings of SPIE pp 1ndash6 International Society forOptics andPhotonics San Diego Calif USA March 2014

[8] Y Zheng ldquoBreast cancer detection with Gabor features fromdigital MammogramsrdquoAlgorithms vol 3 no 1 pp 44ndash62 2010

[9] H-P Chan B Sahiner N Patrick et al ldquoComputerized classi-fication of malignant and benign microcalcifications on mam-mograms texture analysis using an artificial neural networkrdquoPhysics inMedicine and Biology vol 42 no 3 pp 549ndash567 1997

[10] R M Rangayyan T M Nguyen F J Ayres and A K NandildquoEffect of pixel resolution on texture features of breast massesin mammogramsrdquo Journal of Digital Imaging vol 23 no 5 pp547ndash553 2010

[11] H Soltanian-Zadeh F Rafiee-Rad and D S Pourabdollah-Nejad ldquoComparison of multiwavelet wavelet Haralick andshape features for microcalcification classification in mammo-gramsrdquo Pattern Recognition vol 37 no 10 pp 1973ndash1986 2004

[12] A Oliver J Freixenet R Martı et al ldquoA novel breast tissuedensity classificationmethodologyrdquo IEEE Transactions on Infor-mation Technology in Biomedicine vol 12 no 1 pp 55ndash65 2008

[13] A Oliver J Freixenet J Martı et al ldquoA review of automaticmass detection and segmentation in mammographic imagesrdquoMedical Image Analysis vol 14 no 2 pp 87ndash110 2010

[14] M P Sampat A C Bovik G J Whitman andM KMarkey ldquoAmodel-based framework for the detection of spiculated massesonmammographyrdquoMedical Physics vol 35 no 5 pp 2110ndash21232008

[15] E D Pisano S Zong BMHemminger et al ldquoContrast limitedadaptive histogram equalization image processing to improvethe detection of simulated spiculations in densemammogramsrdquoJournal of Digital Imaging vol 11 no 4 pp 193ndash200 1998

[16] S Anand R S S Kumari S Jeeva and TThivya ldquoDirectionlettransform based sharpening and enhancement of mammo-graphic X-ray imagesrdquo Biomedical Signal Processing and Con-trol vol 8 no 4 pp 391ndash399 2013

[17] T S Subashini V Ramalingam and S Palanivel ldquoAutomatedassessment of breast tissue density in digital mammogramsrdquoComputer Vision and Image Understanding vol 114 no 1 pp33ndash43 2010

[18] J Suckling J Parker D Dance et al ldquoThe mammographicimage analysis society digital mammogram databaserdquo in Pro-ceedings of the 2nd InternationalWorkshop on Digital Mammog-raphy pp 375ndash378 York UK July 1994

[19] N Karssemeijer ldquoAutomated classification of parenchymal pat-terns inmammogramsrdquoPhysics inMedicine and Biology vol 43no 2 pp 365ndash378 1998

[20] T Ojala M Pietikainen and T Maenpaa ldquoMultiresolutiongray-scale and rotation invariant texture classificationwith localbinary patternsrdquo IEEE Transactions on Pattern Analysis andMachine Intelligence vol 24 no 7 pp 971ndash987 2002

[21] A R Rivera J R Castillo andO Chae ldquoLocal directional num-ber pattern for face analysis face and expression recognitionrdquoIEEE Transactions on Image Processing vol 22 no 5 pp 1740ndash1752 2013

[22] R A Kirsch ldquoComputer determination of the constituentstructure of biological imagesrdquo Computers and BiomedicalResearch vol 4 no 3 pp 315ndash328 1971


[23] N Dalal and B Triggs ldquoHistograms of oriented gradients forhuman detectionrdquo in Proceedings of the IEEE Computer SocietyConference on Computer Vision and Pattern Recognition (CVPRrsquo05) vol 1 pp 886ndash893 IEEE SanDiego Calif USA June 2005

