Research ArticleLocal Binary Patterns Descriptor Based on Sparse CurveletCoefficients for False-Positive Reduction in Mammograms

Meenakshi M Pawar 1 Sanjay N Talbar2 and Akshay Dudhane2

1Department of Electronics and Telecommunication SVERIrsquos College of Engineering Pandharpur Solapur Maharashtra India2Department of Electronics and Telecommunication Engg SGGSIE amp T Nanded Maharashtra India

Correspondence should be addressed to Meenakshi M Pawar mmpawarcoesveriacin

Received 6 April 2018 Revised 18 June 2018 Accepted 8 August 2018 Published 25 September 2018

Academic Editor Santosh K Vipparthi

Copyright copy 2018 Meenakshi M Pawar et al is is an open access article distributed under the Creative Commons AttributionLicense which permits unrestricted use distribution and reproduction in any medium provided the original work isproperly cited

Breast Cancer is the most prevalent cancer among women across the globe Automatic detection of breast cancer using ComputerAided Diagnosis (CAD) system suffers from false positives (FPs) us reduction of FP is one of the challenging tasks to improvethe performance of the diagnosis systems In the present work new FP reduction technique has been proposed for breast cancerdiagnosis It is based on appropriate integration of preprocessing Self-organizing map (SOM) clustering region of interest (ROI)extraction and FP reduction In preprocessing contrast enhancement of mammograms has been achieved using Local EntropyMaximization algorithme unsupervised SOM clusters an image into number of segments to identify the cancerous region andextracts tumor regions (ie ROIs) However it also detects some FPs which affects the efficiency of the algorithm erefore toreduce the FPs the output of the SOM is given to the FP reduction step which is aimed to classify the extracted ROIs into normaland abnormal class FP reduction consists of feature mining from the ROIs using proposed local sparse curvelet coefficientsfollowed by classification using artificial neural network (ANN)e performance of proposed algorithm has been validated usingthe local datasets as TMCH (Tata Memorial Cancer Hospital) and publicly available MIAS (Suckling et al 1994) and DDSM(Heath et al 2000) databasee proposed technique results in reduction of FPs from 085 to 002 FPimage for MIAS 481 to 016FPimage for DDSM and 232 to 005 FPimage for TMCH reflecting huge improvement in classification of mammograms

1 Introduction

Breast cancer is the most common cancer disease amongwomen across worldwide It is the leading cause of deaths forwomen suffering from cancer disease in India It is estimatedthat breast cancer cases in India would reach to as high as1797900 by 2020 [1] Rising rate of incidences can causehigh mortality is is due to lack of awareness about breastscreening late reporting and insufficient medical access [2]is fact brings a concern and necessity that screening forbreast cancer is prudent in its early stage to confirm longersurvival Among all techniques namely mammographytomosynthesis ultrasonography computed tomographyand magnetic resonance mammography is the most reliableand accepted modality by radiologist for preliminary ex-amination of breast cancer due to cost benefits and acces-sibility [3ndash5] e diagnosis of breast cancer using

mammogram by radiologist varies from expert to expert assymptoms are misinterpreted or overlooked due to thetedious task of screening mammograms Study reveals that10 to 30 of the visible cancers on mammograms areoverlooked and only 20 to 30 of biopsies are positive[6ndash8] Biopsies are traumatic in nature and costly thereforecomputer aided detection and diagnosis (CAD) systemscombined with expert radiologistsrsquo experience would pro-vide more comprehensive diagnosis [9] Detailed surveyabout the research in the design of CAD systems has beengiven in next section

2 Literature Survey

e design and development of CAD system is an importantprogressive area of research for contrast enhancement forbetter visualization and clarification [10ndash12] pectoral

HindawiJournal of Healthcare EngineeringVolume 2018 Article ID 5940436 16 pageshttpsdoiorg10115520185940436

muscle removal segmentation for better delineation of re-gion of interest (ROI) extraction of features and classifi-cation [13 14] e segmentation method is classified asregion based contour-based and clustering method [15]e region and contour-based methods are popularly usedby many researchers Gorgel et al [16] developed Local SeedRegion Growing-Spherical Wavelet Transform (LSRGndashSWT) algorithm using local dataset and MIAS [17] withclassification accuracy of 94 and 9167 respectivelyPereira et al [18] presented segmentation and detection ofmasses in mammogram using wavelet transform and geneticalgorithm that provides FP rate of 135 FPimage andsensitivity of 95 using DDSM [19] Rouhi et al [20] studiedsegmentation using region growing Cellular Neural Net-work (CNN) and ANN e result of classification variedfrom 80 to 96 which is the main weakness of their studyBerber et al [21] proposed Breast Mass Contour Segmen-tation (BMCS) approach and showed 6 FPR for local datasetHybrid level set segmentation method [22] based oncombination of region growing and level set was used tosegment tumor e results showed that the sensitivityvaried from 78 to 100 due to the presence of artifact in theMIAS database e difficulties in region and contour-basedsegmentation methods are the appropriate initialization ofseed point and contour position

Several researchers have implemented clustering methodlike K-means and Fuzzy C-means (FCM) for breast abnor-mality segmentation [3 23] However they have limitationsin terms of learning abilities Learning-based techniques suchas Self-organizing map (SOM) [24] have been successfullyused inmedical image segmentation [25]e success of SOMin medical image segmentation has inspired the researcher tochoose it for mammogram segmentation Many of the timesthe tumor-segmented regions are not the abnormal tissues(cancerous region) and they are known as false positives(FPs)is FP consumesmuch time of radiologists and resultsinto unnecessary biopsies us reducing the FPs is an openresearch problem and various researchers have proposed FPreduction algorithms to improve the specificity of the CADsystems [5 9 23 26ndash31] Usually FP reduction algorithm ispostprocessing step of a CAD system with two stages namelyFeature extraction and Classification Various methods havebeen developed for feature extraction based on wavelets[8 18 32] curvelet [33 34] Gabor [35 36] morphologicaldescriptors [20] textural analysis [26 27 30 32] histogram[4 5 7 29 37ndash40] etce segmentation error can reduce theperformances of morphological descriptor When Gray LevelCo-occurrence Matrix (GLCM) from normal and abnormalregion in dense mammogram is same texture descriptoroverlaps that leads to more number of FPs [37] Ojala et alproposed local binary patterns (LBPs) [41] for textural featureextraction which works well in feature extraction ascompared to morphological descriptor and GLCM-basedtextural descriptor LBP descriptor can be considered aslocal microstructures namely edges flat areas spots etcVariants of LBP have been proposed by various researchersto achieve rotation and intensity invariant features AlsoLBP is computationally efficient and extracts robust fea-tures therefore LBP descriptors have been widely applied

in FP reduction and classification methods for mammo-gram images [29 37 39 40] However LBP descriptor doesnot provide the directional information of local micro-pattern erefore transform technique such as curveletcombined with LBP was used to extract features Variouscurvelet-based approaches have been proposed in the lit-erature [8 33 34 42] which conclude that curvelet out-performs as compared to wavelet transform

In this work novel method of extracting sparse curveletsubband coefficients by incorporating the knowledge ofirregular shape of masses as they appear in sparse matrix andcalculating LBP features has been presented erefore thispaper presents scheme as follows

(1) Preprocessing of mammogram image for contrastenhancement using local entropy maximization-based image fusion algorithm and removal ofbackground noise

(2) Cluster-based segmentation of mammograms usingSOM and extract tumor regions ie ROI)

(3) FP reduction extraction of sparse curvelet subbandcoefficients and computation of LBP descriptor toclassify true positives and false positives to improveperformance of CAD system using MIAS [17]DDSM [19] and Tata Memorial Cancer Hospital(TMCH) datasets

e organization of paper is as follows Sections 1 and 2illustrate the introduction and literature review on auto-matic segmentation and extraction of abnormal masses(ie tumor region) as well as FP reduction methods Section3 presents the proposed methodology for SOM based seg-mentation of mammograms followed by novel false positivereduction in detail Section 4 depicts the experimental resultsand discussions on three benchmark datasets FinallySection 5 concludes the proposed approach for accurateextraction of abnormal masses (ie tumor region) by ex-cluding the FPs

3 Methodology

e block schematic of proposed integrated method forautomatic detection of breast cancer using sparse curveletcoefficient-based LBP descriptor has been shown in Figure 1

31 Preprocessing e mammogram images are low-dosex-ray images so they have poor contrast and suffer fromnoises e preprocessed mammogram image as shown inFigures 2(a)ndash2(d) represents preprocessing of mammogramand Figures 2(e)ndash2(g) represents SOM clustering and ROIextraction

311 Local Entropy Maximization-Based Image FusionContrast Enhancement e contrast enhancement of themammogram is performed using local entropy maximiza-tion [12] for better segmentation Here original image isgiven to the contrast limited adaptive histogram equalization(CLAHE) algorithm to get the second input to our image

2 Journal of Healthcare Engineering

fusion algorithm Further original image along with theCLAHE has been given to the image fusion algorithmProcedure of the image fusion has been given in Algo-rithm 1 We have used local entropy as a fusion rule given bythe following equation

ENT minussum255

k0p(k)log(p(k)) (1)

where ENT is the local entropy and p_org(k) andp_CLAHE(k) are the probability of kth pixel from 5times 5sliding window [12] Here both high frequency componentsfrom original mammogram and CLAHE mammogram havebeen fused using maximum entropy criteria Figure 3(b)presents contrast-enhanced mammogram using local en-tropy maximization-based image fusion

312 Pectoral Muscle Removal Pectoral muscle suppressionhas been performed by dening rectangle as suggested in[14] (Figure 3(c)) It illustrates the rectangle (ABDC) and

xes the points G and has intensity variation and joins themfor pectoral muscle suppression Figure 3(d) illustratespectoral muscle removed image to avoid discrepancies in thealgorithm because of similar intensities present betweenpectoral muscle and masses

32 SOM Clustering SOM is a special type of neural net-work designed to map the input image of sizeNx timesNy toMclusters based on their characteristic features [25] For SOMthe image (I) is converted into a feature vectorf f1 f2 fm where m is the number of features Inthis experiment we have trained SOM with M 4 clustersusing p 9 neighbourhood features such as given a centrepixel (gc) in the image the neighbourhood features arecomputed as given in the following equation

F(1 p) gp p isin [1 n] (2)

where n is the number of neighbourhood (3 times 3 window) gpis the neighbourhoods and F is the feature vector

Mammograms

Contrastenhancement

Pectoralmuscle

removal

SOMclustering

ROIextraction

Preprocessing

Curvelettransform

Scale 1sparse

coefficient

Scale nsparse

coefficientScales 2 to nndash1

sparse coefficient

ANNclassification

LBPpattern

Sparsecurvelet

coefficients

False positive reduction

Figure 1 Schematic architecture for automatic breast cancer detection

MIAS

(a) (b) (c) (d) (e) (f) (g)

Mammogram Preprocessing

Step I

Features forSOM

SOMclustering

9timesmtimesn

Step II

ROI extraction

Original image

Figure 2 Steps for mammogram processing (a) enhanced mammogram (b) binary mask (c) pectoral removal (d) pectoral removedmammogram (e) clustered image (f ) cluster of interest and (g) ROI extraction

Journal of Healthcare Engineering 3

corresponds to centre pixel gc e selection of 3times 3 windowpixel is based on [43] to capture local details

At the start weight vector Wi wi1 wi2 wimminus11113864 1113865 israndom and updated as the network learns e minimumEuclidean distance fminusWi is described as the bestmatching component or winner node fminusWc and de-scribed as

fminusWc

mini fminusWi

1113966 1113967 (3)

Weight vector for winning output neuron and itsneighboring neurons are updated as

Wi(t + 1) Wi(t) + Nci(t) f(t)minusWi(t)( 1113857 (4)

where t 1 2 is time coordinate e function Nci(t) isthe neighbourhood kernel function and expressed as

Nci(t) η(t)exp minusmc minusm2

i

2σ2(t)1113888 1113889 (5)

where η(t) is the learning rate σ(t) is a width of kernel thatcorresponds to neighbourhood neurons around node c andmc and mi corresponds to location vectors of nodes c and i

Figures 4(a) and 4(b) represent cluster map and clusterboundaries marked on mammogram After the severalobservations for known areas it was empirically noticed thatnumber of pixels of range or pixel level threshold (PLT basedon pixel count in TP) as 450 to 31500 16000 to 200000and 4000 to 200000 consist of abnormality for MIASDDSM and TMCH database respectively which is verifiedfrom the expert e size of the tumor is varying because ofthe mammogram size of 1024 times 1024 pixels for MIAS2728 times 3920 pixels to 4608 times 6048 pixels for DDSM and2294 times 1914 or 4096 times 3328 pixels for TMCH datasetserefore cluster regions below or above the specifiedthreshold are discarded and the remaining region is markedas true positive (TP) as shown in Figure 4 Figure 4(a) showsthe clustered image using SOM Figure 4(b) shows thecluster boundaries marked on original image

We can see that there are many FPs along with TP(marked by pink color) which are reduced using pixel levelthreshold (PLT based on pixel count in TP) as explainedabove Figure 4(c) shows the filtered result using PLT

33 ROI Extraction After SOM clustering (initial segmen-tation) the next step is to classify the detected regions intoTP and FP by using proposed local sparse curvelet features(LSCF) followed by ANN classifier To do so initially wehave extracted ROIs from detected regions by SOM clus-tering and manually categorized into TP and FP We col-lected these ROIs from three different datasets according totheir maximum height and maximum width using con-nected components eg region marked in Figure 4(c)erefore their patch size is different as shown in Figure 5ROIs for MIAS DDSM and TMCH dataset Further theseextracted patches have been used to train the ANN for thetask of FP reduction

34 False-Positive (FP) Reduction After ROI extraction FPreduction algorithm performs computation of proposedlocal sparse curvelet features (LSCF) followed by ANNclassifier

341 Proposed Algorithm LBP [43] was proposed as LBPdescriptor computation at circular neighbourhood which iscalled as uniform LBP (ULBP) descriptor and expressed as

ULBP(PR) 1113936

Pminus1

n1S if U LBP(PR)1113872 1113873le 2

P + 1 otherwise

⎧⎪⎪⎨

⎪⎪⎩(6)

where

U LBP(PR)1113872 1113873 S gPminus1 minusgC( 1113857minus S g0 minusgC( 11138571113868111386811138681113868

1113868111386811138681113868

+ 1113944Pminus1

n1S gP minusgC( 1113857minus S gPminus1 minusgC( 1113857

11138681113868111386811138681113868111386811138681113868

(7)

Computation of LBP based on actual shape of massaccording to sparse matrix has been shown in Figure 6 where

(a) (b)

A

B

C

D

F E

G

H

(c) (d)

Figure 3 Preprocessing (a) Original image from MIAS database (b) Contrast-enhanced mammogram using local entropy maximization(c) Process of pectoral muscle removal (d) Pectoral muscle removed mammogram

4 Journal of Healthcare Engineering

it takes pixels related to shape of mass which are called asforeground pixels and rejects the other pixels called asbackground pixels e proposed algorithm uses foregroundpixels only for LBP computation and this will tend to numberof pixel reduction in LBP computations erefore identi-cation of foreground and background pixels is an importantstep which is performed using lookup table approach e

identication of foreground and background pixel is based onnumber of nonzero pixels in the lookup table ie if count ofsliding window nonzero pixels is greater than 2 count(p(i j))gt 2 is identied as foreground and LBP is estimated On theother hand if count of sliding window nonzero pixels is lessthan 2 count(p(i j))lt 2 is identied as background and LBPwould not be estimated and rejected from lookup table

(a) (b) (c)

Figure 4 FP reduction by thresholding (a) clustered image (b) clusters boundaries marked on original image and (c) clusters afterthresholding

MIAS ROI patches from abnormal mammogram

DDSM ROI patches from abnormal mammogram

TMCH ldquoGE Medical Senograph Systemrdquo ROIpatches from abnormal mammogram

TMCH ldquoHologic Selenia Systemrdquo ROI patchesfrom abnormal mammogram

Figure 5 Variable sizes ROIs from MIAS DDSM and TMCH datasets

Input matrixLook-up table LBP code computation

Count

Count=number of (pixel Intensitiesgt0)

Count (pixel intensitiesgt0)is not greater than 2

LBP code27 26 25 24 23 22 21 20

If Count (pixel intensitygt0)gt2calculate LBP code

ElseDiscard

Discard row

Similarly other pixels also compared in look up table Similarly calculate LBP code for other pixel if countgt2Centre pixel

Centre pixel

Shape of mass

g4

g4

g3

g3

g2

g2

3times3 sliding window

g5

g5

gc

gc

63 90 23

2

33

0 0 0 0 100 123 490 10 0 23 0 63 100 123 4 610 11 0 0 23 90 123 4 2 611 0 0 0 0

0 0 0 0 0 0 0 0 0

0 0 0 0 0

0 0 0 0 0 0 00 0 0 0 0 0 000 0 0 0 0 0 0 0

0 0 00 0 00

0 0 0

0 0 000 1

1 1 11

1

1

1 1 11 1

1

1 1 112

131

154

128

0 0 00

0 0 0 00 00 0 0

000

0 0

10 4 2 134 4100 123 90 63 0 0 0 23 64 5123 4 10 90 63 100 23

2323

64

646564

5

55

5

6

6

6

g1

g1

g6

g6

g7

g7

g8

g8

0 0 23 0 0 00 63 90 10 11 00 100 123 4 2 1340 23 64 5 6 00 0 65 0 0 00 0 0 0 0 0

Figure 6 Lookup table approach for LBP computation from shape of mass in ROI

Journal of Healthcare Engineering 5

Nonzero pixels provide actual shape of mass and are taken forLBP computations Graphical representation of proposedalgorithm for LBP descriptor computation using foregroundpixels has been given in Figure 7 and the algorithm has beendescribed in Algorithm 2

342 e Fast Discrete Curvelet Transform (FDCT) eauthors [44] have introduced computationally simple andeiexclcient Fast Discrete Curvelet Transform (FDCT) We havepreferred wrapping-based FDCT approach in proposedwork as it is faster e curvelet coeiexclcients CD(j l k)represented by scale j angle l and spatial location k can bewritten as

CD j l k1 k2( ) sumn1N1

n11sumn2N2

n21I n1 n2[ ]φD

jlk1k2n1 n2[ ] (8)

Figure 8 illustrates LBP code computation based on sparsecurvelet coeiexclcients ROI decomposes using curvelet trans-form with scale orientations l of 16deg and scale of 2 as thedatabase consists of minimum ROI size of 25 times 22 pixelsCurvelet transform with scale orientations l of 16deg and scale of2 produces 1 + 16 17 diyenerent subbands based on subbanddivision Further each curvelet subband coeiexclcients havebeen represented using lookup table using 3 times 3 slidingwindow and if the row in the lookup table identies fore-ground coeiexclcient then LBP is computed with radius R 1and P 8 neighboring pixels as shown in Algorithm 2 total 58LBP features have been obtained from foreground curveletsubband coeiexclcients erefore total 986 LBP features havebeen extracted from 17 curvelet subbands It can be observedfrom Figure 8 curvelet subbands also provide shape of massin 16 diyenerent directions so that the directional informationcan be associated with LBP features Kanadam et al [3] usedconcept of sparse ROI similarly we have extended it forsparse curvelet subband and LBP features computation

35 Classi cation In this work we have analyzed extractedROI from mammogram using normal-abnormal benign-malignant and normal-malignant classes with ANN SVMand KNN classiers e detailed description of ANNclassier has been given in [45 46] To evaluate performance

of the proposed system we have used 3-fold cross validationwhere database is randomly divided into three sets andaccuracy is calculated for each set e nal accuracy of thesystem is average of accuracy of each of three sets Howeverit will not be fair to compare 3-fold cross validation result ofSVM and KNN classier with ANN because ANN classieris tested on only one set of images (33 for training 33 fortesting and 33 for validation)us to do fair comparisonwe have trained ANN using input layer (986 neuron) overthree diyenerent sets (which are considered in SVM and KNN)and calculated its average accuracy Our proposed falsepositive reduction algorithm illustrates in Figures 9(a)ndash9(c)Algorithm 3 summarizes copyow of the proposed method forFP reduction in mammograms

4 Experimental Results and Discussions

e proposed method has been tested and validated usingthree classiers and three clinical mammographic imagedatasets

41 Data Sets

411 Mammographic Image Analysis Society (MIAS)Database e mini-MIAS [17] database consists of 322mammograms each having 1024times1024 pixels and anno-tated like background tissue character class severity centerof abnormality and radius of circle for abnormality isdatabase includes 64 benign 51 malignant and 207 normalcases which have been taken for experimentation

412 Digital Database for Screening Mammography(DDSM) e DDSM [19] dataset consists of 2500 studiesand is composed of cranial-caudal (CC) and mediolateral-oblique (MLO) views of mammographic image for left andright breast annotated with ACR breast density type ofabnormality and ground truth Randomly selected 150abnormal and 100 normal cases from both HOWTEK andLUMISYS scanner of 12 bits per pixel resolution have beensubjected for experimentation

ROI Shape of mass

Mass shape by sparse matrix

Position inlook-up table Nine neighborhood pixels Decision making

Yes

LBP operator on 3times3 neighborhood

Nine neighborhood pixels

120 125 210

105 45

254

10

93 200 250

1 1 1

1 0

1 1 1

Thresholding

LBP code

LBP code =0 + 2 + 4 +8 + 16 + 32+ 64 + 128

No

Pixel count(position(ij))

gt 2

Discard the rowfrom look-up table

eg X P etc

Input

bullbull

bullbull

bullbull

bullbull

bull

X 0 0 0 0 0 0 0 0 1

P 0 0 0 0 0 0 0 0 0

Q 0 0 0 37 200 0 100 102 0

Y 123 90 70 100 23 45 0 144 17

Z 120 105 93 125 45 200 210 10 250

Q position in look-up table

P position

Z position

Y position

X position

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

Figure 7 Process for computation of LBP descriptor from shape of mass in ROI (a) Original image (b) 3times 3 window for selection offoreground pixels (c) lookup table (d) decision making process (e) LBP computation from selected foreground pixels

6 Journal of Healthcare Engineering

413 e Tata Memorial Cancer Hospital (TMCH) isdataset [47] contains 360 full-eld digital mammograms(FFDMs) comprising 180 CC views and 180 MLO viewsfrom right and left breast acquired from 90 randomly se-lected patients It is composed of 180 veried malignant and180 normal breast images It uses biopsy proven breastcancer patientsrsquo pathological data approved by the In-stitutional Research Ethics Committee of Tata MemorialCentre Hospital (TMCH) Mumbai India e ground truthmarking on each abnormal mammogram is performedmanually using the Histopathological Reports (HPR) of therespective patients and expert radiologist from TMCHMumbai Approximately 35 patients are examined usingldquoHologic Selenia Systemrdquo (Scanner1) gives 16-bit

e remaining 55 patients were examined with ldquoGEMedical Senograph Systemrdquo (Scanner2) providing 8-bit truecolor mammogram image in DICOM format of 4096times 3328or 2294times1914 pixels each measuring size 50times 50 μm2

42 Segmentation Evaluation and ROI Extraction e seg-mentation using SOM that detects suspicious mass regions isconsidered as TP whereas from nonmass is taken as FPFrom Table 1 it is clear that total suspicious ROI (includingTP amp FP) of 381 for MIAS 1343 for DDSM and 1009 forTMCH have been taken for evaluation our proposed al-gorithm for FP reduction

From extracted ROIs the minimum patch size is25 times 22 pixels whereas the maximum size is 1152 times1356pixels Tables 2 and 3 represent curvelet subband co-eiexclcients from 17 subbands and reduced coeiexclcientsbased on lookup table approach are used to calculate LBPfeatures It has been observed during experimentationthat the curvelet coeiexclcients on an average are reducedfor sparse LBP by 14 32 33 and 34 for MIASDDSM TMCH Scanner1 and TMCH Scanner2 re-spectively It may be noticed that reduction in curveletcoeiexclcients for every ROI is not xed It completely

bullbullbull

bullbullbull

bullbullbull

Scale 1 approximatecoefficients

Scale 2 1 to 16subband coefficients

Curvelet transform

1

58

58

2

158 LBP

descriptorsfrom

approximatecoefficients

58 lowast 16LBP

descriptorsfrom

1 to 16subband curvelet

coefficients

2

Figure 8 LBP code computation using sparse curvelet subband coeiexclcients

(a) (b) (c)

Figure 9 (a) FP reduction by clusters marked on original image (b) FP reduction by thresholding (c) FP reduction by sparse curveletcoeiexclcient-based LBP and ANN

Journal of Healthcare Engineering 7

depends upon the shape of the ROI as per the sparsematrix Tables 2 and 3 do not represent exact reduction inpixels for complete database but they exhibit pixel re-duction for sample mammograms

43 Classifier Evaluation and False-Positive ReductionFrom Figures 10ndash13 the best classification accuracy of 9857 has been obtained for MIAS in benign versus malignantclassification whereas 9870 for DDSM 9830 for

(1) Load input image (img1)(2) Apply CLAHE algorithm and obtain enhanced image (img2)(3) Decompose img1 and img2 up to 3 level of decomposition using Discrete Wavelet transform (DWT)(4) Use maximum local entropy rule for fusion of img1 and img2 for high frequency subbands(5) Take inverse DWT to obtain the fused image

ALGORITHM 1 Image fusion for contrast enhancement

Input I(m n) m no of rows and nno of columnOutput LBP featuresInitialize Radius R 1 and neighborhood pixels P 8

Mask [1 2 4 8 16 32 64 128 0]Sliding window coordinates kminus1 1Count 1 number of pixels in I(m n)

for i 1 to m dofor j 1 to n do

prepare local circular windowI_local I(i+ k j+ k)center_pixel I(i j)Arrange local neighborhoods of I(i j) pixels in a row col 1 9Lookup_table (i col) reshape(I_local [7 17])count number of pixels greater than zeroa length(find(Lookup_tablegt 0))select pixel position from lookup-table for computation of LBPif agt 2

LBP_code(count) I_localgt center_pixelcount count + 1

endend

endcompute histogram of LBP codesLBP_descriptor LBP_descriptorcountscale invariant

ALGORITHM 2 Algorithm for LBP feature computation based on shape of mass in ROI as

(1) Load input image (img1)(2) Apply CLAHE algorithm and obtain enhanced image (img2)(3) Process img1 and img2 and obtain enhanced image using procedure given in Algorithm 1(4) Remove pectoral muscle using proposed approach (Section 312)(5) Extract neighbourhood features for each pixel and apply SOM clustering(6) Obtain clustered image and separate out the tumorous cluster(7) Extract detected regions ie ROIrsquos from clustered result(8) Extract Sparse Curvelet Coefficients (Subband) up to 2 level from each ROI(9) Extract Sparse LBP code for each subband and obtain a combined feature vector for each ROI(10) Classify each ROI into tumorous and nontumorous class ie TP and FP respectively(11) Map each TP region on original mammogram (img1)(12) end

ALGORITHM 3 Summary of proposed method for FP reduction in mammograms

8 Journal of Healthcare Engineering

TMCH Scanner1 and 100 for TMCH Scanner2 clas-sication accuracies have been obtained in normal versusmalignant classication e classication performance ofANN has improved from 6 to 43 for diyenerent databasesas compared to KNN classier whereas there is littleimprovement about 7 compared with SVM classier eperformances of both proposed sparse LBP and LBPcomputation on curvelet subbands are nearly sametherefore the proposed algorithm can be eiexclciently

implemented in CAD system with lesser number of cur-velet coeiexclcients

Data augmentation has been used for some classes tomaintain balance between two classes to improve perfor-mance and to learn more powerful model Table 4 explainsthe FP reduction with the use of curvelet-based LBP featuresand ANN It has been observed that FP reduced from 085 to002 FPimage in MIAS 481 to 002 FPimage in DDSM and232 to 013 FPimage in TMCH

8000

8500

9000

9500

10000

TMCHS1LBP

TMCHS1sparse_LBP

TMCHS2LBP

TMCHS2sparse_LBP

Normal_Malignant

KNNANNSVM

Figure 10 Average classication rate for TMCH dataset

000

2000

4000

6000

8000

10000

DDSMLBP DDSMsparse_LBP

MIASLBP MIASsparse_LBP

Normal_Malignant

KNNANNSVM

Figure 11 Average classication rate for MIAS and DDSM dataset

000

2000

4000

6000

8000

10000

DDSMLBP DDSMsparse_LBP

MIASLBP MIASsparse_LBP

Benign_Malignant

KNNANNSVM

Figure 12 Average classication rate for MIAS and DDSM dataset

Journal of Healthcare Engineering 9

Similarly Table 5 shows the reduction in FPs as 085to 001 FPimage for MIAS 481 to 003 FPimage forDDSM and 232 to 000 FPimage for TMCH using sparsecurvelet coeiexclcient-based LBP features e results show

the eyenectiveness of sparse curvelet coeiexclcient-based LBPand ANN From Table 6 the best value of AUC 099 isobtained in benign versus malignant classication forMIAS AUC 098 in benign versus malignant in case of

Table 1 Result of SOM segmentation

Datasetused

Result of SOM clustering and threshold TPR (true-positiverate)

TPlesions

FPPI (false-positiveper image)FPimages

MassSegmented nonmass (FP) Total () images

Segmented (TP) LostMIAS 108 7 273 322 (108115) 094 (273322) 085DDSM 140 10 1203 250 (140150) 093 (1203250) 481TMCH 172 8 837 360 (172180) 095 (837360) 232

Table 2 Reduction in curvelet coeiexclcients for sample mammograms from MIAS and DDSM dataset

SrNo

MIAS DDSM

ROI Size

Total number ofcurvelet

coeiexclcients fromsubbands

Total number ofselected curveletcoeiexclcients from

subbands

reductionin curveletcoeiexclcients

ROI Size

Total number ofcurvelet

coeiexclcients fromsubbands

Total number ofselected curveletcoeiexclcients from

subbands

reductionin curveletcoeiexclcients

1 124times138 103911 77133 2577 192times187 216729 168333 22332 179times138 150123 126142 1597 294times 291 518267 366680 29253 51times 116 36815 33421 922 145times 207 181663 148765 18114 83times 83 42449 36653 1365 169times168 171873 154752 9965 84times 76 39115 35815 844 182times 248 272517 219822 19346 74times 83 37767 34448 879 213times 349 449783 301359 33007 53times 64 20969 18621 1120 578times 412 1433195 662072 53808 70times 44 18899 16610 1211 215times 219 286429 242935 15189 80times 66 32409 29454 912 420times 428 1082461 501209 537010 69times 86 36783 33552 878 203times 307 375871 266829 290111 59times116 42019 38442 851 226times 262 357209 276763 225212 81times 101 50637 46122 892 159times194 187563 148741 207013 41times 84 21427 18907 1176 718times 686 2961127 762561 742514 69times141 60641 52427 1354 409times 550 1352439 904829 331015 60times 62 22925 20475 1069 524times 375 1184671 766522 353016 96times101 59647 54235 907 311times 275 517433 397737 231317 136times139 113727 66352 4156 319times 320 614129 412606 328118 55times 94 31359 28305 974 313times 447 842855 510482 394319 157times140 132373 107000 1917 291times 517 907903 637412 297920 156times130 123007 90834 2615 370times 837 1864889 898236 5183

Average 58850 48247 14 Average 788950 437432 32

000

2000

4000

6000

8000

10000

DDSMLBP DDSM sparse_LBP

MIASLBP MIAS sparse_LBP

Normal_Abnormal

KNNANNSVM

Figure 13 Average classication rate for MIAS and DDSM dataset

10 Journal of Healthcare Engineering

DDSM AUC 094 in normal versus malignant in case ofTMCH Scanner1 and AUC 096 in normal versusmalignant classification in TMCH Scanner2 using ANNand curvelet subband-based LBP features e worstperformance of AUC 053 for MIAS is obtained with theproposed algorithm using KNN classifier as shown inTable 7 Similarly from Table 7 the best value ofAUC 098 is obtained in TMCH Scanner1 AUC 1 isobtained in TMCH Scanner2 database for normal versusmalignant classification AUC 098 in benign versusmalignant classification is attained in MIAS database andAUC 098 is achieved for normal versus malignant

classification in DDSM database using ANN classifier forsparse curvelet subband-based LBP features

However from Table 7 it should be noted that the per-formance of proposed algorithm is the best using ANN clas-sifier Figure 14 represents automated CAD system for breastcancer diagnosis with sample mammograms

Table 8 provides comparative study of methods developedfor breast tissue classification e proposed method providesbest results in terms of AUC and reduction of number of FPsas 085 to 001 FPimage for MIAS 481 to 003 FPimage forDDSM and 232 to 000 FPimage for TMCH e earlierreported work uses the fixed patch size-based approach which

Table 3 Reduction in curvelet coefficients for sample mammograms from TMCH Scanner1 and Scanner2 dataset

Srno

TMCH Scanner 1 ldquoGE Medical Senograph Systemrdquo TMCH Scanner 2 ldquoHologic Selenia Systemrdquo

ROI size

Total number ofcurvelet

coefficients fromsubbands

Total number ofselected curveletcoefficients from

subbands

reductionin curveletcoefficients

ROI size

Total number ofcurvelet

coefficients fromsubbands

Total number ofselected curveletcoefficients from

subbands

reductionin curveletcoefficients

1 459times 412 1140617 794885 3031 291times 278 489243 317014 35202 548times 513 1693403 1117873 3399 545times 246 809627 531604 34343 415times 303 758323 445585 4124 560times 483 1630583 1068439 34474 645times 495 1951443 1145580 4129 782times 510 2400137 1275073 46875 437times 691 1816651 985120 4577 87times141 75773 67871 10436 812times 500 2439937 1287065 4725 311times 185 348565 240546 30997 468times 379 1066333 710242 3340 262times 348 550303 282821 48618 673times 582 2355589 1730915 2652 610times 440 1614515 842876 47799 250times 201 304513 235670 2261 949times 391 2227209 1359338 389710 525times 488 1544691 1161942 2478 365times 385 846523 604473 285911 488times 779 2287547 1611385 2956 393times 247 585063 490474 161712 1434times 966 8326581 3742277 5506 341times 301 618111 410542 335813 348times 421 881701 616501 3008 523times 702 2206097 800057 637314 460times 530 1467227 799885 4548 370times 284 632539 423727 330115 398times 450 1078441 828064 2322 344times 202 418955 302427 278116 247times 272 404401 344822 1473 264times188 299997 231926 226917 411times 305 757657 405919 4642 233times 247 346983 267686 228518 286times 344 593155 448566 2438 370times 291 650701 486125 252919 417times 207 523477 429755 1790 680times 483 1979543 1052793 468220 463times 458 1275021 857608 3274 202times 266 324295 215042 3369

Average 1633335 984983 33 Average 952738 563543 34

Table 4 Number of ROIs resulted in FP reduction using curvelet-based LBP (without sparse) amp ANN classification at training andvalidation stage

Class Datasetused

Benignmalignant mass Nonmassbenign mass Total()

images

TPR (true-positiverate)TPlesions

FPPI (false-positiveper image) FP

imagesPreviousstage

Selected(TP)

Lost(FN)

Previousstage

Selected(TN)

Lost(FP)

Normal vsabnormal

MIAS 108lowast 2 216 203 13 273 257 16 315 (203216) 094 (16315) 005DDSM 140lowast 4 560 465 95 1203 1095 108 240 (465560) 083 (108240) 045

Benign vsmalignant

MIAS 49 49 0 59 57 2 108 (4949) 100 (2108) 002DDSM 46lowast 2 92 91 1 94 91 3 140 (9192) 099 (3140) 002

Normal vsmalignant

MIAS 49lowast 4196 184 12 273 254 19 256 (184196) 094 (19256) 007DDSM 46lowast 4184 180 4 1203 1143 60 146 (180184) 098 (60146) 041TMCHScanner1 107lowast 4 428 416 12 605 551 54 217 (416428) 097 (54217) 025

TMCHScanner2 65lowast 4 260 255 5 232 214 18 135 (255260) 098 (18135) 013

lowastAugmentation of image

Journal of Healthcare Engineering 11

limits the automatic CAD system scope whereas proposedsystem provides complete solution to CAD system right fromautomatic tumor patch segmentation to reduction in FPs andfinal representation of mammogram with TP marked on itIt will drastically reduce the radiologist work by locationtumor directly on mammogram

5 Conclusion

A fully automatic CAD system which can accuratelylocate the tumor on a mammogram and reduces FPshas been proposed e developed CAD system consistsof preprocessing SOM clustering ROI extraction

Table 6 Performance evaluation of curvelet-based LBP descriptor algorithm

Dataset Classification Normal-malignant Normal-abnormal Benign-malignantClassifier Sensitivity Specificity AUC Sensitivity Specificity AUC Sensitivity Specificity AUC

MIASANN 094 093 094 094 094 094 100 097 099SVM 085 085 085 083 086 085 088 084 086KNN 067 063 065 058 057 058 062 068 063

DDSMANN 098 095 095 083 091 085 099 097 098SVM 097 088 092 071 091 083 094 089 092KNN 096 064 087 067 090 080 087 073 079

TMCH Scanner1ANN 097 091 094 mdash mdash mdash mdash mdash mdashSVM 096 091 094 mdash mdash mdash mdash mdash mdashKNN 098 082 089 mdash mdash mdash mdash mdash mdash

TMCH Scanner2ANN 098 092 096 mdash mdash mdash mdash mdash mdashSVM 097 090 094 mdash mdash mdash mdash mdash mdashKNN 092 083 088 mdash mdash mdash mdash mdash mdash

Table 7 Performance evaluation of proposed algorithm

Dataset Classification Normal-malignant Normal-abnormal Benign-malignantClassifier Sensitivity Specificity AUC Sensitivity Specificity AUC Sensitivity Specificity AUC

MIASANN 098 095 096 093 097 095 097 100 098SVM 088 083 085 085 082 084 084 092 087KNN 055 051 053 055 063 056 061 067 061

DDSMANN 099 097 098 092 096 093 097 095 096SVM 099 092 096 089 096 092 094 092 093KNN 098 073 092 074 090 082 089 077 083

TMCH Scanner1ANN 099 098 098 mdash mdash mdash mdash mdash mdashSVM 098 096 097 mdash mdash mdash mdash mdash mdashKNN 099 092 095 mdash mdash mdash mdash mdash mdash

TMCH Scanner2ANN 100 100 100 mdash mdash mdash mdash mdash mdashSVM 100 098 099 mdash mdash mdash mdash mdash mdashKNN 096 092 094 mdash mdash mdash mdash mdash mdash

Table 5 Number of ROIs resulted in FP reduction using sparse curvelet coefficient-based LBP amp ANN classification at training andvalidation stage

Class Datasetused

Benignmalignant mass Nonmassbenign massTotal ()images

TPR (true-positiverate)TPlesions

FPPI (false-positiveper image) FP

imagesPreviousstage

Selected(TP)

Lost(FN)

Previousstage

Selected(TN)

Lost(FP)

Normal vsabnormal

MIAS 108lowast 2 216 201 15 273 265 8 315 (201216) 093 (8315) 002DDSM 140lowast 4 560 516 44 1203 1155 48 240 (516560) 092 (48240) 02

Benign vsmalignant

MIAS 49 48 1 59 59 1 108 (4849) 098 (1108) 001DDSM 46lowast 2 92 89 3 94 89 5 140 (8992) 097 (5140) 003

Normal vsmalignant

MIAS 49lowast 4196 192 4 273 259 14 256 (192196) 098 (14256) 005DDSM 46lowast 4184 182 2 1203 1167 36 146 (182184) 099 (36146) 025TMCHScanner1 107lowast 4 428 424 4 605 593 12 217 (424428) 099 (12217) 005

TMCHScanner2 65lowast 4 260 260 0 232 232 0 135 (260260) 100 (0135) 0

lowastAugmentation of image

12 Journal of Healthcare Engineering

sparse LBP feature computation based on sparse Cur-velet coefficients and finally FP reduction using ANNclassifier

e proposed algorithm presents a novel concept ofextraction of curvelet coefficients according to irregularshape of mass is called as sparse curvelet coefficients andcomputation of LBP e analysis proves that the FPs arereduced significantly from 085 to 001 FPimage for MIAS

481 to 003 FPimage for DDSM and 232 to 000 FPimagefor TMCH e ANN classifier showed best results asAUC 098 and accuracy 9857 for MIAS in benign-malignant classification AUC 098 and accuracy 9870for DDSM in normal-malignant classification AUC 098and accuracy 9830 for TMCH Scanner1 and AUC 1and accuracy 100 for TMCH Scanner2 in normal-malignant classification as compared with SVM and KNN

(a) (b) (c) (d) (e) (f )

Figure 14 Representation of fully automatic CAD system for breast cancer using (a) samplemammograms fromMIAS DDSM and TMCHdatasets (b) preprocessed mammograms (c) clustered image (d) TP and FP marked on mammogram (e) TP marked by thresholding (f )TP marked by using LBP descriptor based on sparse curvelet coefficients

Journal of Healthcare Engineering 13

classifier e performance of LBP features and LBP featuresbased on sparse curvelet coefficients are nearly same whichshow that the proposed algorithm is suitable for cancer breasttissue diagnosis

In future the reduced curvelet coefficients can be used toextract local ternary patterns and other local descriptor andlocal directional patterns etc e present work deals withmammogram with single mass this can be further extendedfor multiple mass models with multiple LBP features basedon sparse curvelet coefficients

Data Availability

In this research we have used two publicly available datasetsMIAS and DDSM ese datasets can be found here in [17]and [19] e third database is collected from the localhospital Tata Memorial Cancer Hospital Mumbai whichcan be found at httpeurekasveriacin or available fromthe corresponding author upon request

Conflicts of Interest

e authors declare that they have no conflicts of interest

Acknowledgments

e TMCH database for this work was given by Departmentof Radiodiagnosis Tata Memorial Cancer Hospital Mumbai

References

[1] S Malvia S A Bagadi U S Dubey and S Saxena ldquoEpi-demiology of breast cancer in Indian womenrdquo Asia-PacificJournal of Clinical Oncology vol 13 no 4 pp 289ndash295 2017

[2] A Gupta K Shridhar and P Dhillon ldquoA review of breastcancer awareness among women in India cancer literate orawareness deficitrdquo European Journal of Cancer vol 51no 14 pp 2058ndash2066 2015

[3] K P Kanadam and S R Chereddy ldquoMammogram classifi-cation using sparse-ROI a novel representation to arbitraryshaped massesrdquo Expert Systems with Applications vol 57pp 204ndash213 2016

[4] D O T Bruno M Z do Nascimento R P Ramos et al ldquoLBPoperators on curvelet coefficients as an algorithm to describetexture in breast cancer tissuesrdquo Expert Systems with Appli-cations vol 55 pp 329ndash340 2016

[5] M Hussain ldquoFalse-positive reduction in mammographyusing multiscale spatial Weber law descriptor and supportvector machinesrdquo Neural Computing and Applicationsvol 25 no 1 pp 83ndash93 2014

[6] M M Pawar and S N Talbar ldquoGenetic fuzzy system (GFS)based wavelet co-occurrence feature selection in mammo-gram classification for breast cancer diagnosisrdquo Perspectives inScience vol 8 pp 247ndash250 2016

[7] C Muramatsu T Hara T Endo and H Fujita ldquoBreast massclassification on mammograms using radial local ternarypatternsrdquo Computers in Biology and Medicine vol 72pp 43ndash53 2016

[8] M M Eltoukhy I Faye and B B Samir ldquoA statistical basedfeature extraction method for breast cancer diagnosis indigital mammogram using multiresolution representationrdquo

Table 8 Comparison of classification accuracy AUC and FPimage values from different approaches in breast cancer diagnosis

Author Database Method Classifier Result AUC FPimageEltoukhy et al [33]

MIAS Biggest curvelet coefficients as a featurevector

Euclidean classifier 9407 mdash mdashEltoukhy et al [42] 9859 mdash mdashEltoukhy et al [8] SVM 973 mdash mdash

Dhahbi et al [34] Mini-MIAS Curvelet moments KNN 9127 mdash mdashDDSM 8646 mdash mdash

Bruno et al [4] DDSM Curvelet + LBP SVM 85 085 mdashPL 94 094 mdash

da Rocha et al [40] DDSM LBP SVM 8831 088 mdashKanadam andChereddy [3] MIAS Sparse ROI SVM 9742 mdash mdash

Pereira et al [18] DDSM Wavelet and Wiener filter Multiple thresholding waveletand GA mdash mdash 137

Liu and Zeng [29] DDSMFFDM GLCM CLBP and geometric features SVM mdash mdash 148

De Sampaio et al[39] DDSM LBP DBSCAN 9826 019

Zyout et al [30] DDSM Second order statistics of waveletcoefficients (SOSWC) SVM 968 097 0018

MIAS 952 966 0029

Casti et al [31]DDSM

Differential features Fisher linear discriminantanalysis (FLDA) mdash mdash

168MIAS 212FFDM 082

Proposed method

MIAS

LBP based on sparse curvelet subbandcoefficients ANN

9857 098 001DDSM 9870 098 003TMCHScanner1 9830 098 005

TMCHScanner2 100 1 0

14 Journal of Healthcare Engineering

fusion algorithm Further original image along with theCLAHE has been given to the image fusion algorithmProcedure of the image fusion has been given in Algo-rithm 1 We have used local entropy as a fusion rule given bythe following equation

ENT minussum255

k0p(k)log(p(k)) (1)

where ENT is the local entropy and p_org(k) andp_CLAHE(k) are the probability of kth pixel from 5times 5sliding window [12] Here both high frequency componentsfrom original mammogram and CLAHE mammogram havebeen fused using maximum entropy criteria Figure 3(b)presents contrast-enhanced mammogram using local en-tropy maximization-based image fusion

312 Pectoral Muscle Removal Pectoral muscle suppressionhas been performed by dening rectangle as suggested in[14] (Figure 3(c)) It illustrates the rectangle (ABDC) and

xes the points G and has intensity variation and joins themfor pectoral muscle suppression Figure 3(d) illustratespectoral muscle removed image to avoid discrepancies in thealgorithm because of similar intensities present betweenpectoral muscle and masses

32 SOM Clustering SOM is a special type of neural net-work designed to map the input image of sizeNx timesNy toMclusters based on their characteristic features [25] For SOMthe image (I) is converted into a feature vectorf f1 f2 fm where m is the number of features Inthis experiment we have trained SOM with M 4 clustersusing p 9 neighbourhood features such as given a centrepixel (gc) in the image the neighbourhood features arecomputed as given in the following equation

F(1 p) gp p isin [1 n] (2)

where n is the number of neighbourhood (3 times 3 window) gpis the neighbourhoods and F is the feature vector

Mammograms

Contrastenhancement

Pectoralmuscle

removal

SOMclustering

ROIextraction

Preprocessing

Curvelettransform

Scale 1sparse

coefficient

Scale nsparse

coefficientScales 2 to nndash1

sparse coefficient

ANNclassification

LBPpattern

Sparsecurvelet

coefficients

False positive reduction

Figure 1 Schematic architecture for automatic breast cancer detection

MIAS

(a) (b) (c) (d) (e) (f) (g)

Mammogram Preprocessing

Step I

Features forSOM

SOMclustering

9timesmtimesn

Step II

ROI extraction

Original image

Figure 2 Steps for mammogram processing (a) enhanced mammogram (b) binary mask (c) pectoral removal (d) pectoral removedmammogram (e) clustered image (f ) cluster of interest and (g) ROI extraction

Journal of Healthcare Engineering 3

corresponds to centre pixel gc e selection of 3times 3 windowpixel is based on [43] to capture local details

At the start weight vector Wi wi1 wi2 wimminus11113864 1113865 israndom and updated as the network learns e minimumEuclidean distance fminusWi is described as the bestmatching component or winner node fminusWc and de-scribed as

fminusWc

mini fminusWi

1113966 1113967 (3)

Weight vector for winning output neuron and itsneighboring neurons are updated as

Wi(t + 1) Wi(t) + Nci(t) f(t)minusWi(t)( 1113857 (4)

where t 1 2 is time coordinate e function Nci(t) isthe neighbourhood kernel function and expressed as

Nci(t) η(t)exp minusmc minusm2

i

2σ2(t)1113888 1113889 (5)

where η(t) is the learning rate σ(t) is a width of kernel thatcorresponds to neighbourhood neurons around node c andmc and mi corresponds to location vectors of nodes c and i

Figures 4(a) and 4(b) represent cluster map and clusterboundaries marked on mammogram After the severalobservations for known areas it was empirically noticed thatnumber of pixels of range or pixel level threshold (PLT basedon pixel count in TP) as 450 to 31500 16000 to 200000and 4000 to 200000 consist of abnormality for MIASDDSM and TMCH database respectively which is verifiedfrom the expert e size of the tumor is varying because ofthe mammogram size of 1024 times 1024 pixels for MIAS2728 times 3920 pixels to 4608 times 6048 pixels for DDSM and2294 times 1914 or 4096 times 3328 pixels for TMCH datasetserefore cluster regions below or above the specifiedthreshold are discarded and the remaining region is markedas true positive (TP) as shown in Figure 4 Figure 4(a) showsthe clustered image using SOM Figure 4(b) shows thecluster boundaries marked on original image

We can see that there are many FPs along with TP(marked by pink color) which are reduced using pixel levelthreshold (PLT based on pixel count in TP) as explainedabove Figure 4(c) shows the filtered result using PLT

33 ROI Extraction After SOM clustering (initial segmen-tation) the next step is to classify the detected regions intoTP and FP by using proposed local sparse curvelet features(LSCF) followed by ANN classifier To do so initially wehave extracted ROIs from detected regions by SOM clus-tering and manually categorized into TP and FP We col-lected these ROIs from three different datasets according totheir maximum height and maximum width using con-nected components eg region marked in Figure 4(c)erefore their patch size is different as shown in Figure 5ROIs for MIAS DDSM and TMCH dataset Further theseextracted patches have been used to train the ANN for thetask of FP reduction

34 False-Positive (FP) Reduction After ROI extraction FPreduction algorithm performs computation of proposedlocal sparse curvelet features (LSCF) followed by ANNclassifier

341 Proposed Algorithm LBP [43] was proposed as LBPdescriptor computation at circular neighbourhood which iscalled as uniform LBP (ULBP) descriptor and expressed as

ULBP(PR) 1113936

Pminus1

n1S if U LBP(PR)1113872 1113873le 2

P + 1 otherwise

⎧⎪⎪⎨

⎪⎪⎩(6)

where

U LBP(PR)1113872 1113873 S gPminus1 minusgC( 1113857minus S g0 minusgC( 11138571113868111386811138681113868

1113868111386811138681113868

+ 1113944Pminus1

n1S gP minusgC( 1113857minus S gPminus1 minusgC( 1113857

11138681113868111386811138681113868111386811138681113868

(7)

Computation of LBP based on actual shape of massaccording to sparse matrix has been shown in Figure 6 where

(a) (b)

A

B

C

D

F E

G

H

(c) (d)

Figure 3 Preprocessing (a) Original image from MIAS database (b) Contrast-enhanced mammogram using local entropy maximization(c) Process of pectoral muscle removal (d) Pectoral muscle removed mammogram

4 Journal of Healthcare Engineering

it takes pixels related to shape of mass which are called asforeground pixels and rejects the other pixels called asbackground pixels e proposed algorithm uses foregroundpixels only for LBP computation and this will tend to numberof pixel reduction in LBP computations erefore identi-cation of foreground and background pixels is an importantstep which is performed using lookup table approach e

identication of foreground and background pixel is based onnumber of nonzero pixels in the lookup table ie if count ofsliding window nonzero pixels is greater than 2 count(p(i j))gt 2 is identied as foreground and LBP is estimated On theother hand if count of sliding window nonzero pixels is lessthan 2 count(p(i j))lt 2 is identied as background and LBPwould not be estimated and rejected from lookup table

(a) (b) (c)

Figure 4 FP reduction by thresholding (a) clustered image (b) clusters boundaries marked on original image and (c) clusters afterthresholding

MIAS ROI patches from abnormal mammogram

DDSM ROI patches from abnormal mammogram

TMCH ldquoGE Medical Senograph Systemrdquo ROIpatches from abnormal mammogram

TMCH ldquoHologic Selenia Systemrdquo ROI patchesfrom abnormal mammogram

Figure 5 Variable sizes ROIs from MIAS DDSM and TMCH datasets

Input matrixLook-up table LBP code computation

Count

Count=number of (pixel Intensitiesgt0)

Count (pixel intensitiesgt0)is not greater than 2

LBP code27 26 25 24 23 22 21 20

If Count (pixel intensitygt0)gt2calculate LBP code

ElseDiscard

Discard row

Similarly other pixels also compared in look up table Similarly calculate LBP code for other pixel if countgt2Centre pixel

Centre pixel

Shape of mass

g4

g4

g3

g3

g2

g2

3times3 sliding window

g5

g5

gc

gc

63 90 23

2

33

0 0 0 0 100 123 490 10 0 23 0 63 100 123 4 610 11 0 0 23 90 123 4 2 611 0 0 0 0

0 0 0 0 0 0 0 0 0

0 0 0 0 0

0 0 0 0 0 0 00 0 0 0 0 0 000 0 0 0 0 0 0 0

0 0 00 0 00

0 0 0

0 0 000 1

1 1 11

1

1

1 1 11 1

1

1 1 112

131

154

128

0 0 00

0 0 0 00 00 0 0

000

0 0

10 4 2 134 4100 123 90 63 0 0 0 23 64 5123 4 10 90 63 100 23

2323

64

646564

5

55

5

6

6

6

g1

g1

g6

g6

g7

g7

g8

g8

0 0 23 0 0 00 63 90 10 11 00 100 123 4 2 1340 23 64 5 6 00 0 65 0 0 00 0 0 0 0 0

Figure 6 Lookup table approach for LBP computation from shape of mass in ROI

Journal of Healthcare Engineering 5

Nonzero pixels provide actual shape of mass and are taken forLBP computations Graphical representation of proposedalgorithm for LBP descriptor computation using foregroundpixels has been given in Figure 7 and the algorithm has beendescribed in Algorithm 2

342 e Fast Discrete Curvelet Transform (FDCT) eauthors [44] have introduced computationally simple andeiexclcient Fast Discrete Curvelet Transform (FDCT) We havepreferred wrapping-based FDCT approach in proposedwork as it is faster e curvelet coeiexclcients CD(j l k)represented by scale j angle l and spatial location k can bewritten as

CD j l k1 k2( ) sumn1N1

n11sumn2N2

n21I n1 n2[ ]φD

jlk1k2n1 n2[ ] (8)

Figure 8 illustrates LBP code computation based on sparsecurvelet coeiexclcients ROI decomposes using curvelet trans-form with scale orientations l of 16deg and scale of 2 as thedatabase consists of minimum ROI size of 25 times 22 pixelsCurvelet transform with scale orientations l of 16deg and scale of2 produces 1 + 16 17 diyenerent subbands based on subbanddivision Further each curvelet subband coeiexclcients havebeen represented using lookup table using 3 times 3 slidingwindow and if the row in the lookup table identies fore-ground coeiexclcient then LBP is computed with radius R 1and P 8 neighboring pixels as shown in Algorithm 2 total 58LBP features have been obtained from foreground curveletsubband coeiexclcients erefore total 986 LBP features havebeen extracted from 17 curvelet subbands It can be observedfrom Figure 8 curvelet subbands also provide shape of massin 16 diyenerent directions so that the directional informationcan be associated with LBP features Kanadam et al [3] usedconcept of sparse ROI similarly we have extended it forsparse curvelet subband and LBP features computation

35 Classi cation In this work we have analyzed extractedROI from mammogram using normal-abnormal benign-malignant and normal-malignant classes with ANN SVMand KNN classiers e detailed description of ANNclassier has been given in [45 46] To evaluate performance

of the proposed system we have used 3-fold cross validationwhere database is randomly divided into three sets andaccuracy is calculated for each set e nal accuracy of thesystem is average of accuracy of each of three sets Howeverit will not be fair to compare 3-fold cross validation result ofSVM and KNN classier with ANN because ANN classieris tested on only one set of images (33 for training 33 fortesting and 33 for validation)us to do fair comparisonwe have trained ANN using input layer (986 neuron) overthree diyenerent sets (which are considered in SVM and KNN)and calculated its average accuracy Our proposed falsepositive reduction algorithm illustrates in Figures 9(a)ndash9(c)Algorithm 3 summarizes copyow of the proposed method forFP reduction in mammograms

4 Experimental Results and Discussions

e proposed method has been tested and validated usingthree classiers and three clinical mammographic imagedatasets

41 Data Sets

411 Mammographic Image Analysis Society (MIAS)Database e mini-MIAS [17] database consists of 322mammograms each having 1024times1024 pixels and anno-tated like background tissue character class severity centerof abnormality and radius of circle for abnormality isdatabase includes 64 benign 51 malignant and 207 normalcases which have been taken for experimentation

412 Digital Database for Screening Mammography(DDSM) e DDSM [19] dataset consists of 2500 studiesand is composed of cranial-caudal (CC) and mediolateral-oblique (MLO) views of mammographic image for left andright breast annotated with ACR breast density type ofabnormality and ground truth Randomly selected 150abnormal and 100 normal cases from both HOWTEK andLUMISYS scanner of 12 bits per pixel resolution have beensubjected for experimentation

ROI Shape of mass

Mass shape by sparse matrix

Position inlook-up table Nine neighborhood pixels Decision making

Yes

LBP operator on 3times3 neighborhood

Nine neighborhood pixels

120 125 210

105 45

254

10

93 200 250

1 1 1

1 0

1 1 1

Thresholding

LBP code

LBP code =0 + 2 + 4 +8 + 16 + 32+ 64 + 128

No

Pixel count(position(ij))

gt 2

Discard the rowfrom look-up table

eg X P etc

Input

bullbull

bullbull

bullbull

bullbull

bull

X 0 0 0 0 0 0 0 0 1

P 0 0 0 0 0 0 0 0 0

Q 0 0 0 37 200 0 100 102 0

Y 123 90 70 100 23 45 0 144 17

Z 120 105 93 125 45 200 210 10 250

Q position in look-up table

P position

Z position

Y position

X position

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

Figure 7 Process for computation of LBP descriptor from shape of mass in ROI (a) Original image (b) 3times 3 window for selection offoreground pixels (c) lookup table (d) decision making process (e) LBP computation from selected foreground pixels