[24] R M Haralick I Dinstein and K Shanmugam ldquoTexturalfeatures for image classificationrdquo IEEE Transactions on SystemsMan and Cybernetics vol 3 no 6 pp 610ndash621 1973

[25] W Gomez W C A Pereira and A F C Infantosi ldquoAnalysis ofco-occurrence texture statistics as a function of gray-level quan-tization for classifying breast ultrasoundrdquo IEEE Transactions onMedical Imaging vol 31 no 10 pp 1889ndash1899 2012

[26] R P Ramos M Z do Nascimento and D C Pereira ldquoTextureextraction an evaluation of ridgelet wavelet and co-occurrencebased methods applied to mammogramsrdquo Expert Systems withApplications vol 39 no 12 pp 11036ndash11047 2012

[27] J P Jones and L A Palmer ldquoAn evaluation of the two-dimensional Gabor filter model of simple receptive fields in catstriate cortexrdquo Journal of Neurophysiology vol 58 no 6 pp1233ndash1258 1987

[28] D T Puff E D Pisano K E Muller et al ldquoA method fordetermination of optimal image enhancement for the detectionof mammographic abnormalitiesrdquo Journal of Digital Imagingvol 7 no 4 pp 161ndash171 1994

[29] I Christoyianni A Koutras E Dermatas and G KokkinakisldquoComputer aided diagnosis of breast cancer in digitized mam-mogramsrdquoComputerizedMedical Imaging andGraphics vol 26no 5 pp 309ndash319 2002

[30] P Agrawal M Vatsa and R Singh ldquoSaliency based massdetection from screeningmammogramsrdquo Signal Processing vol99 pp 29ndash47 2014

[31] L de Oliveira Martins A C Silva A C De Paiva and MGattass ldquoDetection of breast masses in mammogram imagesusing growing neural gas algorithm and Ripleyrsquos K functionrdquoJournal of Signal Processing Systems vol 55 no 1ndash3 pp 77ndash902009

[32] D C Moura and M A G Lopez ldquoAn evaluation of imagedescriptors combined with clinical data for breast cancerdiagnosisrdquo International Journal of Computer Assisted Radiologyand Surgery vol 8 no 4 pp 561ndash574 2013

[33] L Nanni S Brahnam and A Lumini ldquoA very high performingsystem to discriminate tissues in mammograms as benign andmalignantrdquo Expert Systems with Applications vol 39 no 2 pp1968ndash1971 2012

[34] L Nanni S Brahnam S Ghidoni E Menegatti and T BarrierldquoDifferent approaches for extracting information from the co-occurrence matrixrdquo PLoS ONE vol 8 no 12 Article ID e835542013

[35] L Nanni S Brahnam S Ghidoni and E Menegatti ldquoRegion-based approaches and descriptors extracted from the co-occurrence matrixrdquo International Journal of Latest Research inScience and Technology vol 3 pp 192ndash200 2014

[36] S Ergin and O Kilic ldquoA new feature extraction frameworkbased on wavelets for breast cancer diagnosisrdquo Computers inBiology and Medicine vol 51 pp 171ndash182 2014

[37] S Dhahbi W Barhoumi and E Zagrouba ldquoBreast cancerdiagnosis in digitized mammograms using curvelet momentsrdquoComputers in Biology and Medicine vol 64 pp 79ndash90 2015

[38] S Aksoy and R M Haralick ldquoFeature normalization andlikelihood-based similarity measures for image retrievalrdquo Pat-tern Recognition Letters vol 22 no 5 pp 563ndash582 2001

[39] P Juszczak D Tax and R Duin ldquoFeature scaling in supportvector data descriptionrdquo in Proceedings of the 8th AnnualConference of the Advanced School for Computing and Imaging(ASCI rsquo02) pp 95ndash102 2002

[40] C Cortes and V Vapnik ldquoSupport-vector networksrdquo MachineLearning vol 20 no 3 pp 273ndash297 1995