6 Journal of Healthcare Engineering

413 e Tata Memorial Cancer Hospital (TMCH) isdataset [47] contains 360 full-eld digital mammograms(FFDMs) comprising 180 CC views and 180 MLO viewsfrom right and left breast acquired from 90 randomly se-lected patients It is composed of 180 veried malignant and180 normal breast images It uses biopsy proven breastcancer patientsrsquo pathological data approved by the In-stitutional Research Ethics Committee of Tata MemorialCentre Hospital (TMCH) Mumbai India e ground truthmarking on each abnormal mammogram is performedmanually using the Histopathological Reports (HPR) of therespective patients and expert radiologist from TMCHMumbai Approximately 35 patients are examined usingldquoHologic Selenia Systemrdquo (Scanner1) gives 16-bit

e remaining 55 patients were examined with ldquoGEMedical Senograph Systemrdquo (Scanner2) providing 8-bit truecolor mammogram image in DICOM format of 4096times 3328or 2294times1914 pixels each measuring size 50times 50 μm2

42 Segmentation Evaluation and ROI Extraction e seg-mentation using SOM that detects suspicious mass regions isconsidered as TP whereas from nonmass is taken as FPFrom Table 1 it is clear that total suspicious ROI (includingTP amp FP) of 381 for MIAS 1343 for DDSM and 1009 forTMCH have been taken for evaluation our proposed al-gorithm for FP reduction

From extracted ROIs the minimum patch size is25 times 22 pixels whereas the maximum size is 1152 times1356pixels Tables 2 and 3 represent curvelet subband co-eiexclcients from 17 subbands and reduced coeiexclcientsbased on lookup table approach are used to calculate LBPfeatures It has been observed during experimentationthat the curvelet coeiexclcients on an average are reducedfor sparse LBP by 14 32 33 and 34 for MIASDDSM TMCH Scanner1 and TMCH Scanner2 re-spectively It may be noticed that reduction in curveletcoeiexclcients for every ROI is not xed It completely

bullbullbull

bullbullbull

bullbullbull

Scale 1 approximatecoefficients

Scale 2 1 to 16subband coefficients

Curvelet transform

1

58

58

2

158 LBP

descriptorsfrom

approximatecoefficients

58 lowast 16LBP

descriptorsfrom

1 to 16subband curvelet

coefficients

2

Figure 8 LBP code computation using sparse curvelet subband coeiexclcients

(a) (b) (c)

Figure 9 (a) FP reduction by clusters marked on original image (b) FP reduction by thresholding (c) FP reduction by sparse curveletcoeiexclcient-based LBP and ANN

Journal of Healthcare Engineering 7

depends upon the shape of the ROI as per the sparsematrix Tables 2 and 3 do not represent exact reduction inpixels for complete database but they exhibit pixel re-duction for sample mammograms

43 Classifier Evaluation and False-Positive ReductionFrom Figures 10ndash13 the best classification accuracy of 9857 has been obtained for MIAS in benign versus malignantclassification whereas 9870 for DDSM 9830 for

(1) Load input image (img1)(2) Apply CLAHE algorithm and obtain enhanced image (img2)(3) Decompose img1 and img2 up to 3 level of decomposition using Discrete Wavelet transform (DWT)(4) Use maximum local entropy rule for fusion of img1 and img2 for high frequency subbands(5) Take inverse DWT to obtain the fused image

ALGORITHM 1 Image fusion for contrast enhancement

Input I(m n) m no of rows and nno of columnOutput LBP featuresInitialize Radius R 1 and neighborhood pixels P 8

Mask [1 2 4 8 16 32 64 128 0]Sliding window coordinates kminus1 1Count 1 number of pixels in I(m n)

for i 1 to m dofor j 1 to n do

prepare local circular windowI_local I(i+ k j+ k)center_pixel I(i j)Arrange local neighborhoods of I(i j) pixels in a row col 1 9Lookup_table (i col) reshape(I_local [7 17])count number of pixels greater than zeroa length(find(Lookup_tablegt 0))select pixel position from lookup-table for computation of LBPif agt 2

LBP_code(count) I_localgt center_pixelcount count + 1

endend

endcompute histogram of LBP codesLBP_descriptor LBP_descriptorcountscale invariant

ALGORITHM 2 Algorithm for LBP feature computation based on shape of mass in ROI as

(1) Load input image (img1)(2) Apply CLAHE algorithm and obtain enhanced image (img2)(3) Process img1 and img2 and obtain enhanced image using procedure given in Algorithm 1(4) Remove pectoral muscle using proposed approach (Section 312)(5) Extract neighbourhood features for each pixel and apply SOM clustering(6) Obtain clustered image and separate out the tumorous cluster(7) Extract detected regions ie ROIrsquos from clustered result(8) Extract Sparse Curvelet Coefficients (Subband) up to 2 level from each ROI(9) Extract Sparse LBP code for each subband and obtain a combined feature vector for each ROI(10) Classify each ROI into tumorous and nontumorous class ie TP and FP respectively(11) Map each TP region on original mammogram (img1)(12) end

ALGORITHM 3 Summary of proposed method for FP reduction in mammograms

8 Journal of Healthcare Engineering

TMCH Scanner1 and 100 for TMCH Scanner2 clas-sication accuracies have been obtained in normal versusmalignant classication e classication performance ofANN has improved from 6 to 43 for diyenerent databasesas compared to KNN classier whereas there is littleimprovement about 7 compared with SVM classier eperformances of both proposed sparse LBP and LBPcomputation on curvelet subbands are nearly sametherefore the proposed algorithm can be eiexclciently

implemented in CAD system with lesser number of cur-velet coeiexclcients

Data augmentation has been used for some classes tomaintain balance between two classes to improve perfor-mance and to learn more powerful model Table 4 explainsthe FP reduction with the use of curvelet-based LBP featuresand ANN It has been observed that FP reduced from 085 to002 FPimage in MIAS 481 to 002 FPimage in DDSM and232 to 013 FPimage in TMCH

8000

8500

9000

9500

10000

TMCHS1LBP

TMCHS1sparse_LBP

TMCHS2LBP

TMCHS2sparse_LBP

Normal_Malignant

KNNANNSVM

Figure 10 Average classication rate for TMCH dataset

000

2000

4000

6000

8000

10000

DDSMLBP DDSMsparse_LBP

MIASLBP MIASsparse_LBP

Normal_Malignant

KNNANNSVM

Figure 11 Average classication rate for MIAS and DDSM dataset

000

2000

4000

6000

8000

10000

DDSMLBP DDSMsparse_LBP

MIASLBP MIASsparse_LBP

Benign_Malignant

KNNANNSVM

Figure 12 Average classication rate for MIAS and DDSM dataset

Journal of Healthcare Engineering 9

Similarly Table 5 shows the reduction in FPs as 085to 001 FPimage for MIAS 481 to 003 FPimage forDDSM and 232 to 000 FPimage for TMCH using sparsecurvelet coeiexclcient-based LBP features e results show

the eyenectiveness of sparse curvelet coeiexclcient-based LBPand ANN From Table 6 the best value of AUC 099 isobtained in benign versus malignant classication forMIAS AUC 098 in benign versus malignant in case of

Table 1 Result of SOM segmentation

Datasetused

Result of SOM clustering and threshold TPR (true-positiverate)

TPlesions

FPPI (false-positiveper image)FPimages

MassSegmented nonmass (FP) Total () images

Segmented (TP) LostMIAS 108 7 273 322 (108115) 094 (273322) 085DDSM 140 10 1203 250 (140150) 093 (1203250) 481TMCH 172 8 837 360 (172180) 095 (837360) 232

Table 2 Reduction in curvelet coeiexclcients for sample mammograms from MIAS and DDSM dataset

SrNo

MIAS DDSM

ROI Size

Total number ofcurvelet

coeiexclcients fromsubbands

Total number ofselected curveletcoeiexclcients from

subbands

reductionin curveletcoeiexclcients

ROI Size

Total number ofcurvelet

coeiexclcients fromsubbands

Total number ofselected curveletcoeiexclcients from

subbands

reductionin curveletcoeiexclcients

1 124times138 103911 77133 2577 192times187 216729 168333 22332 179times138 150123 126142 1597 294times 291 518267 366680 29253 51times 116 36815 33421 922 145times 207 181663 148765 18114 83times 83 42449 36653 1365 169times168 171873 154752 9965 84times 76 39115 35815 844 182times 248 272517 219822 19346 74times 83 37767 34448 879 213times 349 449783 301359 33007 53times 64 20969 18621 1120 578times 412 1433195 662072 53808 70times 44 18899 16610 1211 215times 219 286429 242935 15189 80times 66 32409 29454 912 420times 428 1082461 501209 537010 69times 86 36783 33552 878 203times 307 375871 266829 290111 59times116 42019 38442 851 226times 262 357209 276763 225212 81times 101 50637 46122 892 159times194 187563 148741 207013 41times 84 21427 18907 1176 718times 686 2961127 762561 742514 69times141 60641 52427 1354 409times 550 1352439 904829 331015 60times 62 22925 20475 1069 524times 375 1184671 766522 353016 96times101 59647 54235 907 311times 275 517433 397737 231317 136times139 113727 66352 4156 319times 320 614129 412606 328118 55times 94 31359 28305 974 313times 447 842855 510482 394319 157times140 132373 107000 1917 291times 517 907903 637412 297920 156times130 123007 90834 2615 370times 837 1864889 898236 5183

Average 58850 48247 14 Average 788950 437432 32

000

2000

4000

6000

8000

10000

DDSMLBP DDSM sparse_LBP

MIASLBP MIAS sparse_LBP

Normal_Abnormal

KNNANNSVM

Figure 13 Average classication rate for MIAS and DDSM dataset

10 Journal of Healthcare Engineering

DDSM AUC 094 in normal versus malignant in case ofTMCH Scanner1 and AUC 096 in normal versusmalignant classification in TMCH Scanner2 using ANNand curvelet subband-based LBP features e worstperformance of AUC 053 for MIAS is obtained with theproposed algorithm using KNN classifier as shown inTable 7 Similarly from Table 7 the best value ofAUC 098 is obtained in TMCH Scanner1 AUC 1 isobtained in TMCH Scanner2 database for normal versusmalignant classification AUC 098 in benign versusmalignant classification is attained in MIAS database andAUC 098 is achieved for normal versus malignant

classification in DDSM database using ANN classifier forsparse curvelet subband-based LBP features

However from Table 7 it should be noted that the per-formance of proposed algorithm is the best using ANN clas-sifier Figure 14 represents automated CAD system for breastcancer diagnosis with sample mammograms

Table 8 provides comparative study of methods developedfor breast tissue classification e proposed method providesbest results in terms of AUC and reduction of number of FPsas 085 to 001 FPimage for MIAS 481 to 003 FPimage forDDSM and 232 to 000 FPimage for TMCH e earlierreported work uses the fixed patch size-based approach which

Table 3 Reduction in curvelet coefficients for sample mammograms from TMCH Scanner1 and Scanner2 dataset

Srno

TMCH Scanner 1 ldquoGE Medical Senograph Systemrdquo TMCH Scanner 2 ldquoHologic Selenia Systemrdquo

ROI size

Total number ofcurvelet

coefficients fromsubbands

Total number ofselected curveletcoefficients from

subbands

reductionin curveletcoefficients

ROI size

Total number ofcurvelet

coefficients fromsubbands

Total number ofselected curveletcoefficients from

subbands

reductionin curveletcoefficients

1 459times 412 1140617 794885 3031 291times 278 489243 317014 35202 548times 513 1693403 1117873 3399 545times 246 809627 531604 34343 415times 303 758323 445585 4124 560times 483 1630583 1068439 34474 645times 495 1951443 1145580 4129 782times 510 2400137 1275073 46875 437times 691 1816651 985120 4577 87times141 75773 67871 10436 812times 500 2439937 1287065 4725 311times 185 348565 240546 30997 468times 379 1066333 710242 3340 262times 348 550303 282821 48618 673times 582 2355589 1730915 2652 610times 440 1614515 842876 47799 250times 201 304513 235670 2261 949times 391 2227209 1359338 389710 525times 488 1544691 1161942 2478 365times 385 846523 604473 285911 488times 779 2287547 1611385 2956 393times 247 585063 490474 161712 1434times 966 8326581 3742277 5506 341times 301 618111 410542 335813 348times 421 881701 616501 3008 523times 702 2206097 800057 637314 460times 530 1467227 799885 4548 370times 284 632539 423727 330115 398times 450 1078441 828064 2322 344times 202 418955 302427 278116 247times 272 404401 344822 1473 264times188 299997 231926 226917 411times 305 757657 405919 4642 233times 247 346983 267686 228518 286times 344 593155 448566 2438 370times 291 650701 486125 252919 417times 207 523477 429755 1790 680times 483 1979543 1052793 468220 463times 458 1275021 857608 3274 202times 266 324295 215042 3369

Average 1633335 984983 33 Average 952738 563543 34

Table 4 Number of ROIs resulted in FP reduction using curvelet-based LBP (without sparse) amp ANN classification at training andvalidation stage

Class Datasetused

Benignmalignant mass Nonmassbenign mass Total()

images

TPR (true-positiverate)TPlesions

FPPI (false-positiveper image) FP

imagesPreviousstage

Selected(TP)

Lost(FN)

Previousstage

Selected(TN)

Lost(FP)

Normal vsabnormal

MIAS 108lowast 2 216 203 13 273 257 16 315 (203216) 094 (16315) 005DDSM 140lowast 4 560 465 95 1203 1095 108 240 (465560) 083 (108240) 045

Benign vsmalignant

MIAS 49 49 0 59 57 2 108 (4949) 100 (2108) 002DDSM 46lowast 2 92 91 1 94 91 3 140 (9192) 099 (3140) 002

Normal vsmalignant

MIAS 49lowast 4196 184 12 273 254 19 256 (184196) 094 (19256) 007DDSM 46lowast 4184 180 4 1203 1143 60 146 (180184) 098 (60146) 041TMCHScanner1 107lowast 4 428 416 12 605 551 54 217 (416428) 097 (54217) 025

TMCHScanner2 65lowast 4 260 255 5 232 214 18 135 (255260) 098 (18135) 013

lowastAugmentation of image

Journal of Healthcare Engineering 11

limits the automatic CAD system scope whereas proposedsystem provides complete solution to CAD system right fromautomatic tumor patch segmentation to reduction in FPs andfinal representation of mammogram with TP marked on itIt will drastically reduce the radiologist work by locationtumor directly on mammogram

5 Conclusion

A fully automatic CAD system which can accuratelylocate the tumor on a mammogram and reduces FPshas been proposed e developed CAD system consistsof preprocessing SOM clustering ROI extraction

Table 6 Performance evaluation of curvelet-based LBP descriptor algorithm

Dataset Classification Normal-malignant Normal-abnormal Benign-malignantClassifier Sensitivity Specificity AUC Sensitivity Specificity AUC Sensitivity Specificity AUC

MIASANN 094 093 094 094 094 094 100 097 099SVM 085 085 085 083 086 085 088 084 086KNN 067 063 065 058 057 058 062 068 063

DDSMANN 098 095 095 083 091 085 099 097 098SVM 097 088 092 071 091 083 094 089 092KNN 096 064 087 067 090 080 087 073 079

TMCH Scanner1ANN 097 091 094 mdash mdash mdash mdash mdash mdashSVM 096 091 094 mdash mdash mdash mdash mdash mdashKNN 098 082 089 mdash mdash mdash mdash mdash mdash

TMCH Scanner2ANN 098 092 096 mdash mdash mdash mdash mdash mdashSVM 097 090 094 mdash mdash mdash mdash mdash mdashKNN 092 083 088 mdash mdash mdash mdash mdash mdash

Table 7 Performance evaluation of proposed algorithm

Dataset Classification Normal-malignant Normal-abnormal Benign-malignantClassifier Sensitivity Specificity AUC Sensitivity Specificity AUC Sensitivity Specificity AUC

MIASANN 098 095 096 093 097 095 097 100 098SVM 088 083 085 085 082 084 084 092 087KNN 055 051 053 055 063 056 061 067 061

DDSMANN 099 097 098 092 096 093 097 095 096SVM 099 092 096 089 096 092 094 092 093KNN 098 073 092 074 090 082 089 077 083

TMCH Scanner1ANN 099 098 098 mdash mdash mdash mdash mdash mdashSVM 098 096 097 mdash mdash mdash mdash mdash mdashKNN 099 092 095 mdash mdash mdash mdash mdash mdash

TMCH Scanner2ANN 100 100 100 mdash mdash mdash mdash mdash mdashSVM 100 098 099 mdash mdash mdash mdash mdash mdashKNN 096 092 094 mdash mdash mdash mdash mdash mdash

Table 5 Number of ROIs resulted in FP reduction using sparse curvelet coefficient-based LBP amp ANN classification at training andvalidation stage

Class Datasetused

Benignmalignant mass Nonmassbenign massTotal ()images

TPR (true-positiverate)TPlesions

FPPI (false-positiveper image) FP

imagesPreviousstage

Selected(TP)

Lost(FN)

Previousstage

Selected(TN)

Lost(FP)

Normal vsabnormal

MIAS 108lowast 2 216 201 15 273 265 8 315 (201216) 093 (8315) 002DDSM 140lowast 4 560 516 44 1203 1155 48 240 (516560) 092 (48240) 02

Benign vsmalignant

MIAS 49 48 1 59 59 1 108 (4849) 098 (1108) 001DDSM 46lowast 2 92 89 3 94 89 5 140 (8992) 097 (5140) 003

Normal vsmalignant

MIAS 49lowast 4196 192 4 273 259 14 256 (192196) 098 (14256) 005DDSM 46lowast 4184 182 2 1203 1167 36 146 (182184) 099 (36146) 025TMCHScanner1 107lowast 4 428 424 4 605 593 12 217 (424428) 099 (12217) 005

TMCHScanner2 65lowast 4 260 260 0 232 232 0 135 (260260) 100 (0135) 0

lowastAugmentation of image

12 Journal of Healthcare Engineering

sparse LBP feature computation based on sparse Cur-velet coefficients and finally FP reduction using ANNclassifier

e proposed algorithm presents a novel concept ofextraction of curvelet coefficients according to irregularshape of mass is called as sparse curvelet coefficients andcomputation of LBP e analysis proves that the FPs arereduced significantly from 085 to 001 FPimage for MIAS

481 to 003 FPimage for DDSM and 232 to 000 FPimagefor TMCH e ANN classifier showed best results asAUC 098 and accuracy 9857 for MIAS in benign-malignant classification AUC 098 and accuracy 9870for DDSM in normal-malignant classification AUC 098and accuracy 9830 for TMCH Scanner1 and AUC 1and accuracy 100 for TMCH Scanner2 in normal-malignant classification as compared with SVM and KNN

(a) (b) (c) (d) (e) (f )

Figure 14 Representation of fully automatic CAD system for breast cancer using (a) samplemammograms fromMIAS DDSM and TMCHdatasets (b) preprocessed mammograms (c) clustered image (d) TP and FP marked on mammogram (e) TP marked by thresholding (f )TP marked by using LBP descriptor based on sparse curvelet coefficients

Journal of Healthcare Engineering 13

classifier e performance of LBP features and LBP featuresbased on sparse curvelet coefficients are nearly same whichshow that the proposed algorithm is suitable for cancer breasttissue diagnosis

In future the reduced curvelet coefficients can be used toextract local ternary patterns and other local descriptor andlocal directional patterns etc e present work deals withmammogram with single mass this can be further extendedfor multiple mass models with multiple LBP features basedon sparse curvelet coefficients

Data Availability

In this research we have used two publicly available datasetsMIAS and DDSM ese datasets can be found here in [17]and [19] e third database is collected from the localhospital Tata Memorial Cancer Hospital Mumbai whichcan be found at httpeurekasveriacin or available fromthe corresponding author upon request

Conflicts of Interest

e authors declare that they have no conflicts of interest

Acknowledgments

e TMCH database for this work was given by Departmentof Radiodiagnosis Tata Memorial Cancer Hospital Mumbai

References

[1] S Malvia S A Bagadi U S Dubey and S Saxena ldquoEpi-demiology of breast cancer in Indian womenrdquo Asia-PacificJournal of Clinical Oncology vol 13 no 4 pp 289ndash295 2017

[2] A Gupta K Shridhar and P Dhillon ldquoA review of breastcancer awareness among women in India cancer literate orawareness deficitrdquo European Journal of Cancer vol 51no 14 pp 2058ndash2066 2015

[3] K P Kanadam and S R Chereddy ldquoMammogram classifi-cation using sparse-ROI a novel representation to arbitraryshaped massesrdquo Expert Systems with Applications vol 57pp 204ndash213 2016

[4] D O T Bruno M Z do Nascimento R P Ramos et al ldquoLBPoperators on curvelet coefficients as an algorithm to describetexture in breast cancer tissuesrdquo Expert Systems with Appli-cations vol 55 pp 329ndash340 2016

[5] M Hussain ldquoFalse-positive reduction in mammographyusing multiscale spatial Weber law descriptor and supportvector machinesrdquo Neural Computing and Applicationsvol 25 no 1 pp 83ndash93 2014

[6] M M Pawar and S N Talbar ldquoGenetic fuzzy system (GFS)based wavelet co-occurrence feature selection in mammo-gram classification for breast cancer diagnosisrdquo Perspectives inScience vol 8 pp 247ndash250 2016

[7] C Muramatsu T Hara T Endo and H Fujita ldquoBreast massclassification on mammograms using radial local ternarypatternsrdquo Computers in Biology and Medicine vol 72pp 43ndash53 2016

[8] M M Eltoukhy I Faye and B B Samir ldquoA statistical basedfeature extraction method for breast cancer diagnosis indigital mammogram using multiresolution representationrdquo

Table 8 Comparison of classification accuracy AUC and FPimage values from different approaches in breast cancer diagnosis

Author Database Method Classifier Result AUC FPimageEltoukhy et al [33]

MIAS Biggest curvelet coefficients as a featurevector

Euclidean classifier 9407 mdash mdashEltoukhy et al [42] 9859 mdash mdashEltoukhy et al [8] SVM 973 mdash mdash

Dhahbi et al [34] Mini-MIAS Curvelet moments KNN 9127 mdash mdashDDSM 8646 mdash mdash

Bruno et al [4] DDSM Curvelet + LBP SVM 85 085 mdashPL 94 094 mdash

da Rocha et al [40] DDSM LBP SVM 8831 088 mdashKanadam andChereddy [3] MIAS Sparse ROI SVM 9742 mdash mdash

Pereira et al [18] DDSM Wavelet and Wiener filter Multiple thresholding waveletand GA mdash mdash 137

Liu and Zeng [29] DDSMFFDM GLCM CLBP and geometric features SVM mdash mdash 148

De Sampaio et al[39] DDSM LBP DBSCAN 9826 019

Zyout et al [30] DDSM Second order statistics of waveletcoefficients (SOSWC) SVM 968 097 0018

MIAS 952 966 0029

Casti et al [31]DDSM

Differential features Fisher linear discriminantanalysis (FLDA) mdash mdash

168MIAS 212FFDM 082

Proposed method

MIAS

LBP based on sparse curvelet subbandcoefficients ANN

9857 098 001DDSM 9870 098 003TMCHScanner1 9830 098 005

TMCHScanner2 100 1 0

14 Journal of Healthcare Engineering

corresponds to centre pixel gc e selection of 3times 3 windowpixel is based on [43] to capture local details

At the start weight vector Wi wi1 wi2 wimminus11113864 1113865 israndom and updated as the network learns e minimumEuclidean distance fminusWi is described as the bestmatching component or winner node fminusWc and de-scribed as

fminusWc

mini fminusWi

1113966 1113967 (3)

Weight vector for winning output neuron and itsneighboring neurons are updated as

Wi(t + 1) Wi(t) + Nci(t) f(t)minusWi(t)( 1113857 (4)

where t 1 2 is time coordinate e function Nci(t) isthe neighbourhood kernel function and expressed as

Nci(t) η(t)exp minusmc minusm2

i

2σ2(t)1113888 1113889 (5)

where η(t) is the learning rate σ(t) is a width of kernel thatcorresponds to neighbourhood neurons around node c andmc and mi corresponds to location vectors of nodes c and i

Figures 4(a) and 4(b) represent cluster map and clusterboundaries marked on mammogram After the severalobservations for known areas it was empirically noticed thatnumber of pixels of range or pixel level threshold (PLT basedon pixel count in TP) as 450 to 31500 16000 to 200000and 4000 to 200000 consist of abnormality for MIASDDSM and TMCH database respectively which is verifiedfrom the expert e size of the tumor is varying because ofthe mammogram size of 1024 times 1024 pixels for MIAS2728 times 3920 pixels to 4608 times 6048 pixels for DDSM and2294 times 1914 or 4096 times 3328 pixels for TMCH datasetserefore cluster regions below or above the specifiedthreshold are discarded and the remaining region is markedas true positive (TP) as shown in Figure 4 Figure 4(a) showsthe clustered image using SOM Figure 4(b) shows thecluster boundaries marked on original image

We can see that there are many FPs along with TP(marked by pink color) which are reduced using pixel levelthreshold (PLT based on pixel count in TP) as explainedabove Figure 4(c) shows the filtered result using PLT

33 ROI Extraction After SOM clustering (initial segmen-tation) the next step is to classify the detected regions intoTP and FP by using proposed local sparse curvelet features(LSCF) followed by ANN classifier To do so initially wehave extracted ROIs from detected regions by SOM clus-tering and manually categorized into TP and FP We col-lected these ROIs from three different datasets according totheir maximum height and maximum width using con-nected components eg region marked in Figure 4(c)erefore their patch size is different as shown in Figure 5ROIs for MIAS DDSM and TMCH dataset Further theseextracted patches have been used to train the ANN for thetask of FP reduction

34 False-Positive (FP) Reduction After ROI extraction FPreduction algorithm performs computation of proposedlocal sparse curvelet features (LSCF) followed by ANNclassifier

341 Proposed Algorithm LBP [43] was proposed as LBPdescriptor computation at circular neighbourhood which iscalled as uniform LBP (ULBP) descriptor and expressed as

ULBP(PR) 1113936

Pminus1

n1S if U LBP(PR)1113872 1113873le 2

P + 1 otherwise

⎧⎪⎪⎨

⎪⎪⎩(6)

where

U LBP(PR)1113872 1113873 S gPminus1 minusgC( 1113857minus S g0 minusgC( 11138571113868111386811138681113868

1113868111386811138681113868

+ 1113944Pminus1

n1S gP minusgC( 1113857minus S gPminus1 minusgC( 1113857

11138681113868111386811138681113868111386811138681113868

(7)

Computation of LBP based on actual shape of massaccording to sparse matrix has been shown in Figure 6 where

(a) (b)

A

B

C

D

F E

G

H

(c) (d)

Figure 3 Preprocessing (a) Original image from MIAS database (b) Contrast-enhanced mammogram using local entropy maximization(c) Process of pectoral muscle removal (d) Pectoral muscle removed mammogram

4 Journal of Healthcare Engineering

it takes pixels related to shape of mass which are called asforeground pixels and rejects the other pixels called asbackground pixels e proposed algorithm uses foregroundpixels only for LBP computation and this will tend to numberof pixel reduction in LBP computations erefore identi-cation of foreground and background pixels is an importantstep which is performed using lookup table approach e

identication of foreground and background pixel is based onnumber of nonzero pixels in the lookup table ie if count ofsliding window nonzero pixels is greater than 2 count(p(i j))gt 2 is identied as foreground and LBP is estimated On theother hand if count of sliding window nonzero pixels is lessthan 2 count(p(i j))lt 2 is identied as background and LBPwould not be estimated and rejected from lookup table

(a) (b) (c)

Figure 4 FP reduction by thresholding (a) clustered image (b) clusters boundaries marked on original image and (c) clusters afterthresholding

MIAS ROI patches from abnormal mammogram

DDSM ROI patches from abnormal mammogram

TMCH ldquoGE Medical Senograph Systemrdquo ROIpatches from abnormal mammogram

TMCH ldquoHologic Selenia Systemrdquo ROI patchesfrom abnormal mammogram

Figure 5 Variable sizes ROIs from MIAS DDSM and TMCH datasets

Input matrixLook-up table LBP code computation

Count

Count=number of (pixel Intensitiesgt0)

Count (pixel intensitiesgt0)is not greater than 2

LBP code27 26 25 24 23 22 21 20

If Count (pixel intensitygt0)gt2calculate LBP code

ElseDiscard

Discard row

Similarly other pixels also compared in look up table Similarly calculate LBP code for other pixel if countgt2Centre pixel

Centre pixel

Shape of mass

g4

g4

g3

g3

g2

g2

3times3 sliding window

g5

g5

gc

gc

63 90 23

2

33

0 0 0 0 100 123 490 10 0 23 0 63 100 123 4 610 11 0 0 23 90 123 4 2 611 0 0 0 0

0 0 0 0 0 0 0 0 0

0 0 0 0 0

0 0 0 0 0 0 00 0 0 0 0 0 000 0 0 0 0 0 0 0

0 0 00 0 00

0 0 0

0 0 000 1

1 1 11

1

1

1 1 11 1

1

1 1 112

131

154

128

0 0 00

0 0 0 00 00 0 0

000

0 0

10 4 2 134 4100 123 90 63 0 0 0 23 64 5123 4 10 90 63 100 23

2323

64

646564

5

55

5

6

6

6

g1

g1

g6

g6

g7

g7

g8

g8

0 0 23 0 0 00 63 90 10 11 00 100 123 4 2 1340 23 64 5 6 00 0 65 0 0 00 0 0 0 0 0

Figure 6 Lookup table approach for LBP computation from shape of mass in ROI

Journal of Healthcare Engineering 5

Nonzero pixels provide actual shape of mass and are taken forLBP computations Graphical representation of proposedalgorithm for LBP descriptor computation using foregroundpixels has been given in Figure 7 and the algorithm has beendescribed in Algorithm 2

342 e Fast Discrete Curvelet Transform (FDCT) eauthors [44] have introduced computationally simple andeiexclcient Fast Discrete Curvelet Transform (FDCT) We havepreferred wrapping-based FDCT approach in proposedwork as it is faster e curvelet coeiexclcients CD(j l k)represented by scale j angle l and spatial location k can bewritten as

CD j l k1 k2( ) sumn1N1

n11sumn2N2

n21I n1 n2[ ]φD

jlk1k2n1 n2[ ] (8)

Figure 8 illustrates LBP code computation based on sparsecurvelet coeiexclcients ROI decomposes using curvelet trans-form with scale orientations l of 16deg and scale of 2 as thedatabase consists of minimum ROI size of 25 times 22 pixelsCurvelet transform with scale orientations l of 16deg and scale of2 produces 1 + 16 17 diyenerent subbands based on subbanddivision Further each curvelet subband coeiexclcients havebeen represented using lookup table using 3 times 3 slidingwindow and if the row in the lookup table identies fore-ground coeiexclcient then LBP is computed with radius R 1and P 8 neighboring pixels as shown in Algorithm 2 total 58LBP features have been obtained from foreground curveletsubband coeiexclcients erefore total 986 LBP features havebeen extracted from 17 curvelet subbands It can be observedfrom Figure 8 curvelet subbands also provide shape of massin 16 diyenerent directions so that the directional informationcan be associated with LBP features Kanadam et al [3] usedconcept of sparse ROI similarly we have extended it forsparse curvelet subband and LBP features computation

35 Classi cation In this work we have analyzed extractedROI from mammogram using normal-abnormal benign-malignant and normal-malignant classes with ANN SVMand KNN classiers e detailed description of ANNclassier has been given in [45 46] To evaluate performance

of the proposed system we have used 3-fold cross validationwhere database is randomly divided into three sets andaccuracy is calculated for each set e nal accuracy of thesystem is average of accuracy of each of three sets Howeverit will not be fair to compare 3-fold cross validation result ofSVM and KNN classier with ANN because ANN classieris tested on only one set of images (33 for training 33 fortesting and 33 for validation)us to do fair comparisonwe have trained ANN using input layer (986 neuron) overthree diyenerent sets (which are considered in SVM and KNN)and calculated its average accuracy Our proposed falsepositive reduction algorithm illustrates in Figures 9(a)ndash9(c)Algorithm 3 summarizes copyow of the proposed method forFP reduction in mammograms

4 Experimental Results and Discussions

e proposed method has been tested and validated usingthree classiers and three clinical mammographic imagedatasets

41 Data Sets

411 Mammographic Image Analysis Society (MIAS)Database e mini-MIAS [17] database consists of 322mammograms each having 1024times1024 pixels and anno-tated like background tissue character class severity centerof abnormality and radius of circle for abnormality isdatabase includes 64 benign 51 malignant and 207 normalcases which have been taken for experimentation

412 Digital Database for Screening Mammography(DDSM) e DDSM [19] dataset consists of 2500 studiesand is composed of cranial-caudal (CC) and mediolateral-oblique (MLO) views of mammographic image for left andright breast annotated with ACR breast density type ofabnormality and ground truth Randomly selected 150abnormal and 100 normal cases from both HOWTEK andLUMISYS scanner of 12 bits per pixel resolution have beensubjected for experimentation

ROI Shape of mass

Mass shape by sparse matrix

Position inlook-up table Nine neighborhood pixels Decision making

Yes

LBP operator on 3times3 neighborhood

Nine neighborhood pixels

120 125 210

105 45

254

10

93 200 250

1 1 1

1 0

1 1 1

Thresholding

LBP code

LBP code =0 + 2 + 4 +8 + 16 + 32+ 64 + 128

No

Pixel count(position(ij))

gt 2

Discard the rowfrom look-up table

eg X P etc

Input

bullbull

bullbull

bullbull

bullbull

bull

X 0 0 0 0 0 0 0 0 1

P 0 0 0 0 0 0 0 0 0

Q 0 0 0 37 200 0 100 102 0

Y 123 90 70 100 23 45 0 144 17

Z 120 105 93 125 45 200 210 10 250

Q position in look-up table

P position

Z position

Y position

X position

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

Figure 7 Process for computation of LBP descriptor from shape of mass in ROI (a) Original image (b) 3times 3 window for selection offoreground pixels (c) lookup table (d) decision making process (e) LBP computation from selected foreground pixels

6 Journal of Healthcare Engineering

413 e Tata Memorial Cancer Hospital (TMCH) isdataset [47] contains 360 full-eld digital mammograms(FFDMs) comprising 180 CC views and 180 MLO viewsfrom right and left breast acquired from 90 randomly se-lected patients It is composed of 180 veried malignant and180 normal breast images It uses biopsy proven breastcancer patientsrsquo pathological data approved by the In-stitutional Research Ethics Committee of Tata MemorialCentre Hospital (TMCH) Mumbai India e ground truthmarking on each abnormal mammogram is performedmanually using the Histopathological Reports (HPR) of therespective patients and expert radiologist from TMCHMumbai Approximately 35 patients are examined usingldquoHologic Selenia Systemrdquo (Scanner1) gives 16-bit

e remaining 55 patients were examined with ldquoGEMedical Senograph Systemrdquo (Scanner2) providing 8-bit truecolor mammogram image in DICOM format of 4096times 3328or 2294times1914 pixels each measuring size 50times 50 μm2

42 Segmentation Evaluation and ROI Extraction e seg-mentation using SOM that detects suspicious mass regions isconsidered as TP whereas from nonmass is taken as FPFrom Table 1 it is clear that total suspicious ROI (includingTP amp FP) of 381 for MIAS 1343 for DDSM and 1009 forTMCH have been taken for evaluation our proposed al-gorithm for FP reduction

From extracted ROIs the minimum patch size is25 times 22 pixels whereas the maximum size is 1152 times1356pixels Tables 2 and 3 represent curvelet subband co-eiexclcients from 17 subbands and reduced coeiexclcientsbased on lookup table approach are used to calculate LBPfeatures It has been observed during experimentationthat the curvelet coeiexclcients on an average are reducedfor sparse LBP by 14 32 33 and 34 for MIASDDSM TMCH Scanner1 and TMCH Scanner2 re-spectively It may be noticed that reduction in curveletcoeiexclcients for every ROI is not xed It completely

bullbullbull

bullbullbull

bullbullbull

Scale 1 approximatecoefficients

Scale 2 1 to 16subband coefficients

Curvelet transform

1

58

58

2

158 LBP

descriptorsfrom

approximatecoefficients

58 lowast 16LBP

descriptorsfrom

1 to 16subband curvelet

coefficients

2

Figure 8 LBP code computation using sparse curvelet subband coeiexclcients

(a) (b) (c)

Figure 9 (a) FP reduction by clusters marked on original image (b) FP reduction by thresholding (c) FP reduction by sparse curveletcoeiexclcient-based LBP and ANN

Journal of Healthcare Engineering 7

depends upon the shape of the ROI as per the sparsematrix Tables 2 and 3 do not represent exact reduction inpixels for complete database but they exhibit pixel re-duction for sample mammograms

43 Classifier Evaluation and False-Positive ReductionFrom Figures 10ndash13 the best classification accuracy of 9857 has been obtained for MIAS in benign versus malignantclassification whereas 9870 for DDSM 9830 for

(1) Load input image (img1)(2) Apply CLAHE algorithm and obtain enhanced image (img2)(3) Decompose img1 and img2 up to 3 level of decomposition using Discrete Wavelet transform (DWT)(4) Use maximum local entropy rule for fusion of img1 and img2 for high frequency subbands(5) Take inverse DWT to obtain the fused image

ALGORITHM 1 Image fusion for contrast enhancement

Input I(m n) m no of rows and nno of columnOutput LBP featuresInitialize Radius R 1 and neighborhood pixels P 8

Mask [1 2 4 8 16 32 64 128 0]Sliding window coordinates kminus1 1Count 1 number of pixels in I(m n)

for i 1 to m dofor j 1 to n do

prepare local circular windowI_local I(i+ k j+ k)center_pixel I(i j)Arrange local neighborhoods of I(i j) pixels in a row col 1 9Lookup_table (i col) reshape(I_local [7 17])count number of pixels greater than zeroa length(find(Lookup_tablegt 0))select pixel position from lookup-table for computation of LBPif agt 2

LBP_code(count) I_localgt center_pixelcount count + 1

endend

endcompute histogram of LBP codesLBP_descriptor LBP_descriptorcountscale invariant

ALGORITHM 2 Algorithm for LBP feature computation based on shape of mass in ROI as

(1) Load input image (img1)(2) Apply CLAHE algorithm and obtain enhanced image (img2)(3) Process img1 and img2 and obtain enhanced image using procedure given in Algorithm 1(4) Remove pectoral muscle using proposed approach (Section 312)(5) Extract neighbourhood features for each pixel and apply SOM clustering(6) Obtain clustered image and separate out the tumorous cluster(7) Extract detected regions ie ROIrsquos from clustered result(8) Extract Sparse Curvelet Coefficients (Subband) up to 2 level from each ROI(9) Extract Sparse LBP code for each subband and obtain a combined feature vector for each ROI(10) Classify each ROI into tumorous and nontumorous class ie TP and FP respectively(11) Map each TP region on original mammogram (img1)(12) end

ALGORITHM 3 Summary of proposed method for FP reduction in mammograms

8 Journal of Healthcare Engineering

TMCH Scanner1 and 100 for TMCH Scanner2 clas-sication accuracies have been obtained in normal versusmalignant classication e classication performance ofANN has improved from 6 to 43 for diyenerent databasesas compared to KNN classier whereas there is littleimprovement about 7 compared with SVM classier eperformances of both proposed sparse LBP and LBPcomputation on curvelet subbands are nearly sametherefore the proposed algorithm can be eiexclciently

implemented in CAD system with lesser number of cur-velet coeiexclcients

Data augmentation has been used for some classes tomaintain balance between two classes to improve perfor-mance and to learn more powerful model Table 4 explainsthe FP reduction with the use of curvelet-based LBP featuresand ANN It has been observed that FP reduced from 085 to002 FPimage in MIAS 481 to 002 FPimage in DDSM and232 to 013 FPimage in TMCH

8000

8500

9000

9500

10000

TMCHS1LBP

TMCHS1sparse_LBP

TMCHS2LBP

TMCHS2sparse_LBP

Normal_Malignant

KNNANNSVM

Figure 10 Average classication rate for TMCH dataset

000

2000

4000

6000

8000

10000

DDSMLBP DDSMsparse_LBP

MIASLBP MIASsparse_LBP

Normal_Malignant

KNNANNSVM

Figure 11 Average classication rate for MIAS and DDSM dataset

000

2000

4000

6000

8000

10000

DDSMLBP DDSMsparse_LBP

MIASLBP MIASsparse_LBP

Benign_Malignant

KNNANNSVM

Figure 12 Average classication rate for MIAS and DDSM dataset

Journal of Healthcare Engineering 9

Similarly Table 5 shows the reduction in FPs as 085to 001 FPimage for MIAS 481 to 003 FPimage forDDSM and 232 to 000 FPimage for TMCH using sparsecurvelet coeiexclcient-based LBP features e results show

the eyenectiveness of sparse curvelet coeiexclcient-based LBPand ANN From Table 6 the best value of AUC 099 isobtained in benign versus malignant classication forMIAS AUC 098 in benign versus malignant in case of

Table 1 Result of SOM segmentation

Datasetused

Result of SOM clustering and threshold TPR (true-positiverate)

TPlesions

FPPI (false-positiveper image)FPimages

MassSegmented nonmass (FP) Total () images

Segmented (TP) LostMIAS 108 7 273 322 (108115) 094 (273322) 085DDSM 140 10 1203 250 (140150) 093 (1203250) 481TMCH 172 8 837 360 (172180) 095 (837360) 232

Table 2 Reduction in curvelet coeiexclcients for sample mammograms from MIAS and DDSM dataset

SrNo

MIAS DDSM

ROI Size

Total number ofcurvelet

coeiexclcients fromsubbands

Total number ofselected curveletcoeiexclcients from

subbands

reductionin curveletcoeiexclcients

ROI Size

Total number ofcurvelet

coeiexclcients fromsubbands

Total number ofselected curveletcoeiexclcients from

subbands

reductionin curveletcoeiexclcients

1 124times138 103911 77133 2577 192times187 216729 168333 22332 179times138 150123 126142 1597 294times 291 518267 366680 29253 51times 116 36815 33421 922 145times 207 181663 148765 18114 83times 83 42449 36653 1365 169times168 171873 154752 9965 84times 76 39115 35815 844 182times 248 272517 219822 19346 74times 83 37767 34448 879 213times 349 449783 301359 33007 53times 64 20969 18621 1120 578times 412 1433195 662072 53808 70times 44 18899 16610 1211 215times 219 286429 242935 15189 80times 66 32409 29454 912 420times 428 1082461 501209 537010 69times 86 36783 33552 878 203times 307 375871 266829 290111 59times116 42019 38442 851 226times 262 357209 276763 225212 81times 101 50637 46122 892 159times194 187563 148741 207013 41times 84 21427 18907 1176 718times 686 2961127 762561 742514 69times141 60641 52427 1354 409times 550 1352439 904829 331015 60times 62 22925 20475 1069 524times 375 1184671 766522 353016 96times101 59647 54235 907 311times 275 517433 397737 231317 136times139 113727 66352 4156 319times 320 614129 412606 328118 55times 94 31359 28305 974 313times 447 842855 510482 394319 157times140 132373 107000 1917 291times 517 907903 637412 297920 156times130 123007 90834 2615 370times 837 1864889 898236 5183

Average 58850 48247 14 Average 788950 437432 32

000

2000

4000

6000

8000

10000

DDSMLBP DDSM sparse_LBP

MIASLBP MIAS sparse_LBP

Normal_Abnormal

KNNANNSVM

Figure 13 Average classication rate for MIAS and DDSM dataset

10 Journal of Healthcare Engineering

DDSM AUC 094 in normal versus malignant in case ofTMCH Scanner1 and AUC 096 in normal versusmalignant classification in TMCH Scanner2 using ANNand curvelet subband-based LBP features e worstperformance of AUC 053 for MIAS is obtained with theproposed algorithm using KNN classifier as shown inTable 7 Similarly from Table 7 the best value ofAUC 098 is obtained in TMCH Scanner1 AUC 1 isobtained in TMCH Scanner2 database for normal versusmalignant classification AUC 098 in benign versusmalignant classification is attained in MIAS database andAUC 098 is achieved for normal versus malignant

classification in DDSM database using ANN classifier forsparse curvelet subband-based LBP features

However from Table 7 it should be noted that the per-formance of proposed algorithm is the best using ANN clas-sifier Figure 14 represents automated CAD system for breastcancer diagnosis with sample mammograms

Table 8 provides comparative study of methods developedfor breast tissue classification e proposed method providesbest results in terms of AUC and reduction of number of FPsas 085 to 001 FPimage for MIAS 481 to 003 FPimage forDDSM and 232 to 000 FPimage for TMCH e earlierreported work uses the fixed patch size-based approach which

Table 3 Reduction in curvelet coefficients for sample mammograms from TMCH Scanner1 and Scanner2 dataset

Srno

TMCH Scanner 1 ldquoGE Medical Senograph Systemrdquo TMCH Scanner 2 ldquoHologic Selenia Systemrdquo

ROI size

Total number ofcurvelet

coefficients fromsubbands

Total number ofselected curveletcoefficients from

subbands

reductionin curveletcoefficients

ROI size

Total number ofcurvelet

coefficients fromsubbands

Total number ofselected curveletcoefficients from

subbands

reductionin curveletcoefficients

1 459times 412 1140617 794885 3031 291times 278 489243 317014 35202 548times 513 1693403 1117873 3399 545times 246 809627 531604 34343 415times 303 758323 445585 4124 560times 483 1630583 1068439 34474 645times 495 1951443 1145580 4129 782times 510 2400137 1275073 46875 437times 691 1816651 985120 4577 87times141 75773 67871 10436 812times 500 2439937 1287065 4725 311times 185 348565 240546 30997 468times 379 1066333 710242 3340 262times 348 550303 282821 48618 673times 582 2355589 1730915 2652 610times 440 1614515 842876 47799 250times 201 304513 235670 2261 949times 391 2227209 1359338 389710 525times 488 1544691 1161942 2478 365times 385 846523 604473 285911 488times 779 2287547 1611385 2956 393times 247 585063 490474 161712 1434times 966 8326581 3742277 5506 341times 301 618111 410542 335813 348times 421 881701 616501 3008 523times 702 2206097 800057 637314 460times 530 1467227 799885 4548 370times 284 632539 423727 330115 398times 450 1078441 828064 2322 344times 202 418955 302427 278116 247times 272 404401 344822 1473 264times188 299997 231926 226917 411times 305 757657 405919 4642 233times 247 346983 267686 228518 286times 344 593155 448566 2438 370times 291 650701 486125 252919 417times 207 523477 429755 1790 680times 483 1979543 1052793 468220 463times 458 1275021 857608 3274 202times 266 324295 215042 3369

Average 1633335 984983 33 Average 952738 563543 34

Table 4 Number of ROIs resulted in FP reduction using curvelet-based LBP (without sparse) amp ANN classification at training andvalidation stage

Class Datasetused

Benignmalignant mass Nonmassbenign mass Total()

images

TPR (true-positiverate)TPlesions

FPPI (false-positiveper image) FP

imagesPreviousstage

Selected(TP)

Lost(FN)

Previousstage

Selected(TN)

Lost(FP)

Normal vsabnormal

MIAS 108lowast 2 216 203 13 273 257 16 315 (203216) 094 (16315) 005DDSM 140lowast 4 560 465 95 1203 1095 108 240 (465560) 083 (108240) 045

Benign vsmalignant

MIAS 49 49 0 59 57 2 108 (4949) 100 (2108) 002DDSM 46lowast 2 92 91 1 94 91 3 140 (9192) 099 (3140) 002

Normal vsmalignant

MIAS 49lowast 4196 184 12 273 254 19 256 (184196) 094 (19256) 007DDSM 46lowast 4184 180 4 1203 1143 60 146 (180184) 098 (60146) 041TMCHScanner1 107lowast 4 428 416 12 605 551 54 217 (416428) 097 (54217) 025

TMCHScanner2 65lowast 4 260 255 5 232 214 18 135 (255260) 098 (18135) 013

lowastAugmentation of image

Journal of Healthcare Engineering 11

limits the automatic CAD system scope whereas proposedsystem provides complete solution to CAD system right fromautomatic tumor patch segmentation to reduction in FPs andfinal representation of mammogram with TP marked on itIt will drastically reduce the radiologist work by locationtumor directly on mammogram

5 Conclusion

A fully automatic CAD system which can accuratelylocate the tumor on a mammogram and reduces FPshas been proposed e developed CAD system consistsof preprocessing SOM clustering ROI extraction

Table 6 Performance evaluation of curvelet-based LBP descriptor algorithm

Dataset Classification Normal-malignant Normal-abnormal Benign-malignantClassifier Sensitivity Specificity AUC Sensitivity Specificity AUC Sensitivity Specificity AUC

MIASANN 094 093 094 094 094 094 100 097 099SVM 085 085 085 083 086 085 088 084 086KNN 067 063 065 058 057 058 062 068 063

DDSMANN 098 095 095 083 091 085 099 097 098SVM 097 088 092 071 091 083 094 089 092KNN 096 064 087 067 090 080 087 073 079

TMCH Scanner1ANN 097 091 094 mdash mdash mdash mdash mdash mdashSVM 096 091 094 mdash mdash mdash mdash mdash mdashKNN 098 082 089 mdash mdash mdash mdash mdash mdash

TMCH Scanner2ANN 098 092 096 mdash mdash mdash mdash mdash mdashSVM 097 090 094 mdash mdash mdash mdash mdash mdashKNN 092 083 088 mdash mdash mdash mdash mdash mdash

Table 7 Performance evaluation of proposed algorithm

Dataset Classification Normal-malignant Normal-abnormal Benign-malignantClassifier Sensitivity Specificity AUC Sensitivity Specificity AUC Sensitivity Specificity AUC

MIASANN 098 095 096 093 097 095 097 100 098SVM 088 083 085 085 082 084 084 092 087KNN 055 051 053 055 063 056 061 067 061

DDSMANN 099 097 098 092 096 093 097 095 096SVM 099 092 096 089 096 092 094 092 093KNN 098 073 092 074 090 082 089 077 083

TMCH Scanner1ANN 099 098 098 mdash mdash mdash mdash mdash mdashSVM 098 096 097 mdash mdash mdash mdash mdash mdashKNN 099 092 095 mdash mdash mdash mdash mdash mdash

TMCH Scanner2ANN 100 100 100 mdash mdash mdash mdash mdash mdashSVM 100 098 099 mdash mdash mdash mdash mdash mdashKNN 096 092 094 mdash mdash mdash mdash mdash mdash

Table 5 Number of ROIs resulted in FP reduction using sparse curvelet coefficient-based LBP amp ANN classification at training andvalidation stage

Class Datasetused

Benignmalignant mass Nonmassbenign massTotal ()images

TPR (true-positiverate)TPlesions

FPPI (false-positiveper image) FP

imagesPreviousstage

Selected(TP)

Lost(FN)

Previousstage

Selected(TN)

Lost(FP)

Normal vsabnormal

MIAS 108lowast 2 216 201 15 273 265 8 315 (201216) 093 (8315) 002DDSM 140lowast 4 560 516 44 1203 1155 48 240 (516560) 092 (48240) 02

Benign vsmalignant

MIAS 49 48 1 59 59 1 108 (4849) 098 (1108) 001DDSM 46lowast 2 92 89 3 94 89 5 140 (8992) 097 (5140) 003

Normal vsmalignant

MIAS 49lowast 4196 192 4 273 259 14 256 (192196) 098 (14256) 005DDSM 46lowast 4184 182 2 1203 1167 36 146 (182184) 099 (36146) 025TMCHScanner1 107lowast 4 428 424 4 605 593 12 217 (424428) 099 (12217) 005

TMCHScanner2 65lowast 4 260 260 0 232 232 0 135 (260260) 100 (0135) 0

lowastAugmentation of image

12 Journal of Healthcare Engineering

sparse LBP feature computation based on sparse Cur-velet coefficients and finally FP reduction using ANNclassifier

e proposed algorithm presents a novel concept ofextraction of curvelet coefficients according to irregularshape of mass is called as sparse curvelet coefficients andcomputation of LBP e analysis proves that the FPs arereduced significantly from 085 to 001 FPimage for MIAS

481 to 003 FPimage for DDSM and 232 to 000 FPimagefor TMCH e ANN classifier showed best results asAUC 098 and accuracy 9857 for MIAS in benign-malignant classification AUC 098 and accuracy 9870for DDSM in normal-malignant classification AUC 098and accuracy 9830 for TMCH Scanner1 and AUC 1and accuracy 100 for TMCH Scanner2 in normal-malignant classification as compared with SVM and KNN

(a) (b) (c) (d) (e) (f )

Figure 14 Representation of fully automatic CAD system for breast cancer using (a) samplemammograms fromMIAS DDSM and TMCHdatasets (b) preprocessed mammograms (c) clustered image (d) TP and FP marked on mammogram (e) TP marked by thresholding (f )TP marked by using LBP descriptor based on sparse curvelet coefficients

Journal of Healthcare Engineering 13

classifier e performance of LBP features and LBP featuresbased on sparse curvelet coefficients are nearly same whichshow that the proposed algorithm is suitable for cancer breasttissue diagnosis

In future the reduced curvelet coefficients can be used toextract local ternary patterns and other local descriptor andlocal directional patterns etc e present work deals withmammogram with single mass this can be further extendedfor multiple mass models with multiple LBP features basedon sparse curvelet coefficients

Data Availability

In this research we have used two publicly available datasetsMIAS and DDSM ese datasets can be found here in [17]and [19] e third database is collected from the localhospital Tata Memorial Cancer Hospital Mumbai whichcan be found at httpeurekasveriacin or available fromthe corresponding author upon request

Conflicts of Interest

e authors declare that they have no conflicts of interest

Acknowledgments

e TMCH database for this work was given by Departmentof Radiodiagnosis Tata Memorial Cancer Hospital Mumbai

References

[1] S Malvia S A Bagadi U S Dubey and S Saxena ldquoEpi-demiology of breast cancer in Indian womenrdquo Asia-PacificJournal of Clinical Oncology vol 13 no 4 pp 289ndash295 2017

[2] A Gupta K Shridhar and P Dhillon ldquoA review of breastcancer awareness among women in India cancer literate orawareness deficitrdquo European Journal of Cancer vol 51no 14 pp 2058ndash2066 2015

[3] K P Kanadam and S R Chereddy ldquoMammogram classifi-cation using sparse-ROI a novel representation to arbitraryshaped massesrdquo Expert Systems with Applications vol 57pp 204ndash213 2016

[4] D O T Bruno M Z do Nascimento R P Ramos et al ldquoLBPoperators on curvelet coefficients as an algorithm to describetexture in breast cancer tissuesrdquo Expert Systems with Appli-cations vol 55 pp 329ndash340 2016

[5] M Hussain ldquoFalse-positive reduction in mammographyusing multiscale spatial Weber law descriptor and supportvector machinesrdquo Neural Computing and Applicationsvol 25 no 1 pp 83ndash93 2014

[6] M M Pawar and S N Talbar ldquoGenetic fuzzy system (GFS)based wavelet co-occurrence feature selection in mammo-gram classification for breast cancer diagnosisrdquo Perspectives inScience vol 8 pp 247ndash250 2016

[7] C Muramatsu T Hara T Endo and H Fujita ldquoBreast massclassification on mammograms using radial local ternarypatternsrdquo Computers in Biology and Medicine vol 72pp 43ndash53 2016

[8] M M Eltoukhy I Faye and B B Samir ldquoA statistical basedfeature extraction method for breast cancer diagnosis indigital mammogram using multiresolution representationrdquo

Table 8 Comparison of classification accuracy AUC and FPimage values from different approaches in breast cancer diagnosis

Author Database Method Classifier Result AUC FPimageEltoukhy et al [33]

MIAS Biggest curvelet coefficients as a featurevector

Euclidean classifier 9407 mdash mdashEltoukhy et al [42] 9859 mdash mdashEltoukhy et al [8] SVM 973 mdash mdash

Dhahbi et al [34] Mini-MIAS Curvelet moments KNN 9127 mdash mdashDDSM 8646 mdash mdash

Bruno et al [4] DDSM Curvelet + LBP SVM 85 085 mdashPL 94 094 mdash

da Rocha et al [40] DDSM LBP SVM 8831 088 mdashKanadam andChereddy [3] MIAS Sparse ROI SVM 9742 mdash mdash

Pereira et al [18] DDSM Wavelet and Wiener filter Multiple thresholding waveletand GA mdash mdash 137

Liu and Zeng [29] DDSMFFDM GLCM CLBP and geometric features SVM mdash mdash 148

De Sampaio et al[39] DDSM LBP DBSCAN 9826 019

Zyout et al [30] DDSM Second order statistics of waveletcoefficients (SOSWC) SVM 968 097 0018

MIAS 952 966 0029

Casti et al [31]DDSM

Differential features Fisher linear discriminantanalysis (FLDA) mdash mdash

168MIAS 212FFDM 082

Proposed method

MIAS

LBP based on sparse curvelet subbandcoefficients ANN

9857 098 001DDSM 9870 098 003TMCHScanner1 9830 098 005

TMCHScanner2 100 1 0

14 Journal of Healthcare Engineering

it takes pixels related to shape of mass which are called asforeground pixels and rejects the other pixels called asbackground pixels e proposed algorithm uses foregroundpixels only for LBP computation and this will tend to numberof pixel reduction in LBP computations erefore identi-cation of foreground and background pixels is an importantstep which is performed using lookup table approach e

identication of foreground and background pixel is based onnumber of nonzero pixels in the lookup table ie if count ofsliding window nonzero pixels is greater than 2 count(p(i j))gt 2 is identied as foreground and LBP is estimated On theother hand if count of sliding window nonzero pixels is lessthan 2 count(p(i j))lt 2 is identied as background and LBPwould not be estimated and rejected from lookup table

(a) (b) (c)

Figure 4 FP reduction by thresholding (a) clustered image (b) clusters boundaries marked on original image and (c) clusters afterthresholding

MIAS ROI patches from abnormal mammogram

DDSM ROI patches from abnormal mammogram

TMCH ldquoGE Medical Senograph Systemrdquo ROIpatches from abnormal mammogram

TMCH ldquoHologic Selenia Systemrdquo ROI patchesfrom abnormal mammogram

Figure 5 Variable sizes ROIs from MIAS DDSM and TMCH datasets

Input matrixLook-up table LBP code computation

Count

Count=number of (pixel Intensitiesgt0)

Count (pixel intensitiesgt0)is not greater than 2

LBP code27 26 25 24 23 22 21 20

If Count (pixel intensitygt0)gt2calculate LBP code

ElseDiscard

Discard row

Similarly other pixels also compared in look up table Similarly calculate LBP code for other pixel if countgt2Centre pixel