[41] C-C Chang and C-J Lin ldquoLIBSVM a library for supportvector machinesrdquo ACM Transactions on Intelligent Systems andTechnology vol 2 article 27 2011

[42] G B Junior A C Cardoso de Paiva A Correa Silva and AC M de Oliveira ldquoClassification of breast tissues usingMoranrsquosindex and Gearyrsquos coefficient as texture signatures and SVMrdquoComputers in Biology and Medicine vol 39 no 12 pp 1063ndash1072 2009

[43] T Fawcett ldquoAn introduction to ROC analysisrdquo Pattern Recogni-tion Letters vol 27 no 8 pp 861ndash874 2006

[44] P Armitage G Berry and J N Matthews Statistical Methodsin Medical Research John Wiley amp Sons New York NY USA2002

[45] H Li M L Giger Z Huo et al ldquoComputerized analysisof mammographic parenchymal patterns for assessing breastcancer risk effect of ROI size and locationrdquoMedical Physics vol31 no 3 pp 549ndash555 2004

Submit your manuscripts athttpwwwhindawicom

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

High Energy PhysicsAdvances in

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014


FluidsJournal of

Atomic and Molecular Physics

Journal of



Advances in Condensed Matter Physics

OpticsInternational Journal of



AstronomyAdvances in

International Journal of


Superconductivity


Statistical MechanicsInternational Journal of


GravityJournal of


AstrophysicsJournal of


Physics Research International


Solid State PhysicsJournal of

Computational Methods in Physics

Journal of



Soft MatterJournal of

Hindawi Publishing Corporationhttpwwwhindawicom

AerodynamicsJournal of

Volume 2014


PhotonicsJournal of


Journal of

Biophysics


ThermodynamicsJournal of


Table 1 Summary of texture analysis methods that have been used to analyze mammograms

Method Extracted features Utilized classifiers Purpose[6] Local binary pattern (LBP) Support vector machines (SVMs) Classification of ROIs into massnormal

[7] Histogram of oriented gradients(HOG) SVM Classification of ROIs into massnormal

[11] Haralickrsquos features (HAR) 119896-nearest neighbour (119896-NN) Microcalcification classification[8] Gabor filters (GF) Threshold-based approach Breast cancer detection

[29]Grey levels texture and featuresrelated to independent componentanalysis

Neural network (NN) Classifying ROIs into normalabnormalClassifying ROIs into benignmalignant

[30] A set of texture features SVM Mass detection[31] Ripleyrsquos 119870 function texture measures SVM Detection of breast masses

[9] Texture features derived fromconcurrence matrix NN Microcalcification classification

[32] A set of texture features 119896-NN SVM random forests logisticmodel trees and Naive Bayes Lesion classification

[10] HAR Bayesian classifierFisher linear discriminant

Study the effect of pixel resolution on theperformance of texture methods

[3]LBP robust LBP centre symmetricLBP fuzzy LBP local grey levelappearance LDN HOG HAR and GF

119896-NN linear SVM nonlinear SVMrandom forest and Fisher lineardiscriminant analysis (FLDA)

Finding the best combination among thetexture methods to classify ROIs intomassnormal

[33] Local ternary pattern and local phasequantization SVM Classifying tumors into benignmalignant

[34]Novel sets of texture descriptorsextracted from the cooccurrencematrix

SVM Six medical datasets were used forvalidation one of them for breast cancer

[35]Texture analysis techniques based onthe cooccurrence matrix andregion-based approaches

SVM 15 datasets were used for validation one ofthem for breast cancer

[36]HOG dense scale invariant featuretransform and local configurationpattern

SVM 119896-NN FLDA and decision tree Classifying ROIs into normalabnormalClassifying ROIs into benignmalignant

[37] Curvelet moments 119896-NN Classifying ROIs into normalabnormalClassifying ROIs into benignmalignant