Centre pixel

Shape of mass

g4

g4

g3

g3

g2

g2

3times3 sliding window

g5

g5

gc

gc

63 90 23

2

33

0 0 0 0 100 123 490 10 0 23 0 63 100 123 4 610 11 0 0 23 90 123 4 2 611 0 0 0 0

0 0 0 0 0 0 0 0 0

0 0 0 0 0

0 0 0 0 0 0 00 0 0 0 0 0 000 0 0 0 0 0 0 0

0 0 00 0 00

0 0 0

0 0 000 1

1 1 11

1

1

1 1 11 1

1

1 1 112

131

154

128

0 0 00

0 0 0 00 00 0 0

000

0 0

10 4 2 134 4100 123 90 63 0 0 0 23 64 5123 4 10 90 63 100 23

2323

64

646564

5

55

5

6

6

6

g1

g1

g6

g6

g7

g7

g8

g8

0 0 23 0 0 00 63 90 10 11 00 100 123 4 2 1340 23 64 5 6 00 0 65 0 0 00 0 0 0 0 0

Figure 6 Lookup table approach for LBP computation from shape of mass in ROI

Journal of Healthcare Engineering 5

Nonzero pixels provide actual shape of mass and are taken forLBP computations Graphical representation of proposedalgorithm for LBP descriptor computation using foregroundpixels has been given in Figure 7 and the algorithm has beendescribed in Algorithm 2

342 e Fast Discrete Curvelet Transform (FDCT) eauthors [44] have introduced computationally simple andeiexclcient Fast Discrete Curvelet Transform (FDCT) We havepreferred wrapping-based FDCT approach in proposedwork as it is faster e curvelet coeiexclcients CD(j l k)represented by scale j angle l and spatial location k can bewritten as

CD j l k1 k2( ) sumn1N1

n11sumn2N2

n21I n1 n2[ ]φD

jlk1k2n1 n2[ ] (8)

Figure 8 illustrates LBP code computation based on sparsecurvelet coeiexclcients ROI decomposes using curvelet trans-form with scale orientations l of 16deg and scale of 2 as thedatabase consists of minimum ROI size of 25 times 22 pixelsCurvelet transform with scale orientations l of 16deg and scale of2 produces 1 + 16 17 diyenerent subbands based on subbanddivision Further each curvelet subband coeiexclcients havebeen represented using lookup table using 3 times 3 slidingwindow and if the row in the lookup table identies fore-ground coeiexclcient then LBP is computed with radius R 1and P 8 neighboring pixels as shown in Algorithm 2 total 58LBP features have been obtained from foreground curveletsubband coeiexclcients erefore total 986 LBP features havebeen extracted from 17 curvelet subbands It can be observedfrom Figure 8 curvelet subbands also provide shape of massin 16 diyenerent directions so that the directional informationcan be associated with LBP features Kanadam et al [3] usedconcept of sparse ROI similarly we have extended it forsparse curvelet subband and LBP features computation

35 Classi cation In this work we have analyzed extractedROI from mammogram using normal-abnormal benign-malignant and normal-malignant classes with ANN SVMand KNN classiers e detailed description of ANNclassier has been given in [45 46] To evaluate performance

of the proposed system we have used 3-fold cross validationwhere database is randomly divided into three sets andaccuracy is calculated for each set e nal accuracy of thesystem is average of accuracy of each of three sets Howeverit will not be fair to compare 3-fold cross validation result ofSVM and KNN classier with ANN because ANN classieris tested on only one set of images (33 for training 33 fortesting and 33 for validation)us to do fair comparisonwe have trained ANN using input layer (986 neuron) overthree diyenerent sets (which are considered in SVM and KNN)and calculated its average accuracy Our proposed falsepositive reduction algorithm illustrates in Figures 9(a)ndash9(c)Algorithm 3 summarizes copyow of the proposed method forFP reduction in mammograms

4 Experimental Results and Discussions

e proposed method has been tested and validated usingthree classiers and three clinical mammographic imagedatasets

41 Data Sets

411 Mammographic Image Analysis Society (MIAS)Database e mini-MIAS [17] database consists of 322mammograms each having 1024times1024 pixels and anno-tated like background tissue character class severity centerof abnormality and radius of circle for abnormality isdatabase includes 64 benign 51 malignant and 207 normalcases which have been taken for experimentation

412 Digital Database for Screening Mammography(DDSM) e DDSM [19] dataset consists of 2500 studiesand is composed of cranial-caudal (CC) and mediolateral-oblique (MLO) views of mammographic image for left andright breast annotated with ACR breast density type ofabnormality and ground truth Randomly selected 150abnormal and 100 normal cases from both HOWTEK andLUMISYS scanner of 12 bits per pixel resolution have beensubjected for experimentation

ROI Shape of mass

Mass shape by sparse matrix

Position inlook-up table Nine neighborhood pixels Decision making

Yes

LBP operator on 3times3 neighborhood

Nine neighborhood pixels

120 125 210

105 45

254

10

93 200 250

1 1 1

1 0

1 1 1

Thresholding

LBP code

LBP code =0 + 2 + 4 +8 + 16 + 32+ 64 + 128

No

Pixel count(position(ij))

gt 2

Discard the rowfrom look-up table

eg X P etc

Input

bullbull

bullbull

bullbull

bullbull

bull

X 0 0 0 0 0 0 0 0 1

P 0 0 0 0 0 0 0 0 0

Q 0 0 0 37 200 0 100 102 0

Y 123 90 70 100 23 45 0 144 17

Z 120 105 93 125 45 200 210 10 250

Q position in look-up table

P position

Z position

Y position

X position

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

Figure 7 Process for computation of LBP descriptor from shape of mass in ROI (a) Original image (b) 3times 3 window for selection offoreground pixels (c) lookup table (d) decision making process (e) LBP computation from selected foreground pixels

6 Journal of Healthcare Engineering

413 e Tata Memorial Cancer Hospital (TMCH) isdataset [47] contains 360 full-eld digital mammograms(FFDMs) comprising 180 CC views and 180 MLO viewsfrom right and left breast acquired from 90 randomly se-lected patients It is composed of 180 veried malignant and180 normal breast images It uses biopsy proven breastcancer patientsrsquo pathological data approved by the In-stitutional Research Ethics Committee of Tata MemorialCentre Hospital (TMCH) Mumbai India e ground truthmarking on each abnormal mammogram is performedmanually using the Histopathological Reports (HPR) of therespective patients and expert radiologist from TMCHMumbai Approximately 35 patients are examined usingldquoHologic Selenia Systemrdquo (Scanner1) gives 16-bit

e remaining 55 patients were examined with ldquoGEMedical Senograph Systemrdquo (Scanner2) providing 8-bit truecolor mammogram image in DICOM format of 4096times 3328or 2294times1914 pixels each measuring size 50times 50 μm2

42 Segmentation Evaluation and ROI Extraction e seg-mentation using SOM that detects suspicious mass regions isconsidered as TP whereas from nonmass is taken as FPFrom Table 1 it is clear that total suspicious ROI (includingTP amp FP) of 381 for MIAS 1343 for DDSM and 1009 forTMCH have been taken for evaluation our proposed al-gorithm for FP reduction

From extracted ROIs the minimum patch size is25 times 22 pixels whereas the maximum size is 1152 times1356pixels Tables 2 and 3 represent curvelet subband co-eiexclcients from 17 subbands and reduced coeiexclcientsbased on lookup table approach are used to calculate LBPfeatures It has been observed during experimentationthat the curvelet coeiexclcients on an average are reducedfor sparse LBP by 14 32 33 and 34 for MIASDDSM TMCH Scanner1 and TMCH Scanner2 re-spectively It may be noticed that reduction in curveletcoeiexclcients for every ROI is not xed It completely

bullbullbull

bullbullbull

bullbullbull

Scale 1 approximatecoefficients

Scale 2 1 to 16subband coefficients

Curvelet transform

1

58

58

2

158 LBP

descriptorsfrom

approximatecoefficients

58 lowast 16LBP

descriptorsfrom

1 to 16subband curvelet

coefficients

2

Figure 8 LBP code computation using sparse curvelet subband coeiexclcients

(a) (b) (c)

Figure 9 (a) FP reduction by clusters marked on original image (b) FP reduction by thresholding (c) FP reduction by sparse curveletcoeiexclcient-based LBP and ANN

Journal of Healthcare Engineering 7

depends upon the shape of the ROI as per the sparsematrix Tables 2 and 3 do not represent exact reduction inpixels for complete database but they exhibit pixel re-duction for sample mammograms

43 Classifier Evaluation and False-Positive ReductionFrom Figures 10ndash13 the best classification accuracy of 9857 has been obtained for MIAS in benign versus malignantclassification whereas 9870 for DDSM 9830 for

(1) Load input image (img1)(2) Apply CLAHE algorithm and obtain enhanced image (img2)(3) Decompose img1 and img2 up to 3 level of decomposition using Discrete Wavelet transform (DWT)(4) Use maximum local entropy rule for fusion of img1 and img2 for high frequency subbands(5) Take inverse DWT to obtain the fused image

ALGORITHM 1 Image fusion for contrast enhancement

Input I(m n) m no of rows and nno of columnOutput LBP featuresInitialize Radius R 1 and neighborhood pixels P 8

Mask [1 2 4 8 16 32 64 128 0]Sliding window coordinates kminus1 1Count 1 number of pixels in I(m n)

for i 1 to m dofor j 1 to n do

prepare local circular windowI_local I(i+ k j+ k)center_pixel I(i j)Arrange local neighborhoods of I(i j) pixels in a row col 1 9Lookup_table (i col) reshape(I_local [7 17])count number of pixels greater than zeroa length(find(Lookup_tablegt 0))select pixel position from lookup-table for computation of LBPif agt 2

LBP_code(count) I_localgt center_pixelcount count + 1

endend

endcompute histogram of LBP codesLBP_descriptor LBP_descriptorcountscale invariant

ALGORITHM 2 Algorithm for LBP feature computation based on shape of mass in ROI as

(1) Load input image (img1)(2) Apply CLAHE algorithm and obtain enhanced image (img2)(3) Process img1 and img2 and obtain enhanced image using procedure given in Algorithm 1(4) Remove pectoral muscle using proposed approach (Section 312)(5) Extract neighbourhood features for each pixel and apply SOM clustering(6) Obtain clustered image and separate out the tumorous cluster(7) Extract detected regions ie ROIrsquos from clustered result(8) Extract Sparse Curvelet Coefficients (Subband) up to 2 level from each ROI(9) Extract Sparse LBP code for each subband and obtain a combined feature vector for each ROI(10) Classify each ROI into tumorous and nontumorous class ie TP and FP respectively(11) Map each TP region on original mammogram (img1)(12) end

ALGORITHM 3 Summary of proposed method for FP reduction in mammograms

8 Journal of Healthcare Engineering

TMCH Scanner1 and 100 for TMCH Scanner2 clas-sication accuracies have been obtained in normal versusmalignant classication e classication performance ofANN has improved from 6 to 43 for diyenerent databasesas compared to KNN classier whereas there is littleimprovement about 7 compared with SVM classier eperformances of both proposed sparse LBP and LBPcomputation on curvelet subbands are nearly sametherefore the proposed algorithm can be eiexclciently

implemented in CAD system with lesser number of cur-velet coeiexclcients

Data augmentation has been used for some classes tomaintain balance between two classes to improve perfor-mance and to learn more powerful model Table 4 explainsthe FP reduction with the use of curvelet-based LBP featuresand ANN It has been observed that FP reduced from 085 to002 FPimage in MIAS 481 to 002 FPimage in DDSM and232 to 013 FPimage in TMCH

8000

8500

9000

9500

10000

TMCHS1LBP

TMCHS1sparse_LBP

TMCHS2LBP

TMCHS2sparse_LBP

Normal_Malignant

KNNANNSVM

Figure 10 Average classication rate for TMCH dataset

000

2000

4000

6000

8000

10000

DDSMLBP DDSMsparse_LBP

MIASLBP MIASsparse_LBP

Normal_Malignant

KNNANNSVM

Figure 11 Average classication rate for MIAS and DDSM dataset

000

2000

4000

6000

8000

10000

DDSMLBP DDSMsparse_LBP

MIASLBP MIASsparse_LBP

Benign_Malignant

KNNANNSVM

Figure 12 Average classication rate for MIAS and DDSM dataset

Journal of Healthcare Engineering 9

Similarly Table 5 shows the reduction in FPs as 085to 001 FPimage for MIAS 481 to 003 FPimage forDDSM and 232 to 000 FPimage for TMCH using sparsecurvelet coeiexclcient-based LBP features e results show

the eyenectiveness of sparse curvelet coeiexclcient-based LBPand ANN From Table 6 the best value of AUC 099 isobtained in benign versus malignant classication forMIAS AUC 098 in benign versus malignant in case of

Table 1 Result of SOM segmentation

Datasetused

Result of SOM clustering and threshold TPR (true-positiverate)

TPlesions

FPPI (false-positiveper image)FPimages

MassSegmented nonmass (FP) Total () images

Segmented (TP) LostMIAS 108 7 273 322 (108115) 094 (273322) 085DDSM 140 10 1203 250 (140150) 093 (1203250) 481TMCH 172 8 837 360 (172180) 095 (837360) 232

Table 2 Reduction in curvelet coeiexclcients for sample mammograms from MIAS and DDSM dataset

SrNo

MIAS DDSM

ROI Size

Total number ofcurvelet

coeiexclcients fromsubbands

Total number ofselected curveletcoeiexclcients from

subbands

reductionin curveletcoeiexclcients

ROI Size

Total number ofcurvelet

coeiexclcients fromsubbands

Total number ofselected curveletcoeiexclcients from

subbands

reductionin curveletcoeiexclcients

1 124times138 103911 77133 2577 192times187 216729 168333 22332 179times138 150123 126142 1597 294times 291 518267 366680 29253 51times 116 36815 33421 922 145times 207 181663 148765 18114 83times 83 42449 36653 1365 169times168 171873 154752 9965 84times 76 39115 35815 844 182times 248 272517 219822 19346 74times 83 37767 34448 879 213times 349 449783 301359 33007 53times 64 20969 18621 1120 578times 412 1433195 662072 53808 70times 44 18899 16610 1211 215times 219 286429 242935 15189 80times 66 32409 29454 912 420times 428 1082461 501209 537010 69times 86 36783 33552 878 203times 307 375871 266829 290111 59times116 42019 38442 851 226times 262 357209 276763 225212 81times 101 50637 46122 892 159times194 187563 148741 207013 41times 84 21427 18907 1176 718times 686 2961127 762561 742514 69times141 60641 52427 1354 409times 550 1352439 904829 331015 60times 62 22925 20475 1069 524times 375 1184671 766522 353016 96times101 59647 54235 907 311times 275 517433 397737 231317 136times139 113727 66352 4156 319times 320 614129 412606 328118 55times 94 31359 28305 974 313times 447 842855 510482 394319 157times140 132373 107000 1917 291times 517 907903 637412 297920 156times130 123007 90834 2615 370times 837 1864889 898236 5183

Average 58850 48247 14 Average 788950 437432 32

000

2000

4000

6000

8000

10000

DDSMLBP DDSM sparse_LBP

MIASLBP MIAS sparse_LBP

Normal_Abnormal

KNNANNSVM

Figure 13 Average classication rate for MIAS and DDSM dataset

10 Journal of Healthcare Engineering

DDSM AUC 094 in normal versus malignant in case ofTMCH Scanner1 and AUC 096 in normal versusmalignant classification in TMCH Scanner2 using ANNand curvelet subband-based LBP features e worstperformance of AUC 053 for MIAS is obtained with theproposed algorithm using KNN classifier as shown inTable 7 Similarly from Table 7 the best value ofAUC 098 is obtained in TMCH Scanner1 AUC 1 isobtained in TMCH Scanner2 database for normal versusmalignant classification AUC 098 in benign versusmalignant classification is attained in MIAS database andAUC 098 is achieved for normal versus malignant

classification in DDSM database using ANN classifier forsparse curvelet subband-based LBP features

However from Table 7 it should be noted that the per-formance of proposed algorithm is the best using ANN clas-sifier Figure 14 represents automated CAD system for breastcancer diagnosis with sample mammograms

Table 8 provides comparative study of methods developedfor breast tissue classification e proposed method providesbest results in terms of AUC and reduction of number of FPsas 085 to 001 FPimage for MIAS 481 to 003 FPimage forDDSM and 232 to 000 FPimage for TMCH e earlierreported work uses the fixed patch size-based approach which

Table 3 Reduction in curvelet coefficients for sample mammograms from TMCH Scanner1 and Scanner2 dataset

Srno

TMCH Scanner 1 ldquoGE Medical Senograph Systemrdquo TMCH Scanner 2 ldquoHologic Selenia Systemrdquo

ROI size

Total number ofcurvelet

coefficients fromsubbands

Total number ofselected curveletcoefficients from

subbands

reductionin curveletcoefficients

ROI size

Total number ofcurvelet

coefficients fromsubbands

Total number ofselected curveletcoefficients from

subbands

reductionin curveletcoefficients

1 459times 412 1140617 794885 3031 291times 278 489243 317014 35202 548times 513 1693403 1117873 3399 545times 246 809627 531604 34343 415times 303 758323 445585 4124 560times 483 1630583 1068439 34474 645times 495 1951443 1145580 4129 782times 510 2400137 1275073 46875 437times 691 1816651 985120 4577 87times141 75773 67871 10436 812times 500 2439937 1287065 4725 311times 185 348565 240546 30997 468times 379 1066333 710242 3340 262times 348 550303 282821 48618 673times 582 2355589 1730915 2652 610times 440 1614515 842876 47799 250times 201 304513 235670 2261 949times 391 2227209 1359338 389710 525times 488 1544691 1161942 2478 365times 385 846523 604473 285911 488times 779 2287547 1611385 2956 393times 247 585063 490474 161712 1434times 966 8326581 3742277 5506 341times 301 618111 410542 335813 348times 421 881701 616501 3008 523times 702 2206097 800057 637314 460times 530 1467227 799885 4548 370times 284 632539 423727 330115 398times 450 1078441 828064 2322 344times 202 418955 302427 278116 247times 272 404401 344822 1473 264times188 299997 231926 226917 411times 305 757657 405919 4642 233times 247 346983 267686 228518 286times 344 593155 448566 2438 370times 291 650701 486125 252919 417times 207 523477 429755 1790 680times 483 1979543 1052793 468220 463times 458 1275021 857608 3274 202times 266 324295 215042 3369

Average 1633335 984983 33 Average 952738 563543 34

Table 4 Number of ROIs resulted in FP reduction using curvelet-based LBP (without sparse) amp ANN classification at training andvalidation stage

Class Datasetused

Benignmalignant mass Nonmassbenign mass Total()

images

TPR (true-positiverate)TPlesions

FPPI (false-positiveper image) FP

imagesPreviousstage

Selected(TP)

Lost(FN)

Previousstage

Selected(TN)

Lost(FP)

Normal vsabnormal

MIAS 108lowast 2 216 203 13 273 257 16 315 (203216) 094 (16315) 005DDSM 140lowast 4 560 465 95 1203 1095 108 240 (465560) 083 (108240) 045

Benign vsmalignant

MIAS 49 49 0 59 57 2 108 (4949) 100 (2108) 002DDSM 46lowast 2 92 91 1 94 91 3 140 (9192) 099 (3140) 002

Normal vsmalignant

MIAS 49lowast 4196 184 12 273 254 19 256 (184196) 094 (19256) 007DDSM 46lowast 4184 180 4 1203 1143 60 146 (180184) 098 (60146) 041TMCHScanner1 107lowast 4 428 416 12 605 551 54 217 (416428) 097 (54217) 025

TMCHScanner2 65lowast 4 260 255 5 232 214 18 135 (255260) 098 (18135) 013

lowastAugmentation of image

Journal of Healthcare Engineering 11

limits the automatic CAD system scope whereas proposedsystem provides complete solution to CAD system right fromautomatic tumor patch segmentation to reduction in FPs andfinal representation of mammogram with TP marked on itIt will drastically reduce the radiologist work by locationtumor directly on mammogram

5 Conclusion

A fully automatic CAD system which can accuratelylocate the tumor on a mammogram and reduces FPshas been proposed e developed CAD system consistsof preprocessing SOM clustering ROI extraction

Table 6 Performance evaluation of curvelet-based LBP descriptor algorithm

Dataset Classification Normal-malignant Normal-abnormal Benign-malignantClassifier Sensitivity Specificity AUC Sensitivity Specificity AUC Sensitivity Specificity AUC

MIASANN 094 093 094 094 094 094 100 097 099SVM 085 085 085 083 086 085 088 084 086KNN 067 063 065 058 057 058 062 068 063

DDSMANN 098 095 095 083 091 085 099 097 098SVM 097 088 092 071 091 083 094 089 092KNN 096 064 087 067 090 080 087 073 079

TMCH Scanner1ANN 097 091 094 mdash mdash mdash mdash mdash mdashSVM 096 091 094 mdash mdash mdash mdash mdash mdashKNN 098 082 089 mdash mdash mdash mdash mdash mdash

TMCH Scanner2ANN 098 092 096 mdash mdash mdash mdash mdash mdashSVM 097 090 094 mdash mdash mdash mdash mdash mdashKNN 092 083 088 mdash mdash mdash mdash mdash mdash

Table 7 Performance evaluation of proposed algorithm

Dataset Classification Normal-malignant Normal-abnormal Benign-malignantClassifier Sensitivity Specificity AUC Sensitivity Specificity AUC Sensitivity Specificity AUC

MIASANN 098 095 096 093 097 095 097 100 098SVM 088 083 085 085 082 084 084 092 087KNN 055 051 053 055 063 056 061 067 061

DDSMANN 099 097 098 092 096 093 097 095 096SVM 099 092 096 089 096 092 094 092 093KNN 098 073 092 074 090 082 089 077 083

TMCH Scanner1ANN 099 098 098 mdash mdash mdash mdash mdash mdashSVM 098 096 097 mdash mdash mdash mdash mdash mdashKNN 099 092 095 mdash mdash mdash mdash mdash mdash

TMCH Scanner2ANN 100 100 100 mdash mdash mdash mdash mdash mdashSVM 100 098 099 mdash mdash mdash mdash mdash mdashKNN 096 092 094 mdash mdash mdash mdash mdash mdash

Table 5 Number of ROIs resulted in FP reduction using sparse curvelet coefficient-based LBP amp ANN classification at training andvalidation stage

Class Datasetused

Benignmalignant mass Nonmassbenign massTotal ()images

TPR (true-positiverate)TPlesions

FPPI (false-positiveper image) FP

imagesPreviousstage

Selected(TP)

Lost(FN)

Previousstage

Selected(TN)

Lost(FP)

Normal vsabnormal

MIAS 108lowast 2 216 201 15 273 265 8 315 (201216) 093 (8315) 002DDSM 140lowast 4 560 516 44 1203 1155 48 240 (516560) 092 (48240) 02

Benign vsmalignant

MIAS 49 48 1 59 59 1 108 (4849) 098 (1108) 001DDSM 46lowast 2 92 89 3 94 89 5 140 (8992) 097 (5140) 003

Normal vsmalignant

MIAS 49lowast 4196 192 4 273 259 14 256 (192196) 098 (14256) 005DDSM 46lowast 4184 182 2 1203 1167 36 146 (182184) 099 (36146) 025TMCHScanner1 107lowast 4 428 424 4 605 593 12 217 (424428) 099 (12217) 005

TMCHScanner2 65lowast 4 260 260 0 232 232 0 135 (260260) 100 (0135) 0

lowastAugmentation of image

12 Journal of Healthcare Engineering

sparse LBP feature computation based on sparse Cur-velet coefficients and finally FP reduction using ANNclassifier

e proposed algorithm presents a novel concept ofextraction of curvelet coefficients according to irregularshape of mass is called as sparse curvelet coefficients andcomputation of LBP e analysis proves that the FPs arereduced significantly from 085 to 001 FPimage for MIAS

481 to 003 FPimage for DDSM and 232 to 000 FPimagefor TMCH e ANN classifier showed best results asAUC 098 and accuracy 9857 for MIAS in benign-malignant classification AUC 098 and accuracy 9870for DDSM in normal-malignant classification AUC 098and accuracy 9830 for TMCH Scanner1 and AUC 1and accuracy 100 for TMCH Scanner2 in normal-malignant classification as compared with SVM and KNN

(a) (b) (c) (d) (e) (f )

Figure 14 Representation of fully automatic CAD system for breast cancer using (a) samplemammograms fromMIAS DDSM and TMCHdatasets (b) preprocessed mammograms (c) clustered image (d) TP and FP marked on mammogram (e) TP marked by thresholding (f )TP marked by using LBP descriptor based on sparse curvelet coefficients

Journal of Healthcare Engineering 13

classifier e performance of LBP features and LBP featuresbased on sparse curvelet coefficients are nearly same whichshow that the proposed algorithm is suitable for cancer breasttissue diagnosis

In future the reduced curvelet coefficients can be used toextract local ternary patterns and other local descriptor andlocal directional patterns etc e present work deals withmammogram with single mass this can be further extendedfor multiple mass models with multiple LBP features basedon sparse curvelet coefficients

Data Availability

In this research we have used two publicly available datasetsMIAS and DDSM ese datasets can be found here in [17]and [19] e third database is collected from the localhospital Tata Memorial Cancer Hospital Mumbai whichcan be found at httpeurekasveriacin or available fromthe corresponding author upon request

Conflicts of Interest

e authors declare that they have no conflicts of interest

Acknowledgments

e TMCH database for this work was given by Departmentof Radiodiagnosis Tata Memorial Cancer Hospital Mumbai

References

[1] S Malvia S A Bagadi U S Dubey and S Saxena ldquoEpi-demiology of breast cancer in Indian womenrdquo Asia-PacificJournal of Clinical Oncology vol 13 no 4 pp 289ndash295 2017

[2] A Gupta K Shridhar and P Dhillon ldquoA review of breastcancer awareness among women in India cancer literate orawareness deficitrdquo European Journal of Cancer vol 51no 14 pp 2058ndash2066 2015

[3] K P Kanadam and S R Chereddy ldquoMammogram classifi-cation using sparse-ROI a novel representation to arbitraryshaped massesrdquo Expert Systems with Applications vol 57pp 204ndash213 2016

[4] D O T Bruno M Z do Nascimento R P Ramos et al ldquoLBPoperators on curvelet coefficients as an algorithm to describetexture in breast cancer tissuesrdquo Expert Systems with Appli-cations vol 55 pp 329ndash340 2016

[5] M Hussain ldquoFalse-positive reduction in mammographyusing multiscale spatial Weber law descriptor and supportvector machinesrdquo Neural Computing and Applicationsvol 25 no 1 pp 83ndash93 2014

[6] M M Pawar and S N Talbar ldquoGenetic fuzzy system (GFS)based wavelet co-occurrence feature selection in mammo-gram classification for breast cancer diagnosisrdquo Perspectives inScience vol 8 pp 247ndash250 2016

[7] C Muramatsu T Hara T Endo and H Fujita ldquoBreast massclassification on mammograms using radial local ternarypatternsrdquo Computers in Biology and Medicine vol 72pp 43ndash53 2016

[8] M M Eltoukhy I Faye and B B Samir ldquoA statistical basedfeature extraction method for breast cancer diagnosis indigital mammogram using multiresolution representationrdquo

Table 8 Comparison of classification accuracy AUC and FPimage values from different approaches in breast cancer diagnosis

Author Database Method Classifier Result AUC FPimageEltoukhy et al [33]

MIAS Biggest curvelet coefficients as a featurevector

Euclidean classifier 9407 mdash mdashEltoukhy et al [42] 9859 mdash mdashEltoukhy et al [8] SVM 973 mdash mdash

Dhahbi et al [34] Mini-MIAS Curvelet moments KNN 9127 mdash mdashDDSM 8646 mdash mdash

Bruno et al [4] DDSM Curvelet + LBP SVM 85 085 mdashPL 94 094 mdash

da Rocha et al [40] DDSM LBP SVM 8831 088 mdashKanadam andChereddy [3] MIAS Sparse ROI SVM 9742 mdash mdash

Pereira et al [18] DDSM Wavelet and Wiener filter Multiple thresholding waveletand GA mdash mdash 137

Liu and Zeng [29] DDSMFFDM GLCM CLBP and geometric features SVM mdash mdash 148

De Sampaio et al[39] DDSM LBP DBSCAN 9826 019

Zyout et al [30] DDSM Second order statistics of waveletcoefficients (SOSWC) SVM 968 097 0018

MIAS 952 966 0029

Casti et al [31]DDSM

Differential features Fisher linear discriminantanalysis (FLDA) mdash mdash

168MIAS 212FFDM 082

Proposed method

MIAS

LBP based on sparse curvelet subbandcoefficients ANN

9857 098 001DDSM 9870 098 003TMCHScanner1 9830 098 005

TMCHScanner2 100 1 0

14 Journal of Healthcare Engineering

Nonzero pixels provide actual shape of mass and are taken forLBP computations Graphical representation of proposedalgorithm for LBP descriptor computation using foregroundpixels has been given in Figure 7 and the algorithm has beendescribed in Algorithm 2

342 e Fast Discrete Curvelet Transform (FDCT) eauthors [44] have introduced computationally simple andeiexclcient Fast Discrete Curvelet Transform (FDCT) We havepreferred wrapping-based FDCT approach in proposedwork as it is faster e curvelet coeiexclcients CD(j l k)represented by scale j angle l and spatial location k can bewritten as

CD j l k1 k2( ) sumn1N1

n11sumn2N2

n21I n1 n2[ ]φD

jlk1k2n1 n2[ ] (8)

Figure 8 illustrates LBP code computation based on sparsecurvelet coeiexclcients ROI decomposes using curvelet trans-form with scale orientations l of 16deg and scale of 2 as thedatabase consists of minimum ROI size of 25 times 22 pixelsCurvelet transform with scale orientations l of 16deg and scale of2 produces 1 + 16 17 diyenerent subbands based on subbanddivision Further each curvelet subband coeiexclcients havebeen represented using lookup table using 3 times 3 slidingwindow and if the row in the lookup table identies fore-ground coeiexclcient then LBP is computed with radius R 1and P 8 neighboring pixels as shown in Algorithm 2 total 58LBP features have been obtained from foreground curveletsubband coeiexclcients erefore total 986 LBP features havebeen extracted from 17 curvelet subbands It can be observedfrom Figure 8 curvelet subbands also provide shape of massin 16 diyenerent directions so that the directional informationcan be associated with LBP features Kanadam et al [3] usedconcept of sparse ROI similarly we have extended it forsparse curvelet subband and LBP features computation

35 Classi cation In this work we have analyzed extractedROI from mammogram using normal-abnormal benign-malignant and normal-malignant classes with ANN SVMand KNN classiers e detailed description of ANNclassier has been given in [45 46] To evaluate performance

of the proposed system we have used 3-fold cross validationwhere database is randomly divided into three sets andaccuracy is calculated for each set e nal accuracy of thesystem is average of accuracy of each of three sets Howeverit will not be fair to compare 3-fold cross validation result ofSVM and KNN classier with ANN because ANN classieris tested on only one set of images (33 for training 33 fortesting and 33 for validation)us to do fair comparisonwe have trained ANN using input layer (986 neuron) overthree diyenerent sets (which are considered in SVM and KNN)and calculated its average accuracy Our proposed falsepositive reduction algorithm illustrates in Figures 9(a)ndash9(c)Algorithm 3 summarizes copyow of the proposed method forFP reduction in mammograms

4 Experimental Results and Discussions

e proposed method has been tested and validated usingthree classiers and three clinical mammographic imagedatasets

41 Data Sets

411 Mammographic Image Analysis Society (MIAS)Database e mini-MIAS [17] database consists of 322mammograms each having 1024times1024 pixels and anno-tated like background tissue character class severity centerof abnormality and radius of circle for abnormality isdatabase includes 64 benign 51 malignant and 207 normalcases which have been taken for experimentation

412 Digital Database for Screening Mammography(DDSM) e DDSM [19] dataset consists of 2500 studiesand is composed of cranial-caudal (CC) and mediolateral-oblique (MLO) views of mammographic image for left andright breast annotated with ACR breast density type ofabnormality and ground truth Randomly selected 150abnormal and 100 normal cases from both HOWTEK andLUMISYS scanner of 12 bits per pixel resolution have beensubjected for experimentation

ROI Shape of mass

Mass shape by sparse matrix

Position inlook-up table Nine neighborhood pixels Decision making

Yes

LBP operator on 3times3 neighborhood

Nine neighborhood pixels

120 125 210

105 45

254

10

93 200 250

1 1 1

1 0

1 1 1

Thresholding

LBP code

LBP code =0 + 2 + 4 +8 + 16 + 32+ 64 + 128

No

Pixel count(position(ij))

gt 2

Discard the rowfrom look-up table

eg X P etc

Input

bullbull

bullbull

bullbull

bullbull

bull

X 0 0 0 0 0 0 0 0 1

P 0 0 0 0 0 0 0 0 0

Q 0 0 0 37 200 0 100 102 0

Y 123 90 70 100 23 45 0 144 17

Z 120 105 93 125 45 200 210 10 250

Q position in look-up table

P position

Z position

Y position

X position

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

Figure 7 Process for computation of LBP descriptor from shape of mass in ROI (a) Original image (b) 3times 3 window for selection offoreground pixels (c) lookup table (d) decision making process (e) LBP computation from selected foreground pixels

6 Journal of Healthcare Engineering

413 e Tata Memorial Cancer Hospital (TMCH) isdataset [47] contains 360 full-eld digital mammograms(FFDMs) comprising 180 CC views and 180 MLO viewsfrom right and left breast acquired from 90 randomly se-lected patients It is composed of 180 veried malignant and180 normal breast images It uses biopsy proven breastcancer patientsrsquo pathological data approved by the In-stitutional Research Ethics Committee of Tata MemorialCentre Hospital (TMCH) Mumbai India e ground truthmarking on each abnormal mammogram is performedmanually using the Histopathological Reports (HPR) of therespective patients and expert radiologist from TMCHMumbai Approximately 35 patients are examined usingldquoHologic Selenia Systemrdquo (Scanner1) gives 16-bit

e remaining 55 patients were examined with ldquoGEMedical Senograph Systemrdquo (Scanner2) providing 8-bit truecolor mammogram image in DICOM format of 4096times 3328or 2294times1914 pixels each measuring size 50times 50 μm2

42 Segmentation Evaluation and ROI Extraction e seg-mentation using SOM that detects suspicious mass regions isconsidered as TP whereas from nonmass is taken as FPFrom Table 1 it is clear that total suspicious ROI (includingTP amp FP) of 381 for MIAS 1343 for DDSM and 1009 forTMCH have been taken for evaluation our proposed al-gorithm for FP reduction

From extracted ROIs the minimum patch size is25 times 22 pixels whereas the maximum size is 1152 times1356pixels Tables 2 and 3 represent curvelet subband co-eiexclcients from 17 subbands and reduced coeiexclcientsbased on lookup table approach are used to calculate LBPfeatures It has been observed during experimentationthat the curvelet coeiexclcients on an average are reducedfor sparse LBP by 14 32 33 and 34 for MIASDDSM TMCH Scanner1 and TMCH Scanner2 re-spectively It may be noticed that reduction in curveletcoeiexclcients for every ROI is not xed It completely

bullbullbull

bullbullbull

bullbullbull

Scale 1 approximatecoefficients

Scale 2 1 to 16subband coefficients

Curvelet transform

1

58

58

2

158 LBP

descriptorsfrom

approximatecoefficients

58 lowast 16LBP

descriptorsfrom

1 to 16subband curvelet

coefficients

2

Figure 8 LBP code computation using sparse curvelet subband coeiexclcients

(a) (b) (c)

Figure 9 (a) FP reduction by clusters marked on original image (b) FP reduction by thresholding (c) FP reduction by sparse curveletcoeiexclcient-based LBP and ANN

Journal of Healthcare Engineering 7

depends upon the shape of the ROI as per the sparsematrix Tables 2 and 3 do not represent exact reduction inpixels for complete database but they exhibit pixel re-duction for sample mammograms

43 Classifier Evaluation and False-Positive ReductionFrom Figures 10ndash13 the best classification accuracy of 9857 has been obtained for MIAS in benign versus malignantclassification whereas 9870 for DDSM 9830 for

(1) Load input image (img1)(2) Apply CLAHE algorithm and obtain enhanced image (img2)(3) Decompose img1 and img2 up to 3 level of decomposition using Discrete Wavelet transform (DWT)(4) Use maximum local entropy rule for fusion of img1 and img2 for high frequency subbands(5) Take inverse DWT to obtain the fused image

ALGORITHM 1 Image fusion for contrast enhancement

Input I(m n) m no of rows and nno of columnOutput LBP featuresInitialize Radius R 1 and neighborhood pixels P 8

Mask [1 2 4 8 16 32 64 128 0]Sliding window coordinates kminus1 1Count 1 number of pixels in I(m n)

for i 1 to m dofor j 1 to n do

prepare local circular windowI_local I(i+ k j+ k)center_pixel I(i j)Arrange local neighborhoods of I(i j) pixels in a row col 1 9Lookup_table (i col) reshape(I_local [7 17])count number of pixels greater than zeroa length(find(Lookup_tablegt 0))select pixel position from lookup-table for computation of LBPif agt 2

LBP_code(count) I_localgt center_pixelcount count + 1

endend

endcompute histogram of LBP codesLBP_descriptor LBP_descriptorcountscale invariant

ALGORITHM 2 Algorithm for LBP feature computation based on shape of mass in ROI as

(1) Load input image (img1)(2) Apply CLAHE algorithm and obtain enhanced image (img2)(3) Process img1 and img2 and obtain enhanced image using procedure given in Algorithm 1(4) Remove pectoral muscle using proposed approach (Section 312)(5) Extract neighbourhood features for each pixel and apply SOM clustering(6) Obtain clustered image and separate out the tumorous cluster(7) Extract detected regions ie ROIrsquos from clustered result(8) Extract Sparse Curvelet Coefficients (Subband) up to 2 level from each ROI(9) Extract Sparse LBP code for each subband and obtain a combined feature vector for each ROI(10) Classify each ROI into tumorous and nontumorous class ie TP and FP respectively(11) Map each TP region on original mammogram (img1)(12) end

ALGORITHM 3 Summary of proposed method for FP reduction in mammograms

8 Journal of Healthcare Engineering

TMCH Scanner1 and 100 for TMCH Scanner2 clas-sication accuracies have been obtained in normal versusmalignant classication e classication performance ofANN has improved from 6 to 43 for diyenerent databasesas compared to KNN classier whereas there is littleimprovement about 7 compared with SVM classier eperformances of both proposed sparse LBP and LBPcomputation on curvelet subbands are nearly sametherefore the proposed algorithm can be eiexclciently

implemented in CAD system with lesser number of cur-velet coeiexclcients

Data augmentation has been used for some classes tomaintain balance between two classes to improve perfor-mance and to learn more powerful model Table 4 explainsthe FP reduction with the use of curvelet-based LBP featuresand ANN It has been observed that FP reduced from 085 to002 FPimage in MIAS 481 to 002 FPimage in DDSM and232 to 013 FPimage in TMCH

8000

8500

9000

9500

10000

TMCHS1LBP

TMCHS1sparse_LBP

TMCHS2LBP

TMCHS2sparse_LBP

Normal_Malignant

KNNANNSVM

Figure 10 Average classication rate for TMCH dataset

000

2000

4000

6000

8000

10000

DDSMLBP DDSMsparse_LBP

MIASLBP MIASsparse_LBP

Normal_Malignant

KNNANNSVM

Figure 11 Average classication rate for MIAS and DDSM dataset

000

2000

4000

6000

8000

10000

DDSMLBP DDSMsparse_LBP

MIASLBP MIASsparse_LBP

Benign_Malignant

KNNANNSVM

Figure 12 Average classication rate for MIAS and DDSM dataset

Journal of Healthcare Engineering 9

Similarly Table 5 shows the reduction in FPs as 085to 001 FPimage for MIAS 481 to 003 FPimage forDDSM and 232 to 000 FPimage for TMCH using sparsecurvelet coeiexclcient-based LBP features e results show

the eyenectiveness of sparse curvelet coeiexclcient-based LBPand ANN From Table 6 the best value of AUC 099 isobtained in benign versus malignant classication forMIAS AUC 098 in benign versus malignant in case of

Table 1 Result of SOM segmentation

Datasetused

Result of SOM clustering and threshold TPR (true-positiverate)

TPlesions

FPPI (false-positiveper image)FPimages

MassSegmented nonmass (FP) Total () images

Segmented (TP) LostMIAS 108 7 273 322 (108115) 094 (273322) 085DDSM 140 10 1203 250 (140150) 093 (1203250) 481TMCH 172 8 837 360 (172180) 095 (837360) 232

Table 2 Reduction in curvelet coeiexclcients for sample mammograms from MIAS and DDSM dataset

SrNo

MIAS DDSM

ROI Size

Total number ofcurvelet

coeiexclcients fromsubbands

Total number ofselected curveletcoeiexclcients from

subbands

reductionin curveletcoeiexclcients

ROI Size

Total number ofcurvelet

coeiexclcients fromsubbands

Total number ofselected curveletcoeiexclcients from

subbands

reductionin curveletcoeiexclcients

1 124times138 103911 77133 2577 192times187 216729 168333 22332 179times138 150123 126142 1597 294times 291 518267 366680 29253 51times 116 36815 33421 922 145times 207 181663 148765 18114 83times 83 42449 36653 1365 169times168 171873 154752 9965 84times 76 39115 35815 844 182times 248 272517 219822 19346 74times 83 37767 34448 879 213times 349 449783 301359 33007 53times 64 20969 18621 1120 578times 412 1433195 662072 53808 70times 44 18899 16610 1211 215times 219 286429 242935 15189 80times 66 32409 29454 912 420times 428 1082461 501209 537010 69times 86 36783 33552 878 203times 307 375871 266829 290111 59times116 42019 38442 851 226times 262 357209 276763 225212 81times 101 50637 46122 892 159times194 187563 148741 207013 41times 84 21427 18907 1176 718times 686 2961127 762561 742514 69times141 60641 52427 1354 409times 550 1352439 904829 331015 60times 62 22925 20475 1069 524times 375 1184671 766522 353016 96times101 59647 54235 907 311times 275 517433 397737 231317 136times139 113727 66352 4156 319times 320 614129 412606 328118 55times 94 31359 28305 974 313times 447 842855 510482 394319 157times140 132373 107000 1917 291times 517 907903 637412 297920 156times130 123007 90834 2615 370times 837 1864889 898236 5183

Average 58850 48247 14 Average 788950 437432 32

000

2000

4000

6000

8000

10000

DDSMLBP DDSM sparse_LBP

MIASLBP MIAS sparse_LBP

Normal_Abnormal

KNNANNSVM

Figure 13 Average classication rate for MIAS and DDSM dataset

10 Journal of Healthcare Engineering

DDSM AUC 094 in normal versus malignant in case ofTMCH Scanner1 and AUC 096 in normal versusmalignant classification in TMCH Scanner2 using ANNand curvelet subband-based LBP features e worstperformance of AUC 053 for MIAS is obtained with theproposed algorithm using KNN classifier as shown inTable 7 Similarly from Table 7 the best value ofAUC 098 is obtained in TMCH Scanner1 AUC 1 isobtained in TMCH Scanner2 database for normal versusmalignant classification AUC 098 in benign versusmalignant classification is attained in MIAS database andAUC 098 is achieved for normal versus malignant

classification in DDSM database using ANN classifier forsparse curvelet subband-based LBP features

However from Table 7 it should be noted that the per-formance of proposed algorithm is the best using ANN clas-sifier Figure 14 represents automated CAD system for breastcancer diagnosis with sample mammograms

Table 8 provides comparative study of methods developedfor breast tissue classification e proposed method providesbest results in terms of AUC and reduction of number of FPsas 085 to 001 FPimage for MIAS 481 to 003 FPimage forDDSM and 232 to 000 FPimage for TMCH e earlierreported work uses the fixed patch size-based approach which

Table 3 Reduction in curvelet coefficients for sample mammograms from TMCH Scanner1 and Scanner2 dataset

Srno

TMCH Scanner 1 ldquoGE Medical Senograph Systemrdquo TMCH Scanner 2 ldquoHologic Selenia Systemrdquo

ROI size

Total number ofcurvelet

coefficients fromsubbands

Total number ofselected curveletcoefficients from

subbands

reductionin curveletcoefficients

ROI size

Total number ofcurvelet

coefficients fromsubbands

Total number ofselected curveletcoefficients from

subbands

reductionin curveletcoefficients

1 459times 412 1140617 794885 3031 291times 278 489243 317014 35202 548times 513 1693403 1117873 3399 545times 246 809627 531604 34343 415times 303 758323 445585 4124 560times 483 1630583 1068439 34474 645times 495 1951443 1145580 4129 782times 510 2400137 1275073 46875 437times 691 1816651 985120 4577 87times141 75773 67871 10436 812times 500 2439937 1287065 4725 311times 185 348565 240546 30997 468times 379 1066333 710242 3340 262times 348 550303 282821 48618 673times 582 2355589 1730915 2652 610times 440 1614515 842876 47799 250times 201 304513 235670 2261 949times 391 2227209 1359338 389710 525times 488 1544691 1161942 2478 365times 385 846523 604473 285911 488times 779 2287547 1611385 2956 393times 247 585063 490474 161712 1434times 966 8326581 3742277 5506 341times 301 618111 410542 335813 348times 421 881701 616501 3008 523times 702 2206097 800057 637314 460times 530 1467227 799885 4548 370times 284 632539 423727 330115 398times 450 1078441 828064 2322 344times 202 418955 302427 278116 247times 272 404401 344822 1473 264times188 299997 231926 226917 411times 305 757657 405919 4642 233times 247 346983 267686 228518 286times 344 593155 448566 2438 370times 291 650701 486125 252919 417times 207 523477 429755 1790 680times 483 1979543 1052793 468220 463times 458 1275021 857608 3274 202times 266 324295 215042 3369

Average 1633335 984983 33 Average 952738 563543 34

Table 4 Number of ROIs resulted in FP reduction using curvelet-based LBP (without sparse) amp ANN classification at training andvalidation stage

Class Datasetused

Benignmalignant mass Nonmassbenign mass Total()

images

TPR (true-positiverate)TPlesions

FPPI (false-positiveper image) FP

imagesPreviousstage

Selected(TP)

Lost(FN)

Previousstage

Selected(TN)

Lost(FP)

Normal vsabnormal

MIAS 108lowast 2 216 203 13 273 257 16 315 (203216) 094 (16315) 005DDSM 140lowast 4 560 465 95 1203 1095 108 240 (465560) 083 (108240) 045

Benign vsmalignant

MIAS 49 49 0 59 57 2 108 (4949) 100 (2108) 002DDSM 46lowast 2 92 91 1 94 91 3 140 (9192) 099 (3140) 002

Normal vsmalignant

MIAS 49lowast 4196 184 12 273 254 19 256 (184196) 094 (19256) 007DDSM 46lowast 4184 180 4 1203 1143 60 146 (180184) 098 (60146) 041TMCHScanner1 107lowast 4 428 416 12 605 551 54 217 (416428) 097 (54217) 025

TMCHScanner2 65lowast 4 260 255 5 232 214 18 135 (255260) 098 (18135) 013

lowastAugmentation of image

Journal of Healthcare Engineering 11

limits the automatic CAD system scope whereas proposedsystem provides complete solution to CAD system right fromautomatic tumor patch segmentation to reduction in FPs andfinal representation of mammogram with TP marked on itIt will drastically reduce the radiologist work by locationtumor directly on mammogram

5 Conclusion

A fully automatic CAD system which can accuratelylocate the tumor on a mammogram and reduces FPshas been proposed e developed CAD system consistsof preprocessing SOM clustering ROI extraction

Table 6 Performance evaluation of curvelet-based LBP descriptor algorithm

Dataset Classification Normal-malignant Normal-abnormal Benign-malignantClassifier Sensitivity Specificity AUC Sensitivity Specificity AUC Sensitivity Specificity AUC

MIASANN 094 093 094 094 094 094 100 097 099SVM 085 085 085 083 086 085 088 084 086KNN 067 063 065 058 057 058 062 068 063

DDSMANN 098 095 095 083 091 085 099 097 098SVM 097 088 092 071 091 083 094 089 092KNN 096 064 087 067 090 080 087 073 079

TMCH Scanner1ANN 097 091 094 mdash mdash mdash mdash mdash mdashSVM 096 091 094 mdash mdash mdash mdash mdash mdashKNN 098 082 089 mdash mdash mdash mdash mdash mdash

TMCH Scanner2ANN 098 092 096 mdash mdash mdash mdash mdash mdashSVM 097 090 094 mdash mdash mdash mdash mdash mdashKNN 092 083 088 mdash mdash mdash mdash mdash mdash

Table 7 Performance evaluation of proposed algorithm

Dataset Classification Normal-malignant Normal-abnormal Benign-malignantClassifier Sensitivity Specificity AUC Sensitivity Specificity AUC Sensitivity Specificity AUC

MIASANN 098 095 096 093 097 095 097 100 098SVM 088 083 085 085 082 084 084 092 087KNN 055 051 053 055 063 056 061 067 061

DDSMANN 099 097 098 092 096 093 097 095 096SVM 099 092 096 089 096 092 094 092 093KNN 098 073 092 074 090 082 089 077 083

TMCH Scanner1ANN 099 098 098 mdash mdash mdash mdash mdash mdashSVM 098 096 097 mdash mdash mdash mdash mdash mdashKNN 099 092 095 mdash mdash mdash mdash mdash mdash

TMCH Scanner2ANN 100 100 100 mdash mdash mdash mdash mdash mdashSVM 100 098 099 mdash mdash mdash mdash mdash mdashKNN 096 092 094 mdash mdash mdash mdash mdash mdash

Table 5 Number of ROIs resulted in FP reduction using sparse curvelet coefficient-based LBP amp ANN classification at training andvalidation stage

Class Datasetused

Benignmalignant mass Nonmassbenign massTotal ()images

TPR (true-positiverate)TPlesions

FPPI (false-positiveper image) FP

imagesPreviousstage

Selected(TP)

Lost(FN)

Previousstage

Selected(TN)

Lost(FP)

Normal vsabnormal

MIAS 108lowast 2 216 201 15 273 265 8 315 (201216) 093 (8315) 002DDSM 140lowast 4 560 516 44 1203 1155 48 240 (516560) 092 (48240) 02

Benign vsmalignant

MIAS 49 48 1 59 59 1 108 (4849) 098 (1108) 001DDSM 46lowast 2 92 89 3 94 89 5 140 (8992) 097 (5140) 003

Normal vsmalignant

MIAS 49lowast 4196 192 4 273 259 14 256 (192196) 098 (14256) 005DDSM 46lowast 4184 182 2 1203 1167 36 146 (182184) 099 (36146) 025TMCHScanner1 107lowast 4 428 424 4 605 593 12 217 (424428) 099 (12217) 005

TMCHScanner2 65lowast 4 260 260 0 232 232 0 135 (260260) 100 (0135) 0

lowastAugmentation of image

12 Journal of Healthcare Engineering

sparse LBP feature computation based on sparse Cur-velet coefficients and finally FP reduction using ANNclassifier

e proposed algorithm presents a novel concept ofextraction of curvelet coefficients according to irregularshape of mass is called as sparse curvelet coefficients andcomputation of LBP e analysis proves that the FPs arereduced significantly from 085 to 001 FPimage for MIAS

481 to 003 FPimage for DDSM and 232 to 000 FPimagefor TMCH e ANN classifier showed best results asAUC 098 and accuracy 9857 for MIAS in benign-malignant classification AUC 098 and accuracy 9870for DDSM in normal-malignant classification AUC 098and accuracy 9830 for TMCH Scanner1 and AUC 1and accuracy 100 for TMCH Scanner2 in normal-malignant classification as compared with SVM and KNN

(a) (b) (c) (d) (e) (f )

Figure 14 Representation of fully automatic CAD system for breast cancer using (a) samplemammograms fromMIAS DDSM and TMCHdatasets (b) preprocessed mammograms (c) clustered image (d) TP and FP marked on mammogram (e) TP marked by thresholding (f )TP marked by using LBP descriptor based on sparse curvelet coefficients

Journal of Healthcare Engineering 13

classifier e performance of LBP features and LBP featuresbased on sparse curvelet coefficients are nearly same whichshow that the proposed algorithm is suitable for cancer breasttissue diagnosis

In future the reduced curvelet coefficients can be used toextract local ternary patterns and other local descriptor andlocal directional patterns etc e present work deals withmammogram with single mass this can be further extendedfor multiple mass models with multiple LBP features basedon sparse curvelet coefficients

Data Availability

In this research we have used two publicly available datasetsMIAS and DDSM ese datasets can be found here in [17]and [19] e third database is collected from the localhospital Tata Memorial Cancer Hospital Mumbai whichcan be found at httpeurekasveriacin or available fromthe corresponding author upon request

Conflicts of Interest

e authors declare that they have no conflicts of interest

Acknowledgments

e TMCH database for this work was given by Departmentof Radiodiagnosis Tata Memorial Cancer Hospital Mumbai

References

[1] S Malvia S A Bagadi U S Dubey and S Saxena ldquoEpi-demiology of breast cancer in Indian womenrdquo Asia-PacificJournal of Clinical Oncology vol 13 no 4 pp 289ndash295 2017

[2] A Gupta K Shridhar and P Dhillon ldquoA review of breastcancer awareness among women in India cancer literate orawareness deficitrdquo European Journal of Cancer vol 51no 14 pp 2058ndash2066 2015

[3] K P Kanadam and S R Chereddy ldquoMammogram classifi-cation using sparse-ROI a novel representation to arbitraryshaped massesrdquo Expert Systems with Applications vol 57pp 204ndash213 2016

[4] D O T Bruno M Z do Nascimento R P Ramos et al ldquoLBPoperators on curvelet coefficients as an algorithm to describetexture in breast cancer tissuesrdquo Expert Systems with Appli-cations vol 55 pp 329ndash340 2016

[5] M Hussain ldquoFalse-positive reduction in mammographyusing multiscale spatial Weber law descriptor and supportvector machinesrdquo Neural Computing and Applicationsvol 25 no 1 pp 83ndash93 2014

[6] M M Pawar and S N Talbar ldquoGenetic fuzzy system (GFS)based wavelet co-occurrence feature selection in mammo-gram classification for breast cancer diagnosisrdquo Perspectives inScience vol 8 pp 247ndash250 2016

[7] C Muramatsu T Hara T Endo and H Fujita ldquoBreast massclassification on mammograms using radial local ternarypatternsrdquo Computers in Biology and Medicine vol 72pp 43ndash53 2016

[8] M M Eltoukhy I Faye and B B Samir ldquoA statistical basedfeature extraction method for breast cancer diagnosis indigital mammogram using multiresolution representationrdquo

Table 8 Comparison of classification accuracy AUC and FPimage values from different approaches in breast cancer diagnosis

Author Database Method Classifier Result AUC FPimageEltoukhy et al [33]

MIAS Biggest curvelet coefficients as a featurevector

Euclidean classifier 9407 mdash mdashEltoukhy et al [42] 9859 mdash mdashEltoukhy et al [8] SVM 973 mdash mdash

Dhahbi et al [34] Mini-MIAS Curvelet moments KNN 9127 mdash mdashDDSM 8646 mdash mdash

Bruno et al [4] DDSM Curvelet + LBP SVM 85 085 mdashPL 94 094 mdash

da Rocha et al [40] DDSM LBP SVM 8831 088 mdashKanadam andChereddy [3] MIAS Sparse ROI SVM 9742 mdash mdash

Pereira et al [18] DDSM Wavelet and Wiener filter Multiple thresholding waveletand GA mdash mdash 137

Liu and Zeng [29] DDSMFFDM GLCM CLBP and geometric features SVM mdash mdash 148

De Sampaio et al[39] DDSM LBP DBSCAN 9826 019

Zyout et al [30] DDSM Second order statistics of waveletcoefficients (SOSWC) SVM 968 097 0018

MIAS 952 966 0029

Casti et al [31]DDSM

Differential features Fisher linear discriminantanalysis (FLDA) mdash mdash

168MIAS 212FFDM 082

Proposed method

MIAS

LBP based on sparse curvelet subbandcoefficients ANN

9857 098 001DDSM 9870 098 003TMCHScanner1 9830 098 005

TMCHScanner2 100 1 0

14 Journal of Healthcare Engineering

413 e Tata Memorial Cancer Hospital (TMCH) isdataset [47] contains 360 full-eld digital mammograms(FFDMs) comprising 180 CC views and 180 MLO viewsfrom right and left breast acquired from 90 randomly se-lected patients It is composed of 180 veried malignant and180 normal breast images It uses biopsy proven breastcancer patientsrsquo pathological data approved by the In-stitutional Research Ethics Committee of Tata MemorialCentre Hospital (TMCH) Mumbai India e ground truthmarking on each abnormal mammogram is performedmanually using the Histopathological Reports (HPR) of therespective patients and expert radiologist from TMCHMumbai Approximately 35 patients are examined usingldquoHologic Selenia Systemrdquo (Scanner1) gives 16-bit

e remaining 55 patients were examined with ldquoGEMedical Senograph Systemrdquo (Scanner2) providing 8-bit truecolor mammogram image in DICOM format of 4096times 3328or 2294times1914 pixels each measuring size 50times 50 μm2

42 Segmentation Evaluation and ROI Extraction e seg-mentation using SOM that detects suspicious mass regions isconsidered as TP whereas from nonmass is taken as FPFrom Table 1 it is clear that total suspicious ROI (includingTP amp FP) of 381 for MIAS 1343 for DDSM and 1009 forTMCH have been taken for evaluation our proposed al-gorithm for FP reduction

From extracted ROIs the minimum patch size is25 times 22 pixels whereas the maximum size is 1152 times1356pixels Tables 2 and 3 represent curvelet subband co-eiexclcients from 17 subbands and reduced coeiexclcientsbased on lookup table approach are used to calculate LBPfeatures It has been observed during experimentationthat the curvelet coeiexclcients on an average are reducedfor sparse LBP by 14 32 33 and 34 for MIASDDSM TMCH Scanner1 and TMCH Scanner2 re-spectively It may be noticed that reduction in curveletcoeiexclcients for every ROI is not xed It completely

bullbullbull

bullbullbull

bullbullbull

Scale 1 approximatecoefficients

Scale 2 1 to 16subband coefficients

Curvelet transform

1

58

58

2

158 LBP

descriptorsfrom

approximatecoefficients

58 lowast 16LBP

descriptorsfrom

1 to 16subband curvelet

coefficients

2

Figure 8 LBP code computation using sparse curvelet subband coeiexclcients

(a) (b) (c)