CAD systems usually focus on a ROI to study breastmassesThe texture of this ROI describes the pattern of spatialvariation of gray levels in a neighbourhood that is smallcompared to the breast area but big enough to include themasses In other words texture must be analyzed in a regionand the size of this region should be tuned Thus we shouldanswer the question what is the optimal neighbourhood size(integration scale) for texture analysis In addition the sizeof a mammogram is usually in the range of thousands ofpixels Consequently several works have reduced the originalresolution of a mammogram to reduce the computationalcomplexity and the execution time of their algorithms [14] orto save resources (eg memory and storage space) Howeverimage downsampling may also affect the performance of thetexture analysis methods Therefore we should answer thequestion how far can we downsample the image while keepingthe performance of the texture methods

In breast cancer CAD systems several preprocessingoperations such as image filtering or enhancement are usuallyapplied tomammograms Pisano et al show that the contrast-limited adaptive histogram equalization (CLAHE) applied toa mammogram before it is displayed can make the indicative

structures of breast cancermore visible [15] Sharpening (SH)is used to improve the detection of clustered calcifications[16] The median filter (MF) is used to remove the noisefrom the mammograms [17] Preprocessing may affect theperformance of texture analysis methods because it effectivelychanges the gray levels of the images This effect should beassessed After extracting the texture features from a givenmammogram they are usually normalized before proceedingto the classification stage The utilized normalization methodmay also affect the final classification results

In this paper we study the effect of pixel resolutionintegration scale preprocessing and feature normalizationon the performance of texture analysis methods when usedto classify masses in mammograms For that purpose wehave chosen five widelyrecently used texture methodslocal binary pattern (LBP) local directional number (LDN)histogram of oriented gradients (HOG) Haralickrsquos features(HAR) and Gabor filters (GF) In order to evaluate theperformance of the aforementioned methods we extracteda set of regions of interest (ROIs) containing lesions fromthe mini-MIAS database [18] and we used each tex-ture analysis method to classify the ROIs into benign or
















(a) (b)




119892 (119909 119910) = exp(minus12)(11990921205902119909+119910

21205902119910)expminus1198952120587(1199060119909+V0119910) (1)




















1205962120596120585 + 119862119896

sum119894=1

120585119894

st 119910119894 (120596119879120601 (119909119894) + 119887) ge 1 minus 120585119894 120585119894 ge 0(2)



119870(119909119894 119909119895) = exp (minus120574 10038171003817100381710038171003817119909119894 minus 119909119895100381710038171003817100381710038172

2) (3)




3 Experiments





















02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(a)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(b)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(c)AU

C

02

04

06

08

1


1000 120583m800120583m600120583m

400120583m200120583m

(d)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(e)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(f)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(g)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(h)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(i)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(j)



LBP LDN HOG HAR GF0

02

04

06

08

1AU

C

NPCLAHE

MFSH

(a)

LBP LDN HOG HAR GF0

02

04

06

08

1

AUC

NPCLAHE

MFSH

(b)


LBP LDN HOG HAR GF0

02

04

06

08

1

AUC

nh

l2

l1mn

zs

(a)

LBP LDN HOG HAR GF0

02

04

06

08

1

AUC

nh

l2

l1mn

zs

(b)






4 Discussion






























5 Conclusion


Competing Interests


Acknowledgments


References





















































FluidsJournal of


Journal of










Superconductivity




GravityJournal of








Journal of






Volume 2014


PhotonicsJournal of


Journal of

Biophysics


















(a) (b)




119892 (119909 119910) = exp(minus12)(11990921205902119909+119910

21205902119910)expminus1198952120587(1199060119909+V0119910) (1)




















1205962120596120585 + 119862119896

sum119894=1

120585119894

st 119910119894 (120596119879120601 (119909119894) + 119887) ge 1 minus 120585119894 120585119894 ge 0(2)