Figure 9 (a) FP reduction by clusters marked on original image (b) FP reduction by thresholding (c) FP reduction by sparse curveletcoeiexclcient-based LBP and ANN

Journal of Healthcare Engineering 7

depends upon the shape of the ROI as per the sparsematrix Tables 2 and 3 do not represent exact reduction inpixels for complete database but they exhibit pixel re-duction for sample mammograms

43 Classifier Evaluation and False-Positive ReductionFrom Figures 10ndash13 the best classification accuracy of 9857 has been obtained for MIAS in benign versus malignantclassification whereas 9870 for DDSM 9830 for

(1) Load input image (img1)(2) Apply CLAHE algorithm and obtain enhanced image (img2)(3) Decompose img1 and img2 up to 3 level of decomposition using Discrete Wavelet transform (DWT)(4) Use maximum local entropy rule for fusion of img1 and img2 for high frequency subbands(5) Take inverse DWT to obtain the fused image

ALGORITHM 1 Image fusion for contrast enhancement

Input I(m n) m no of rows and nno of columnOutput LBP featuresInitialize Radius R 1 and neighborhood pixels P 8

Mask [1 2 4 8 16 32 64 128 0]Sliding window coordinates kminus1 1Count 1 number of pixels in I(m n)

for i 1 to m dofor j 1 to n do

prepare local circular windowI_local I(i+ k j+ k)center_pixel I(i j)Arrange local neighborhoods of I(i j) pixels in a row col 1 9Lookup_table (i col) reshape(I_local [7 17])count number of pixels greater than zeroa length(find(Lookup_tablegt 0))select pixel position from lookup-table for computation of LBPif agt 2

LBP_code(count) I_localgt center_pixelcount count + 1

endend

endcompute histogram of LBP codesLBP_descriptor LBP_descriptorcountscale invariant

ALGORITHM 2 Algorithm for LBP feature computation based on shape of mass in ROI as

(1) Load input image (img1)(2) Apply CLAHE algorithm and obtain enhanced image (img2)(3) Process img1 and img2 and obtain enhanced image using procedure given in Algorithm 1(4) Remove pectoral muscle using proposed approach (Section 312)(5) Extract neighbourhood features for each pixel and apply SOM clustering(6) Obtain clustered image and separate out the tumorous cluster(7) Extract detected regions ie ROIrsquos from clustered result(8) Extract Sparse Curvelet Coefficients (Subband) up to 2 level from each ROI(9) Extract Sparse LBP code for each subband and obtain a combined feature vector for each ROI(10) Classify each ROI into tumorous and nontumorous class ie TP and FP respectively(11) Map each TP region on original mammogram (img1)(12) end

ALGORITHM 3 Summary of proposed method for FP reduction in mammograms

8 Journal of Healthcare Engineering

TMCH Scanner1 and 100 for TMCH Scanner2 clas-sication accuracies have been obtained in normal versusmalignant classication e classication performance ofANN has improved from 6 to 43 for diyenerent databasesas compared to KNN classier whereas there is littleimprovement about 7 compared with SVM classier eperformances of both proposed sparse LBP and LBPcomputation on curvelet subbands are nearly sametherefore the proposed algorithm can be eiexclciently

implemented in CAD system with lesser number of cur-velet coeiexclcients

Data augmentation has been used for some classes tomaintain balance between two classes to improve perfor-mance and to learn more powerful model Table 4 explainsthe FP reduction with the use of curvelet-based LBP featuresand ANN It has been observed that FP reduced from 085 to002 FPimage in MIAS 481 to 002 FPimage in DDSM and232 to 013 FPimage in TMCH

8000

8500

9000

9500

10000

TMCHS1LBP

TMCHS1sparse_LBP

TMCHS2LBP

TMCHS2sparse_LBP

Normal_Malignant

KNNANNSVM

Figure 10 Average classication rate for TMCH dataset

000

2000

4000

6000

8000

10000

DDSMLBP DDSMsparse_LBP

MIASLBP MIASsparse_LBP

Normal_Malignant

KNNANNSVM

Figure 11 Average classication rate for MIAS and DDSM dataset

000

2000

4000

6000

8000

10000

DDSMLBP DDSMsparse_LBP

MIASLBP MIASsparse_LBP

Benign_Malignant

KNNANNSVM

Figure 12 Average classication rate for MIAS and DDSM dataset

Journal of Healthcare Engineering 9

Similarly Table 5 shows the reduction in FPs as 085to 001 FPimage for MIAS 481 to 003 FPimage forDDSM and 232 to 000 FPimage for TMCH using sparsecurvelet coeiexclcient-based LBP features e results show

the eyenectiveness of sparse curvelet coeiexclcient-based LBPand ANN From Table 6 the best value of AUC 099 isobtained in benign versus malignant classication forMIAS AUC 098 in benign versus malignant in case of

Table 1 Result of SOM segmentation

Datasetused

Result of SOM clustering and threshold TPR (true-positiverate)

TPlesions

FPPI (false-positiveper image)FPimages

MassSegmented nonmass (FP) Total () images

Segmented (TP) LostMIAS 108 7 273 322 (108115) 094 (273322) 085DDSM 140 10 1203 250 (140150) 093 (1203250) 481TMCH 172 8 837 360 (172180) 095 (837360) 232

Table 2 Reduction in curvelet coeiexclcients for sample mammograms from MIAS and DDSM dataset

SrNo

MIAS DDSM

ROI Size

Total number ofcurvelet

coeiexclcients fromsubbands

Total number ofselected curveletcoeiexclcients from

subbands

reductionin curveletcoeiexclcients

ROI Size

Total number ofcurvelet

coeiexclcients fromsubbands

Total number ofselected curveletcoeiexclcients from

subbands

reductionin curveletcoeiexclcients

1 124times138 103911 77133 2577 192times187 216729 168333 22332 179times138 150123 126142 1597 294times 291 518267 366680 29253 51times 116 36815 33421 922 145times 207 181663 148765 18114 83times 83 42449 36653 1365 169times168 171873 154752 9965 84times 76 39115 35815 844 182times 248 272517 219822 19346 74times 83 37767 34448 879 213times 349 449783 301359 33007 53times 64 20969 18621 1120 578times 412 1433195 662072 53808 70times 44 18899 16610 1211 215times 219 286429 242935 15189 80times 66 32409 29454 912 420times 428 1082461 501209 537010 69times 86 36783 33552 878 203times 307 375871 266829 290111 59times116 42019 38442 851 226times 262 357209 276763 225212 81times 101 50637 46122 892 159times194 187563 148741 207013 41times 84 21427 18907 1176 718times 686 2961127 762561 742514 69times141 60641 52427 1354 409times 550 1352439 904829 331015 60times 62 22925 20475 1069 524times 375 1184671 766522 353016 96times101 59647 54235 907 311times 275 517433 397737 231317 136times139 113727 66352 4156 319times 320 614129 412606 328118 55times 94 31359 28305 974 313times 447 842855 510482 394319 157times140 132373 107000 1917 291times 517 907903 637412 297920 156times130 123007 90834 2615 370times 837 1864889 898236 5183

Average 58850 48247 14 Average 788950 437432 32

000

2000

4000

6000

8000

10000

DDSMLBP DDSM sparse_LBP

MIASLBP MIAS sparse_LBP

Normal_Abnormal

KNNANNSVM

Figure 13 Average classication rate for MIAS and DDSM dataset

10 Journal of Healthcare Engineering

DDSM AUC 094 in normal versus malignant in case ofTMCH Scanner1 and AUC 096 in normal versusmalignant classification in TMCH Scanner2 using ANNand curvelet subband-based LBP features e worstperformance of AUC 053 for MIAS is obtained with theproposed algorithm using KNN classifier as shown inTable 7 Similarly from Table 7 the best value ofAUC 098 is obtained in TMCH Scanner1 AUC 1 isobtained in TMCH Scanner2 database for normal versusmalignant classification AUC 098 in benign versusmalignant classification is attained in MIAS database andAUC 098 is achieved for normal versus malignant

classification in DDSM database using ANN classifier forsparse curvelet subband-based LBP features

However from Table 7 it should be noted that the per-formance of proposed algorithm is the best using ANN clas-sifier Figure 14 represents automated CAD system for breastcancer diagnosis with sample mammograms

Table 8 provides comparative study of methods developedfor breast tissue classification e proposed method providesbest results in terms of AUC and reduction of number of FPsas 085 to 001 FPimage for MIAS 481 to 003 FPimage forDDSM and 232 to 000 FPimage for TMCH e earlierreported work uses the fixed patch size-based approach which

Table 3 Reduction in curvelet coefficients for sample mammograms from TMCH Scanner1 and Scanner2 dataset

Srno

TMCH Scanner 1 ldquoGE Medical Senograph Systemrdquo TMCH Scanner 2 ldquoHologic Selenia Systemrdquo

ROI size

Total number ofcurvelet

coefficients fromsubbands

Total number ofselected curveletcoefficients from

subbands

reductionin curveletcoefficients

ROI size

Total number ofcurvelet

coefficients fromsubbands

Total number ofselected curveletcoefficients from

subbands

reductionin curveletcoefficients

1 459times 412 1140617 794885 3031 291times 278 489243 317014 35202 548times 513 1693403 1117873 3399 545times 246 809627 531604 34343 415times 303 758323 445585 4124 560times 483 1630583 1068439 34474 645times 495 1951443 1145580 4129 782times 510 2400137 1275073 46875 437times 691 1816651 985120 4577 87times141 75773 67871 10436 812times 500 2439937 1287065 4725 311times 185 348565 240546 30997 468times 379 1066333 710242 3340 262times 348 550303 282821 48618 673times 582 2355589 1730915 2652 610times 440 1614515 842876 47799 250times 201 304513 235670 2261 949times 391 2227209 1359338 389710 525times 488 1544691 1161942 2478 365times 385 846523 604473 285911 488times 779 2287547 1611385 2956 393times 247 585063 490474 161712 1434times 966 8326581 3742277 5506 341times 301 618111 410542 335813 348times 421 881701 616501 3008 523times 702 2206097 800057 637314 460times 530 1467227 799885 4548 370times 284 632539 423727 330115 398times 450 1078441 828064 2322 344times 202 418955 302427 278116 247times 272 404401 344822 1473 264times188 299997 231926 226917 411times 305 757657 405919 4642 233times 247 346983 267686 228518 286times 344 593155 448566 2438 370times 291 650701 486125 252919 417times 207 523477 429755 1790 680times 483 1979543 1052793 468220 463times 458 1275021 857608 3274 202times 266 324295 215042 3369

Average 1633335 984983 33 Average 952738 563543 34

Table 4 Number of ROIs resulted in FP reduction using curvelet-based LBP (without sparse) amp ANN classification at training andvalidation stage

Class Datasetused

Benignmalignant mass Nonmassbenign mass Total()

images

TPR (true-positiverate)TPlesions

FPPI (false-positiveper image) FP

imagesPreviousstage

Selected(TP)

Lost(FN)

Previousstage

Selected(TN)

Lost(FP)

Normal vsabnormal

MIAS 108lowast 2 216 203 13 273 257 16 315 (203216) 094 (16315) 005DDSM 140lowast 4 560 465 95 1203 1095 108 240 (465560) 083 (108240) 045

Benign vsmalignant

MIAS 49 49 0 59 57 2 108 (4949) 100 (2108) 002DDSM 46lowast 2 92 91 1 94 91 3 140 (9192) 099 (3140) 002

Normal vsmalignant

MIAS 49lowast 4196 184 12 273 254 19 256 (184196) 094 (19256) 007DDSM 46lowast 4184 180 4 1203 1143 60 146 (180184) 098 (60146) 041TMCHScanner1 107lowast 4 428 416 12 605 551 54 217 (416428) 097 (54217) 025

TMCHScanner2 65lowast 4 260 255 5 232 214 18 135 (255260) 098 (18135) 013

lowastAugmentation of image

Journal of Healthcare Engineering 11

limits the automatic CAD system scope whereas proposedsystem provides complete solution to CAD system right fromautomatic tumor patch segmentation to reduction in FPs andfinal representation of mammogram with TP marked on itIt will drastically reduce the radiologist work by locationtumor directly on mammogram

5 Conclusion

A fully automatic CAD system which can accuratelylocate the tumor on a mammogram and reduces FPshas been proposed e developed CAD system consistsof preprocessing SOM clustering ROI extraction

Table 6 Performance evaluation of curvelet-based LBP descriptor algorithm

Dataset Classification Normal-malignant Normal-abnormal Benign-malignantClassifier Sensitivity Specificity AUC Sensitivity Specificity AUC Sensitivity Specificity AUC

MIASANN 094 093 094 094 094 094 100 097 099SVM 085 085 085 083 086 085 088 084 086KNN 067 063 065 058 057 058 062 068 063

DDSMANN 098 095 095 083 091 085 099 097 098SVM 097 088 092 071 091 083 094 089 092KNN 096 064 087 067 090 080 087 073 079

TMCH Scanner1ANN 097 091 094 mdash mdash mdash mdash mdash mdashSVM 096 091 094 mdash mdash mdash mdash mdash mdashKNN 098 082 089 mdash mdash mdash mdash mdash mdash

TMCH Scanner2ANN 098 092 096 mdash mdash mdash mdash mdash mdashSVM 097 090 094 mdash mdash mdash mdash mdash mdashKNN 092 083 088 mdash mdash mdash mdash mdash mdash

Table 7 Performance evaluation of proposed algorithm

Dataset Classification Normal-malignant Normal-abnormal Benign-malignantClassifier Sensitivity Specificity AUC Sensitivity Specificity AUC Sensitivity Specificity AUC

MIASANN 098 095 096 093 097 095 097 100 098SVM 088 083 085 085 082 084 084 092 087KNN 055 051 053 055 063 056 061 067 061

DDSMANN 099 097 098 092 096 093 097 095 096SVM 099 092 096 089 096 092 094 092 093KNN 098 073 092 074 090 082 089 077 083

TMCH Scanner1ANN 099 098 098 mdash mdash mdash mdash mdash mdashSVM 098 096 097 mdash mdash mdash mdash mdash mdashKNN 099 092 095 mdash mdash mdash mdash mdash mdash

TMCH Scanner2ANN 100 100 100 mdash mdash mdash mdash mdash mdashSVM 100 098 099 mdash mdash mdash mdash mdash mdashKNN 096 092 094 mdash mdash mdash mdash mdash mdash

Table 5 Number of ROIs resulted in FP reduction using sparse curvelet coefficient-based LBP amp ANN classification at training andvalidation stage

Class Datasetused

Benignmalignant mass Nonmassbenign massTotal ()images

TPR (true-positiverate)TPlesions

FPPI (false-positiveper image) FP

imagesPreviousstage

Selected(TP)

Lost(FN)

Previousstage

Selected(TN)

Lost(FP)

Normal vsabnormal

MIAS 108lowast 2 216 201 15 273 265 8 315 (201216) 093 (8315) 002DDSM 140lowast 4 560 516 44 1203 1155 48 240 (516560) 092 (48240) 02

Benign vsmalignant

MIAS 49 48 1 59 59 1 108 (4849) 098 (1108) 001DDSM 46lowast 2 92 89 3 94 89 5 140 (8992) 097 (5140) 003

Normal vsmalignant

MIAS 49lowast 4196 192 4 273 259 14 256 (192196) 098 (14256) 005DDSM 46lowast 4184 182 2 1203 1167 36 146 (182184) 099 (36146) 025TMCHScanner1 107lowast 4 428 424 4 605 593 12 217 (424428) 099 (12217) 005

TMCHScanner2 65lowast 4 260 260 0 232 232 0 135 (260260) 100 (0135) 0

lowastAugmentation of image

12 Journal of Healthcare Engineering

sparse LBP feature computation based on sparse Cur-velet coefficients and finally FP reduction using ANNclassifier

e proposed algorithm presents a novel concept ofextraction of curvelet coefficients according to irregularshape of mass is called as sparse curvelet coefficients andcomputation of LBP e analysis proves that the FPs arereduced significantly from 085 to 001 FPimage for MIAS

481 to 003 FPimage for DDSM and 232 to 000 FPimagefor TMCH e ANN classifier showed best results asAUC 098 and accuracy 9857 for MIAS in benign-malignant classification AUC 098 and accuracy 9870for DDSM in normal-malignant classification AUC 098and accuracy 9830 for TMCH Scanner1 and AUC 1and accuracy 100 for TMCH Scanner2 in normal-malignant classification as compared with SVM and KNN

(a) (b) (c) (d) (e) (f )

Figure 14 Representation of fully automatic CAD system for breast cancer using (a) samplemammograms fromMIAS DDSM and TMCHdatasets (b) preprocessed mammograms (c) clustered image (d) TP and FP marked on mammogram (e) TP marked by thresholding (f )TP marked by using LBP descriptor based on sparse curvelet coefficients

Journal of Healthcare Engineering 13

classifier e performance of LBP features and LBP featuresbased on sparse curvelet coefficients are nearly same whichshow that the proposed algorithm is suitable for cancer breasttissue diagnosis

In future the reduced curvelet coefficients can be used toextract local ternary patterns and other local descriptor andlocal directional patterns etc e present work deals withmammogram with single mass this can be further extendedfor multiple mass models with multiple LBP features basedon sparse curvelet coefficients

Data Availability

In this research we have used two publicly available datasetsMIAS and DDSM ese datasets can be found here in [17]and [19] e third database is collected from the localhospital Tata Memorial Cancer Hospital Mumbai whichcan be found at httpeurekasveriacin or available fromthe corresponding author upon request

Conflicts of Interest

e authors declare that they have no conflicts of interest

Acknowledgments

e TMCH database for this work was given by Departmentof Radiodiagnosis Tata Memorial Cancer Hospital Mumbai

References

[1] S Malvia S A Bagadi U S Dubey and S Saxena ldquoEpi-demiology of breast cancer in Indian womenrdquo Asia-PacificJournal of Clinical Oncology vol 13 no 4 pp 289ndash295 2017

[2] A Gupta K Shridhar and P Dhillon ldquoA review of breastcancer awareness among women in India cancer literate orawareness deficitrdquo European Journal of Cancer vol 51no 14 pp 2058ndash2066 2015

[3] K P Kanadam and S R Chereddy ldquoMammogram classifi-cation using sparse-ROI a novel representation to arbitraryshaped massesrdquo Expert Systems with Applications vol 57pp 204ndash213 2016

[4] D O T Bruno M Z do Nascimento R P Ramos et al ldquoLBPoperators on curvelet coefficients as an algorithm to describetexture in breast cancer tissuesrdquo Expert Systems with Appli-cations vol 55 pp 329ndash340 2016

[5] M Hussain ldquoFalse-positive reduction in mammographyusing multiscale spatial Weber law descriptor and supportvector machinesrdquo Neural Computing and Applicationsvol 25 no 1 pp 83ndash93 2014

[6] M M Pawar and S N Talbar ldquoGenetic fuzzy system (GFS)based wavelet co-occurrence feature selection in mammo-gram classification for breast cancer diagnosisrdquo Perspectives inScience vol 8 pp 247ndash250 2016

[7] C Muramatsu T Hara T Endo and H Fujita ldquoBreast massclassification on mammograms using radial local ternarypatternsrdquo Computers in Biology and Medicine vol 72pp 43ndash53 2016

[8] M M Eltoukhy I Faye and B B Samir ldquoA statistical basedfeature extraction method for breast cancer diagnosis indigital mammogram using multiresolution representationrdquo

Table 8 Comparison of classification accuracy AUC and FPimage values from different approaches in breast cancer diagnosis

Author Database Method Classifier Result AUC FPimageEltoukhy et al [33]

MIAS Biggest curvelet coefficients as a featurevector

Euclidean classifier 9407 mdash mdashEltoukhy et al [42] 9859 mdash mdashEltoukhy et al [8] SVM 973 mdash mdash

Dhahbi et al [34] Mini-MIAS Curvelet moments KNN 9127 mdash mdashDDSM 8646 mdash mdash

Bruno et al [4] DDSM Curvelet + LBP SVM 85 085 mdashPL 94 094 mdash

da Rocha et al [40] DDSM LBP SVM 8831 088 mdashKanadam andChereddy [3] MIAS Sparse ROI SVM 9742 mdash mdash

Pereira et al [18] DDSM Wavelet and Wiener filter Multiple thresholding waveletand GA mdash mdash 137

Liu and Zeng [29] DDSMFFDM GLCM CLBP and geometric features SVM mdash mdash 148

De Sampaio et al[39] DDSM LBP DBSCAN 9826 019

Zyout et al [30] DDSM Second order statistics of waveletcoefficients (SOSWC) SVM 968 097 0018

MIAS 952 966 0029

Casti et al [31]DDSM

Differential features Fisher linear discriminantanalysis (FLDA) mdash mdash

168MIAS 212FFDM 082

Proposed method

MIAS

LBP based on sparse curvelet subbandcoefficients ANN

9857 098 001DDSM 9870 098 003TMCHScanner1 9830 098 005

TMCHScanner2 100 1 0

14 Journal of Healthcare Engineering

depends upon the shape of the ROI as per the sparsematrix Tables 2 and 3 do not represent exact reduction inpixels for complete database but they exhibit pixel re-duction for sample mammograms

43 Classifier Evaluation and False-Positive ReductionFrom Figures 10ndash13 the best classification accuracy of 9857 has been obtained for MIAS in benign versus malignantclassification whereas 9870 for DDSM 9830 for

(1) Load input image (img1)(2) Apply CLAHE algorithm and obtain enhanced image (img2)(3) Decompose img1 and img2 up to 3 level of decomposition using Discrete Wavelet transform (DWT)(4) Use maximum local entropy rule for fusion of img1 and img2 for high frequency subbands(5) Take inverse DWT to obtain the fused image

ALGORITHM 1 Image fusion for contrast enhancement

Input I(m n) m no of rows and nno of columnOutput LBP featuresInitialize Radius R 1 and neighborhood pixels P 8

Mask [1 2 4 8 16 32 64 128 0]Sliding window coordinates kminus1 1Count 1 number of pixels in I(m n)

for i 1 to m dofor j 1 to n do

prepare local circular windowI_local I(i+ k j+ k)center_pixel I(i j)Arrange local neighborhoods of I(i j) pixels in a row col 1 9Lookup_table (i col) reshape(I_local [7 17])count number of pixels greater than zeroa length(find(Lookup_tablegt 0))select pixel position from lookup-table for computation of LBPif agt 2

LBP_code(count) I_localgt center_pixelcount count + 1

endend

endcompute histogram of LBP codesLBP_descriptor LBP_descriptorcountscale invariant

ALGORITHM 2 Algorithm for LBP feature computation based on shape of mass in ROI as

(1) Load input image (img1)(2) Apply CLAHE algorithm and obtain enhanced image (img2)(3) Process img1 and img2 and obtain enhanced image using procedure given in Algorithm 1(4) Remove pectoral muscle using proposed approach (Section 312)(5) Extract neighbourhood features for each pixel and apply SOM clustering(6) Obtain clustered image and separate out the tumorous cluster(7) Extract detected regions ie ROIrsquos from clustered result(8) Extract Sparse Curvelet Coefficients (Subband) up to 2 level from each ROI(9) Extract Sparse LBP code for each subband and obtain a combined feature vector for each ROI(10) Classify each ROI into tumorous and nontumorous class ie TP and FP respectively(11) Map each TP region on original mammogram (img1)(12) end

ALGORITHM 3 Summary of proposed method for FP reduction in mammograms

8 Journal of Healthcare Engineering

TMCH Scanner1 and 100 for TMCH Scanner2 clas-sication accuracies have been obtained in normal versusmalignant classication e classication performance ofANN has improved from 6 to 43 for diyenerent databasesas compared to KNN classier whereas there is littleimprovement about 7 compared with SVM classier eperformances of both proposed sparse LBP and LBPcomputation on curvelet subbands are nearly sametherefore the proposed algorithm can be eiexclciently

implemented in CAD system with lesser number of cur-velet coeiexclcients

Data augmentation has been used for some classes tomaintain balance between two classes to improve perfor-mance and to learn more powerful model Table 4 explainsthe FP reduction with the use of curvelet-based LBP featuresand ANN It has been observed that FP reduced from 085 to002 FPimage in MIAS 481 to 002 FPimage in DDSM and232 to 013 FPimage in TMCH

8000

8500

9000

9500

10000

TMCHS1LBP

TMCHS1sparse_LBP

TMCHS2LBP

TMCHS2sparse_LBP

Normal_Malignant

KNNANNSVM

Figure 10 Average classication rate for TMCH dataset

000

2000

4000

6000

8000

10000

DDSMLBP DDSMsparse_LBP

MIASLBP MIASsparse_LBP

Normal_Malignant

KNNANNSVM

Figure 11 Average classication rate for MIAS and DDSM dataset

000

2000

4000

6000

8000

10000

DDSMLBP DDSMsparse_LBP

MIASLBP MIASsparse_LBP

Benign_Malignant

KNNANNSVM

Figure 12 Average classication rate for MIAS and DDSM dataset

Journal of Healthcare Engineering 9

Similarly Table 5 shows the reduction in FPs as 085to 001 FPimage for MIAS 481 to 003 FPimage forDDSM and 232 to 000 FPimage for TMCH using sparsecurvelet coeiexclcient-based LBP features e results show

the eyenectiveness of sparse curvelet coeiexclcient-based LBPand ANN From Table 6 the best value of AUC 099 isobtained in benign versus malignant classication forMIAS AUC 098 in benign versus malignant in case of

Table 1 Result of SOM segmentation

Datasetused

Result of SOM clustering and threshold TPR (true-positiverate)

TPlesions

FPPI (false-positiveper image)FPimages

MassSegmented nonmass (FP) Total () images

Segmented (TP) LostMIAS 108 7 273 322 (108115) 094 (273322) 085DDSM 140 10 1203 250 (140150) 093 (1203250) 481TMCH 172 8 837 360 (172180) 095 (837360) 232

Table 2 Reduction in curvelet coeiexclcients for sample mammograms from MIAS and DDSM dataset

SrNo

MIAS DDSM

ROI Size

Total number ofcurvelet

coeiexclcients fromsubbands

Total number ofselected curveletcoeiexclcients from

subbands

reductionin curveletcoeiexclcients

ROI Size

Total number ofcurvelet

coeiexclcients fromsubbands

Total number ofselected curveletcoeiexclcients from

subbands

reductionin curveletcoeiexclcients

1 124times138 103911 77133 2577 192times187 216729 168333 22332 179times138 150123 126142 1597 294times 291 518267 366680 29253 51times 116 36815 33421 922 145times 207 181663 148765 18114 83times 83 42449 36653 1365 169times168 171873 154752 9965 84times 76 39115 35815 844 182times 248 272517 219822 19346 74times 83 37767 34448 879 213times 349 449783 301359 33007 53times 64 20969 18621 1120 578times 412 1433195 662072 53808 70times 44 18899 16610 1211 215times 219 286429 242935 15189 80times 66 32409 29454 912 420times 428 1082461 501209 537010 69times 86 36783 33552 878 203times 307 375871 266829 290111 59times116 42019 38442 851 226times 262 357209 276763 225212 81times 101 50637 46122 892 159times194 187563 148741 207013 41times 84 21427 18907 1176 718times 686 2961127 762561 742514 69times141 60641 52427 1354 409times 550 1352439 904829 331015 60times 62 22925 20475 1069 524times 375 1184671 766522 353016 96times101 59647 54235 907 311times 275 517433 397737 231317 136times139 113727 66352 4156 319times 320 614129 412606 328118 55times 94 31359 28305 974 313times 447 842855 510482 394319 157times140 132373 107000 1917 291times 517 907903 637412 297920 156times130 123007 90834 2615 370times 837 1864889 898236 5183

Average 58850 48247 14 Average 788950 437432 32

000

2000

4000

6000

8000

10000

DDSMLBP DDSM sparse_LBP

MIASLBP MIAS sparse_LBP

Normal_Abnormal

KNNANNSVM

Figure 13 Average classication rate for MIAS and DDSM dataset

10 Journal of Healthcare Engineering

DDSM AUC 094 in normal versus malignant in case ofTMCH Scanner1 and AUC 096 in normal versusmalignant classification in TMCH Scanner2 using ANNand curvelet subband-based LBP features e worstperformance of AUC 053 for MIAS is obtained with theproposed algorithm using KNN classifier as shown inTable 7 Similarly from Table 7 the best value ofAUC 098 is obtained in TMCH Scanner1 AUC 1 isobtained in TMCH Scanner2 database for normal versusmalignant classification AUC 098 in benign versusmalignant classification is attained in MIAS database andAUC 098 is achieved for normal versus malignant

classification in DDSM database using ANN classifier forsparse curvelet subband-based LBP features

However from Table 7 it should be noted that the per-formance of proposed algorithm is the best using ANN clas-sifier Figure 14 represents automated CAD system for breastcancer diagnosis with sample mammograms

Table 8 provides comparative study of methods developedfor breast tissue classification e proposed method providesbest results in terms of AUC and reduction of number of FPsas 085 to 001 FPimage for MIAS 481 to 003 FPimage forDDSM and 232 to 000 FPimage for TMCH e earlierreported work uses the fixed patch size-based approach which

Table 3 Reduction in curvelet coefficients for sample mammograms from TMCH Scanner1 and Scanner2 dataset

Srno

TMCH Scanner 1 ldquoGE Medical Senograph Systemrdquo TMCH Scanner 2 ldquoHologic Selenia Systemrdquo