119870(119909119894 119909119895) = exp (minus120574 10038171003817100381710038171003817119909119894 minus 119909119895100381710038171003817100381710038172

2) (3)




3 Experiments





















02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(a)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(b)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(c)AU

C

02

04

06

08

1


1000 120583m800120583m600120583m

400120583m200120583m

(d)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(e)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(f)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(g)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(h)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(i)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(j)



LBP LDN HOG HAR GF0

02

04

06

08

1AU

C

NPCLAHE

MFSH

(a)

LBP LDN HOG HAR GF0

02

04

06

08

1

AUC

NPCLAHE

MFSH

(b)


LBP LDN HOG HAR GF0

02

04

06

08

1

AUC

nh

l2

l1mn

zs

(a)

LBP LDN HOG HAR GF0

02

04

06

08

1

AUC

nh

l2

l1mn

zs

(b)






4 Discussion






























5 Conclusion


Competing Interests


Acknowledgments


References





















































FluidsJournal of


Journal of










Superconductivity




GravityJournal of








Journal of






Volume 2014


PhotonicsJournal of


Journal of

Biophysics




(a) (b)




119892 (119909 119910) = exp(minus12)(11990921205902119909+119910

21205902119910)expminus1198952120587(1199060119909+V0119910) (1)




















1205962120596120585 + 119862119896

sum119894=1

120585119894

st 119910119894 (120596119879120601 (119909119894) + 119887) ge 1 minus 120585119894 120585119894 ge 0(2)



119870(119909119894 119909119895) = exp (minus120574 10038171003817100381710038171003817119909119894 minus 119909119895100381710038171003817100381710038172

2) (3)




3 Experiments





















02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(a)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(b)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(c)AU

C

02

04

06

08

1


1000 120583m800120583m600120583m

400120583m200120583m

(d)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(e)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(f)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(g)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(h)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(i)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(j)



LBP LDN HOG HAR GF0

02

04

06

08

1AU

C

NPCLAHE

MFSH

(a)

LBP LDN HOG HAR GF0

02

04

06

08

1

AUC

NPCLAHE

MFSH

(b)


LBP LDN HOG HAR GF0

02

04

06

08

1

AUC

nh

l2

l1mn

zs

(a)

LBP LDN HOG HAR GF0

02

04

06

08

1

AUC

nh

l2

l1mn

zs

(b)






4 Discussion






























5 Conclusion


Competing Interests


Acknowledgments


References





















































FluidsJournal of


Journal of










Superconductivity




GravityJournal of








Journal of






Volume 2014


PhotonicsJournal of


Journal of

Biophysics








1205962120596120585 + 119862119896

sum119894=1

120585119894

st 119910119894 (120596119879120601 (119909119894) + 119887) ge 1 minus 120585119894 120585119894 ge 0(2)



119870(119909119894 119909119895) = exp (minus120574 10038171003817100381710038171003817119909119894 minus 119909119895100381710038171003817100381710038172

2) (3)




3 Experiments





















02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(a)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(b)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(c)AU

C

02

04

06

08

1


1000 120583m800120583m600120583m

400120583m200120583m

(d)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(e)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(f)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(g)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(h)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(i)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(j)



LBP LDN HOG HAR GF0

02

04

06

08

1AU

C

NPCLAHE

MFSH

(a)

LBP LDN HOG HAR GF0

02

04

06

08

1

AUC

NPCLAHE

MFSH

(b)


LBP LDN HOG HAR GF0

02

04

06

08

1

AUC

nh

l2

l1mn

zs

(a)

LBP LDN HOG HAR GF0

02

04

06

08

1

AUC

nh

l2

l1mn

zs

(b)






4 Discussion






























5 Conclusion


Competing Interests


Acknowledgments


References





















































FluidsJournal of


Journal of










Superconductivity




GravityJournal of








Journal of






Volume 2014


PhotonicsJournal of


Journal of

Biophysics





















02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(a)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(b)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(c)AU

C

02

04

06

08

1


1000 120583m800120583m600120583m

400120583m200120583m

(d)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(e)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(f)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(g)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(h)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(i)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(j)



LBP LDN HOG HAR GF0

02

04

06

08

1AU

C

NPCLAHE

MFSH

(a)

LBP LDN HOG HAR GF0

02

04

06

08

1

AUC

NPCLAHE

MFSH

(b)


LBP LDN HOG HAR GF0

02

04

06

08

1

AUC

nh

l2

l1mn

zs

(a)

LBP LDN HOG HAR GF0

02

04

06

08

1

AUC

nh

l2

l1mn

zs

(b)






4 Discussion






























5 Conclusion


Competing Interests


Acknowledgments


References





















































FluidsJournal of


Journal of










Superconductivity




GravityJournal of








Journal of






Volume 2014


PhotonicsJournal of


Journal of

Biophysics




02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(a)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(b)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(c)AU

C

02

04

06

08

1


1000 120583m800120583m600120583m

400120583m200120583m

(d)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(e)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(f)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(g)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(h)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(i)

02

04

06

08

1

AUC


1000 120583m800120583m600120583m

400120583m200120583m

(j)



LBP LDN HOG HAR GF0

02

04

06

08

1AU

C

NPCLAHE

MFSH

(a)

LBP LDN HOG HAR GF0

02

04

06

08

1

AUC

NPCLAHE

MFSH

(b)


LBP LDN HOG HAR GF0

02

04

06

08

1

AUC

nh

l2

l1mn

zs

(a)

LBP LDN HOG HAR GF0

02

04

06

08

1

AUC

nh

l2

l1mn

zs

(b)






4 Discussion






























5 Conclusion


Competing Interests


Acknowledgments


References





















































FluidsJournal of


Journal of










Superconductivity




GravityJournal of








Journal of






Volume 2014


PhotonicsJournal of


Journal of

Biophysics




LBP LDN HOG HAR GF0

02

04

06

08

1AU

C

NPCLAHE

MFSH

(a)

LBP LDN HOG HAR GF0

02

04

06

08

1

AUC

NPCLAHE

MFSH

(b)


LBP LDN HOG HAR GF0

02

04

06

08

1

AUC

nh

l2

l1mn

zs

(a)

LBP LDN HOG HAR GF0

02

04

06

08

1

AUC

nh

l2

l1mn

zs

(b)






4 Discussion






























5 Conclusion


Competing Interests


Acknowledgments


References





















































FluidsJournal of


Journal of










Superconductivity




GravityJournal of








Journal of






Volume 2014


PhotonicsJournal of


Journal of

Biophysics































5 Conclusion


Competing Interests


Acknowledgments


References





















































FluidsJournal of


Journal of










Superconductivity




GravityJournal of








Journal of






Volume 2014


PhotonicsJournal of


Journal of

Biophysics
















5 Conclusion


Competing Interests


Acknowledgments


References





















































FluidsJournal of


Journal of










Superconductivity




GravityJournal of








Journal of






Volume 2014


PhotonicsJournal of


Journal of

Biophysics





5 Conclusion


Competing Interests


Acknowledgments


References





















































FluidsJournal of


Journal of










Superconductivity




GravityJournal of








Journal of






Volume 2014


PhotonicsJournal of


Journal of

Biophysics
































FluidsJournal of


Journal of










Superconductivity




GravityJournal of








Journal of






Volume 2014


PhotonicsJournal of


Journal of

Biophysics








FluidsJournal of


Journal of










Superconductivity




GravityJournal of








Journal of






Volume 2014


PhotonicsJournal of


Journal of

Biophysics



Documents

Research Article The Impact of Pixel Resolution ...downloads.hindawi.com/journals/ijo/2016/1370259.pdf · several studies have used the pixel resolution masa baseline resolution in