ROI size

Total number ofcurvelet

coefficients fromsubbands

Total number ofselected curveletcoefficients from

subbands

reductionin curveletcoefficients

ROI size

Total number ofcurvelet

coefficients fromsubbands

Total number ofselected curveletcoefficients from

subbands

reductionin curveletcoefficients

1 459times 412 1140617 794885 3031 291times 278 489243 317014 35202 548times 513 1693403 1117873 3399 545times 246 809627 531604 34343 415times 303 758323 445585 4124 560times 483 1630583 1068439 34474 645times 495 1951443 1145580 4129 782times 510 2400137 1275073 46875 437times 691 1816651 985120 4577 87times141 75773 67871 10436 812times 500 2439937 1287065 4725 311times 185 348565 240546 30997 468times 379 1066333 710242 3340 262times 348 550303 282821 48618 673times 582 2355589 1730915 2652 610times 440 1614515 842876 47799 250times 201 304513 235670 2261 949times 391 2227209 1359338 389710 525times 488 1544691 1161942 2478 365times 385 846523 604473 285911 488times 779 2287547 1611385 2956 393times 247 585063 490474 161712 1434times 966 8326581 3742277 5506 341times 301 618111 410542 335813 348times 421 881701 616501 3008 523times 702 2206097 800057 637314 460times 530 1467227 799885 4548 370times 284 632539 423727 330115 398times 450 1078441 828064 2322 344times 202 418955 302427 278116 247times 272 404401 344822 1473 264times188 299997 231926 226917 411times 305 757657 405919 4642 233times 247 346983 267686 228518 286times 344 593155 448566 2438 370times 291 650701 486125 252919 417times 207 523477 429755 1790 680times 483 1979543 1052793 468220 463times 458 1275021 857608 3274 202times 266 324295 215042 3369

Average 1633335 984983 33 Average 952738 563543 34

Table 4 Number of ROIs resulted in FP reduction using curvelet-based LBP (without sparse) amp ANN classification at training andvalidation stage

Class Datasetused

Benignmalignant mass Nonmassbenign mass Total()

images

TPR (true-positiverate)TPlesions

FPPI (false-positiveper image) FP

imagesPreviousstage

Selected(TP)

Lost(FN)

Previousstage

Selected(TN)

Lost(FP)

Normal vsabnormal

MIAS 108lowast 2 216 203 13 273 257 16 315 (203216) 094 (16315) 005DDSM 140lowast 4 560 465 95 1203 1095 108 240 (465560) 083 (108240) 045

Benign vsmalignant

MIAS 49 49 0 59 57 2 108 (4949) 100 (2108) 002DDSM 46lowast 2 92 91 1 94 91 3 140 (9192) 099 (3140) 002

Normal vsmalignant

MIAS 49lowast 4196 184 12 273 254 19 256 (184196) 094 (19256) 007DDSM 46lowast 4184 180 4 1203 1143 60 146 (180184) 098 (60146) 041TMCHScanner1 107lowast 4 428 416 12 605 551 54 217 (416428) 097 (54217) 025

TMCHScanner2 65lowast 4 260 255 5 232 214 18 135 (255260) 098 (18135) 013

lowastAugmentation of image

Journal of Healthcare Engineering 11

limits the automatic CAD system scope whereas proposedsystem provides complete solution to CAD system right fromautomatic tumor patch segmentation to reduction in FPs andfinal representation of mammogram with TP marked on itIt will drastically reduce the radiologist work by locationtumor directly on mammogram

5 Conclusion

A fully automatic CAD system which can accuratelylocate the tumor on a mammogram and reduces FPshas been proposed e developed CAD system consistsof preprocessing SOM clustering ROI extraction

Table 6 Performance evaluation of curvelet-based LBP descriptor algorithm

Dataset Classification Normal-malignant Normal-abnormal Benign-malignantClassifier Sensitivity Specificity AUC Sensitivity Specificity AUC Sensitivity Specificity AUC

MIASANN 094 093 094 094 094 094 100 097 099SVM 085 085 085 083 086 085 088 084 086KNN 067 063 065 058 057 058 062 068 063

DDSMANN 098 095 095 083 091 085 099 097 098SVM 097 088 092 071 091 083 094 089 092KNN 096 064 087 067 090 080 087 073 079

TMCH Scanner1ANN 097 091 094 mdash mdash mdash mdash mdash mdashSVM 096 091 094 mdash mdash mdash mdash mdash mdashKNN 098 082 089 mdash mdash mdash mdash mdash mdash

TMCH Scanner2ANN 098 092 096 mdash mdash mdash mdash mdash mdashSVM 097 090 094 mdash mdash mdash mdash mdash mdashKNN 092 083 088 mdash mdash mdash mdash mdash mdash

Table 7 Performance evaluation of proposed algorithm

Dataset Classification Normal-malignant Normal-abnormal Benign-malignantClassifier Sensitivity Specificity AUC Sensitivity Specificity AUC Sensitivity Specificity AUC

MIASANN 098 095 096 093 097 095 097 100 098SVM 088 083 085 085 082 084 084 092 087KNN 055 051 053 055 063 056 061 067 061

DDSMANN 099 097 098 092 096 093 097 095 096SVM 099 092 096 089 096 092 094 092 093KNN 098 073 092 074 090 082 089 077 083

TMCH Scanner1ANN 099 098 098 mdash mdash mdash mdash mdash mdashSVM 098 096 097 mdash mdash mdash mdash mdash mdashKNN 099 092 095 mdash mdash mdash mdash mdash mdash

TMCH Scanner2ANN 100 100 100 mdash mdash mdash mdash mdash mdashSVM 100 098 099 mdash mdash mdash mdash mdash mdashKNN 096 092 094 mdash mdash mdash mdash mdash mdash

Table 5 Number of ROIs resulted in FP reduction using sparse curvelet coefficient-based LBP amp ANN classification at training andvalidation stage

Class Datasetused

Benignmalignant mass Nonmassbenign massTotal ()images

TPR (true-positiverate)TPlesions

FPPI (false-positiveper image) FP

imagesPreviousstage

Selected(TP)

Lost(FN)

Previousstage

Selected(TN)

Lost(FP)

Normal vsabnormal

MIAS 108lowast 2 216 201 15 273 265 8 315 (201216) 093 (8315) 002DDSM 140lowast 4 560 516 44 1203 1155 48 240 (516560) 092 (48240) 02

Benign vsmalignant

MIAS 49 48 1 59 59 1 108 (4849) 098 (1108) 001DDSM 46lowast 2 92 89 3 94 89 5 140 (8992) 097 (5140) 003

Normal vsmalignant

MIAS 49lowast 4196 192 4 273 259 14 256 (192196) 098 (14256) 005DDSM 46lowast 4184 182 2 1203 1167 36 146 (182184) 099 (36146) 025TMCHScanner1 107lowast 4 428 424 4 605 593 12 217 (424428) 099 (12217) 005

TMCHScanner2 65lowast 4 260 260 0 232 232 0 135 (260260) 100 (0135) 0

lowastAugmentation of image

12 Journal of Healthcare Engineering

sparse LBP feature computation based on sparse Cur-velet coefficients and finally FP reduction using ANNclassifier

e proposed algorithm presents a novel concept ofextraction of curvelet coefficients according to irregularshape of mass is called as sparse curvelet coefficients andcomputation of LBP e analysis proves that the FPs arereduced significantly from 085 to 001 FPimage for MIAS

481 to 003 FPimage for DDSM and 232 to 000 FPimagefor TMCH e ANN classifier showed best results asAUC 098 and accuracy 9857 for MIAS in benign-malignant classification AUC 098 and accuracy 9870for DDSM in normal-malignant classification AUC 098and accuracy 9830 for TMCH Scanner1 and AUC 1and accuracy 100 for TMCH Scanner2 in normal-malignant classification as compared with SVM and KNN

(a) (b) (c) (d) (e) (f )

Figure 14 Representation of fully automatic CAD system for breast cancer using (a) samplemammograms fromMIAS DDSM and TMCHdatasets (b) preprocessed mammograms (c) clustered image (d) TP and FP marked on mammogram (e) TP marked by thresholding (f )TP marked by using LBP descriptor based on sparse curvelet coefficients

Journal of Healthcare Engineering 13

classifier e performance of LBP features and LBP featuresbased on sparse curvelet coefficients are nearly same whichshow that the proposed algorithm is suitable for cancer breasttissue diagnosis

In future the reduced curvelet coefficients can be used toextract local ternary patterns and other local descriptor andlocal directional patterns etc e present work deals withmammogram with single mass this can be further extendedfor multiple mass models with multiple LBP features basedon sparse curvelet coefficients

Data Availability

In this research we have used two publicly available datasetsMIAS and DDSM ese datasets can be found here in [17]and [19] e third database is collected from the localhospital Tata Memorial Cancer Hospital Mumbai whichcan be found at httpeurekasveriacin or available fromthe corresponding author upon request

Conflicts of Interest

e authors declare that they have no conflicts of interest

Acknowledgments

e TMCH database for this work was given by Departmentof Radiodiagnosis Tata Memorial Cancer Hospital Mumbai

References

[1] S Malvia S A Bagadi U S Dubey and S Saxena ldquoEpi-demiology of breast cancer in Indian womenrdquo Asia-PacificJournal of Clinical Oncology vol 13 no 4 pp 289ndash295 2017

[2] A Gupta K Shridhar and P Dhillon ldquoA review of breastcancer awareness among women in India cancer literate orawareness deficitrdquo European Journal of Cancer vol 51no 14 pp 2058ndash2066 2015

[3] K P Kanadam and S R Chereddy ldquoMammogram classifi-cation using sparse-ROI a novel representation to arbitraryshaped massesrdquo Expert Systems with Applications vol 57pp 204ndash213 2016

[4] D O T Bruno M Z do Nascimento R P Ramos et al ldquoLBPoperators on curvelet coefficients as an algorithm to describetexture in breast cancer tissuesrdquo Expert Systems with Appli-cations vol 55 pp 329ndash340 2016

[5] M Hussain ldquoFalse-positive reduction in mammographyusing multiscale spatial Weber law descriptor and supportvector machinesrdquo Neural Computing and Applicationsvol 25 no 1 pp 83ndash93 2014

[6] M M Pawar and S N Talbar ldquoGenetic fuzzy system (GFS)based wavelet co-occurrence feature selection in mammo-gram classification for breast cancer diagnosisrdquo Perspectives inScience vol 8 pp 247ndash250 2016

[7] C Muramatsu T Hara T Endo and H Fujita ldquoBreast massclassification on mammograms using radial local ternarypatternsrdquo Computers in Biology and Medicine vol 72pp 43ndash53 2016

[8] M M Eltoukhy I Faye and B B Samir ldquoA statistical basedfeature extraction method for breast cancer diagnosis indigital mammogram using multiresolution representationrdquo

Table 8 Comparison of classification accuracy AUC and FPimage values from different approaches in breast cancer diagnosis

Author Database Method Classifier Result AUC FPimageEltoukhy et al [33]

MIAS Biggest curvelet coefficients as a featurevector

Euclidean classifier 9407 mdash mdashEltoukhy et al [42] 9859 mdash mdashEltoukhy et al [8] SVM 973 mdash mdash

Dhahbi et al [34] Mini-MIAS Curvelet moments KNN 9127 mdash mdashDDSM 8646 mdash mdash

Bruno et al [4] DDSM Curvelet + LBP SVM 85 085 mdashPL 94 094 mdash

da Rocha et al [40] DDSM LBP SVM 8831 088 mdashKanadam andChereddy [3] MIAS Sparse ROI SVM 9742 mdash mdash

Pereira et al [18] DDSM Wavelet and Wiener filter Multiple thresholding waveletand GA mdash mdash 137

Liu and Zeng [29] DDSMFFDM GLCM CLBP and geometric features SVM mdash mdash 148

De Sampaio et al[39] DDSM LBP DBSCAN 9826 019

Zyout et al [30] DDSM Second order statistics of waveletcoefficients (SOSWC) SVM 968 097 0018

MIAS 952 966 0029

Casti et al [31]DDSM

Differential features Fisher linear discriminantanalysis (FLDA) mdash mdash

168MIAS 212FFDM 082

Proposed method

MIAS

LBP based on sparse curvelet subbandcoefficients ANN

9857 098 001DDSM 9870 098 003TMCHScanner1 9830 098 005

TMCHScanner2 100 1 0

14 Journal of Healthcare Engineering

TMCH Scanner1 and 100 for TMCH Scanner2 clas-sication accuracies have been obtained in normal versusmalignant classication e classication performance ofANN has improved from 6 to 43 for diyenerent databasesas compared to KNN classier whereas there is littleimprovement about 7 compared with SVM classier eperformances of both proposed sparse LBP and LBPcomputation on curvelet subbands are nearly sametherefore the proposed algorithm can be eiexclciently

implemented in CAD system with lesser number of cur-velet coeiexclcients

Data augmentation has been used for some classes tomaintain balance between two classes to improve perfor-mance and to learn more powerful model Table 4 explainsthe FP reduction with the use of curvelet-based LBP featuresand ANN It has been observed that FP reduced from 085 to002 FPimage in MIAS 481 to 002 FPimage in DDSM and232 to 013 FPimage in TMCH

8000

8500

9000

9500

10000

TMCHS1LBP

TMCHS1sparse_LBP

TMCHS2LBP

TMCHS2sparse_LBP

Normal_Malignant

KNNANNSVM

Figure 10 Average classication rate for TMCH dataset

000

2000

4000

6000

8000

10000

DDSMLBP DDSMsparse_LBP

MIASLBP MIASsparse_LBP

Normal_Malignant

KNNANNSVM

Figure 11 Average classication rate for MIAS and DDSM dataset

000

2000

4000

6000

8000

10000

DDSMLBP DDSMsparse_LBP

MIASLBP MIASsparse_LBP

Benign_Malignant

KNNANNSVM

Figure 12 Average classication rate for MIAS and DDSM dataset

Journal of Healthcare Engineering 9

Similarly Table 5 shows the reduction in FPs as 085to 001 FPimage for MIAS 481 to 003 FPimage forDDSM and 232 to 000 FPimage for TMCH using sparsecurvelet coeiexclcient-based LBP features e results show

the eyenectiveness of sparse curvelet coeiexclcient-based LBPand ANN From Table 6 the best value of AUC 099 isobtained in benign versus malignant classication forMIAS AUC 098 in benign versus malignant in case of

Table 1 Result of SOM segmentation

Datasetused

Result of SOM clustering and threshold TPR (true-positiverate)

TPlesions

FPPI (false-positiveper image)FPimages

MassSegmented nonmass (FP) Total () images

Segmented (TP) LostMIAS 108 7 273 322 (108115) 094 (273322) 085DDSM 140 10 1203 250 (140150) 093 (1203250) 481TMCH 172 8 837 360 (172180) 095 (837360) 232

Table 2 Reduction in curvelet coeiexclcients for sample mammograms from MIAS and DDSM dataset

SrNo

MIAS DDSM

ROI Size

Total number ofcurvelet

coeiexclcients fromsubbands

Total number ofselected curveletcoeiexclcients from

subbands

reductionin curveletcoeiexclcients

ROI Size

Total number ofcurvelet

coeiexclcients fromsubbands

Total number ofselected curveletcoeiexclcients from

subbands

reductionin curveletcoeiexclcients

1 124times138 103911 77133 2577 192times187 216729 168333 22332 179times138 150123 126142 1597 294times 291 518267 366680 29253 51times 116 36815 33421 922 145times 207 181663 148765 18114 83times 83 42449 36653 1365 169times168 171873 154752 9965 84times 76 39115 35815 844 182times 248 272517 219822 19346 74times 83 37767 34448 879 213times 349 449783 301359 33007 53times 64 20969 18621 1120 578times 412 1433195 662072 53808 70times 44 18899 16610 1211 215times 219 286429 242935 15189 80times 66 32409 29454 912 420times 428 1082461 501209 537010 69times 86 36783 33552 878 203times 307 375871 266829 290111 59times116 42019 38442 851 226times 262 357209 276763 225212 81times 101 50637 46122 892 159times194 187563 148741 207013 41times 84 21427 18907 1176 718times 686 2961127 762561 742514 69times141 60641 52427 1354 409times 550 1352439 904829 331015 60times 62 22925 20475 1069 524times 375 1184671 766522 353016 96times101 59647 54235 907 311times 275 517433 397737 231317 136times139 113727 66352 4156 319times 320 614129 412606 328118 55times 94 31359 28305 974 313times 447 842855 510482 394319 157times140 132373 107000 1917 291times 517 907903 637412 297920 156times130 123007 90834 2615 370times 837 1864889 898236 5183

Average 58850 48247 14 Average 788950 437432 32

000

2000

4000

6000

8000

10000

DDSMLBP DDSM sparse_LBP

MIASLBP MIAS sparse_LBP

Normal_Abnormal

KNNANNSVM

Figure 13 Average classication rate for MIAS and DDSM dataset

10 Journal of Healthcare Engineering

DDSM AUC 094 in normal versus malignant in case ofTMCH Scanner1 and AUC 096 in normal versusmalignant classification in TMCH Scanner2 using ANNand curvelet subband-based LBP features e worstperformance of AUC 053 for MIAS is obtained with theproposed algorithm using KNN classifier as shown inTable 7 Similarly from Table 7 the best value ofAUC 098 is obtained in TMCH Scanner1 AUC 1 isobtained in TMCH Scanner2 database for normal versusmalignant classification AUC 098 in benign versusmalignant classification is attained in MIAS database andAUC 098 is achieved for normal versus malignant

classification in DDSM database using ANN classifier forsparse curvelet subband-based LBP features

However from Table 7 it should be noted that the per-formance of proposed algorithm is the best using ANN clas-sifier Figure 14 represents automated CAD system for breastcancer diagnosis with sample mammograms

Table 8 provides comparative study of methods developedfor breast tissue classification e proposed method providesbest results in terms of AUC and reduction of number of FPsas 085 to 001 FPimage for MIAS 481 to 003 FPimage forDDSM and 232 to 000 FPimage for TMCH e earlierreported work uses the fixed patch size-based approach which

Table 3 Reduction in curvelet coefficients for sample mammograms from TMCH Scanner1 and Scanner2 dataset

Srno

TMCH Scanner 1 ldquoGE Medical Senograph Systemrdquo TMCH Scanner 2 ldquoHologic Selenia Systemrdquo

ROI size

Total number ofcurvelet

coefficients fromsubbands

Total number ofselected curveletcoefficients from

subbands

reductionin curveletcoefficients

ROI size

Total number ofcurvelet

coefficients fromsubbands

Total number ofselected curveletcoefficients from

subbands

reductionin curveletcoefficients

1 459times 412 1140617 794885 3031 291times 278 489243 317014 35202 548times 513 1693403 1117873 3399 545times 246 809627 531604 34343 415times 303 758323 445585 4124 560times 483 1630583 1068439 34474 645times 495 1951443 1145580 4129 782times 510 2400137 1275073 46875 437times 691 1816651 985120 4577 87times141 75773 67871 10436 812times 500 2439937 1287065 4725 311times 185 348565 240546 30997 468times 379 1066333 710242 3340 262times 348 550303 282821 48618 673times 582 2355589 1730915 2652 610times 440 1614515 842876 47799 250times 201 304513 235670 2261 949times 391 2227209 1359338 389710 525times 488 1544691 1161942 2478 365times 385 846523 604473 285911 488times 779 2287547 1611385 2956 393times 247 585063 490474 161712 1434times 966 8326581 3742277 5506 341times 301 618111 410542 335813 348times 421 881701 616501 3008 523times 702 2206097 800057 637314 460times 530 1467227 799885 4548 370times 284 632539 423727 330115 398times 450 1078441 828064 2322 344times 202 418955 302427 278116 247times 272 404401 344822 1473 264times188 299997 231926 226917 411times 305 757657 405919 4642 233times 247 346983 267686 228518 286times 344 593155 448566 2438 370times 291 650701 486125 252919 417times 207 523477 429755 1790 680times 483 1979543 1052793 468220 463times 458 1275021 857608 3274 202times 266 324295 215042 3369

Average 1633335 984983 33 Average 952738 563543 34

Table 4 Number of ROIs resulted in FP reduction using curvelet-based LBP (without sparse) amp ANN classification at training andvalidation stage

Class Datasetused

Benignmalignant mass Nonmassbenign mass Total()

images

TPR (true-positiverate)TPlesions

FPPI (false-positiveper image) FP

imagesPreviousstage

Selected(TP)

Lost(FN)

Previousstage

Selected(TN)

Lost(FP)

Normal vsabnormal

MIAS 108lowast 2 216 203 13 273 257 16 315 (203216) 094 (16315) 005DDSM 140lowast 4 560 465 95 1203 1095 108 240 (465560) 083 (108240) 045

Benign vsmalignant

MIAS 49 49 0 59 57 2 108 (4949) 100 (2108) 002DDSM 46lowast 2 92 91 1 94 91 3 140 (9192) 099 (3140) 002

Normal vsmalignant

MIAS 49lowast 4196 184 12 273 254 19 256 (184196) 094 (19256) 007DDSM 46lowast 4184 180 4 1203 1143 60 146 (180184) 098 (60146) 041TMCHScanner1 107lowast 4 428 416 12 605 551 54 217 (416428) 097 (54217) 025

TMCHScanner2 65lowast 4 260 255 5 232 214 18 135 (255260) 098 (18135) 013

lowastAugmentation of image

Journal of Healthcare Engineering 11

limits the automatic CAD system scope whereas proposedsystem provides complete solution to CAD system right fromautomatic tumor patch segmentation to reduction in FPs andfinal representation of mammogram with TP marked on itIt will drastically reduce the radiologist work by locationtumor directly on mammogram

5 Conclusion

A fully automatic CAD system which can accuratelylocate the tumor on a mammogram and reduces FPshas been proposed e developed CAD system consistsof preprocessing SOM clustering ROI extraction

Table 6 Performance evaluation of curvelet-based LBP descriptor algorithm

Dataset Classification Normal-malignant Normal-abnormal Benign-malignantClassifier Sensitivity Specificity AUC Sensitivity Specificity AUC Sensitivity Specificity AUC

MIASANN 094 093 094 094 094 094 100 097 099SVM 085 085 085 083 086 085 088 084 086KNN 067 063 065 058 057 058 062 068 063

DDSMANN 098 095 095 083 091 085 099 097 098SVM 097 088 092 071 091 083 094 089 092KNN 096 064 087 067 090 080 087 073 079

TMCH Scanner1ANN 097 091 094 mdash mdash mdash mdash mdash mdashSVM 096 091 094 mdash mdash mdash mdash mdash mdashKNN 098 082 089 mdash mdash mdash mdash mdash mdash

TMCH Scanner2ANN 098 092 096 mdash mdash mdash mdash mdash mdashSVM 097 090 094 mdash mdash mdash mdash mdash mdashKNN 092 083 088 mdash mdash mdash mdash mdash mdash

Table 7 Performance evaluation of proposed algorithm

Dataset Classification Normal-malignant Normal-abnormal Benign-malignantClassifier Sensitivity Specificity AUC Sensitivity Specificity AUC Sensitivity Specificity AUC

MIASANN 098 095 096 093 097 095 097 100 098SVM 088 083 085 085 082 084 084 092 087KNN 055 051 053 055 063 056 061 067 061

DDSMANN 099 097 098 092 096 093 097 095 096SVM 099 092 096 089 096 092 094 092 093KNN 098 073 092 074 090 082 089 077 083

TMCH Scanner1ANN 099 098 098 mdash mdash mdash mdash mdash mdashSVM 098 096 097 mdash mdash mdash mdash mdash mdashKNN 099 092 095 mdash mdash mdash mdash mdash mdash

TMCH Scanner2ANN 100 100 100 mdash mdash mdash mdash mdash mdashSVM 100 098 099 mdash mdash mdash mdash mdash mdashKNN 096 092 094 mdash mdash mdash mdash mdash mdash

Table 5 Number of ROIs resulted in FP reduction using sparse curvelet coefficient-based LBP amp ANN classification at training andvalidation stage

Class Datasetused

Benignmalignant mass Nonmassbenign massTotal ()images

TPR (true-positiverate)TPlesions

FPPI (false-positiveper image) FP

imagesPreviousstage

Selected(TP)

Lost(FN)

Previousstage

Selected(TN)

Lost(FP)

Normal vsabnormal

MIAS 108lowast 2 216 201 15 273 265 8 315 (201216) 093 (8315) 002DDSM 140lowast 4 560 516 44 1203 1155 48 240 (516560) 092 (48240) 02

Benign vsmalignant

MIAS 49 48 1 59 59 1 108 (4849) 098 (1108) 001DDSM 46lowast 2 92 89 3 94 89 5 140 (8992) 097 (5140) 003

Normal vsmalignant

MIAS 49lowast 4196 192 4 273 259 14 256 (192196) 098 (14256) 005DDSM 46lowast 4184 182 2 1203 1167 36 146 (182184) 099 (36146) 025TMCHScanner1 107lowast 4 428 424 4 605 593 12 217 (424428) 099 (12217) 005

TMCHScanner2 65lowast 4 260 260 0 232 232 0 135 (260260) 100 (0135) 0

lowastAugmentation of image

12 Journal of Healthcare Engineering

sparse LBP feature computation based on sparse Cur-velet coefficients and finally FP reduction using ANNclassifier

e proposed algorithm presents a novel concept ofextraction of curvelet coefficients according to irregularshape of mass is called as sparse curvelet coefficients andcomputation of LBP e analysis proves that the FPs arereduced significantly from 085 to 001 FPimage for MIAS

481 to 003 FPimage for DDSM and 232 to 000 FPimagefor TMCH e ANN classifier showed best results asAUC 098 and accuracy 9857 for MIAS in benign-malignant classification AUC 098 and accuracy 9870for DDSM in normal-malignant classification AUC 098and accuracy 9830 for TMCH Scanner1 and AUC 1and accuracy 100 for TMCH Scanner2 in normal-malignant classification as compared with SVM and KNN

(a) (b) (c) (d) (e) (f )

Figure 14 Representation of fully automatic CAD system for breast cancer using (a) samplemammograms fromMIAS DDSM and TMCHdatasets (b) preprocessed mammograms (c) clustered image (d) TP and FP marked on mammogram (e) TP marked by thresholding (f )TP marked by using LBP descriptor based on sparse curvelet coefficients

Journal of Healthcare Engineering 13

classifier e performance of LBP features and LBP featuresbased on sparse curvelet coefficients are nearly same whichshow that the proposed algorithm is suitable for cancer breasttissue diagnosis

In future the reduced curvelet coefficients can be used toextract local ternary patterns and other local descriptor andlocal directional patterns etc e present work deals withmammogram with single mass this can be further extendedfor multiple mass models with multiple LBP features basedon sparse curvelet coefficients

Data Availability

In this research we have used two publicly available datasetsMIAS and DDSM ese datasets can be found here in [17]and [19] e third database is collected from the localhospital Tata Memorial Cancer Hospital Mumbai whichcan be found at httpeurekasveriacin or available fromthe corresponding author upon request

Conflicts of Interest

e authors declare that they have no conflicts of interest

Acknowledgments

e TMCH database for this work was given by Departmentof Radiodiagnosis Tata Memorial Cancer Hospital Mumbai

References

[1] S Malvia S A Bagadi U S Dubey and S Saxena ldquoEpi-demiology of breast cancer in Indian womenrdquo Asia-PacificJournal of Clinical Oncology vol 13 no 4 pp 289ndash295 2017

[2] A Gupta K Shridhar and P Dhillon ldquoA review of breastcancer awareness among women in India cancer literate orawareness deficitrdquo European Journal of Cancer vol 51no 14 pp 2058ndash2066 2015

[3] K P Kanadam and S R Chereddy ldquoMammogram classifi-cation using sparse-ROI a novel representation to arbitraryshaped massesrdquo Expert Systems with Applications vol 57pp 204ndash213 2016

[4] D O T Bruno M Z do Nascimento R P Ramos et al ldquoLBPoperators on curvelet coefficients as an algorithm to describetexture in breast cancer tissuesrdquo Expert Systems with Appli-cations vol 55 pp 329ndash340 2016

[5] M Hussain ldquoFalse-positive reduction in mammographyusing multiscale spatial Weber law descriptor and supportvector machinesrdquo Neural Computing and Applicationsvol 25 no 1 pp 83ndash93 2014

[6] M M Pawar and S N Talbar ldquoGenetic fuzzy system (GFS)based wavelet co-occurrence feature selection in mammo-gram classification for breast cancer diagnosisrdquo Perspectives inScience vol 8 pp 247ndash250 2016

[7] C Muramatsu T Hara T Endo and H Fujita ldquoBreast massclassification on mammograms using radial local ternarypatternsrdquo Computers in Biology and Medicine vol 72pp 43ndash53 2016

[8] M M Eltoukhy I Faye and B B Samir ldquoA statistical basedfeature extraction method for breast cancer diagnosis indigital mammogram using multiresolution representationrdquo

Table 8 Comparison of classification accuracy AUC and FPimage values from different approaches in breast cancer diagnosis

Author Database Method Classifier Result AUC FPimageEltoukhy et al [33]

MIAS Biggest curvelet coefficients as a featurevector

Euclidean classifier 9407 mdash mdashEltoukhy et al [42] 9859 mdash mdashEltoukhy et al [8] SVM 973 mdash mdash

Dhahbi et al [34] Mini-MIAS Curvelet moments KNN 9127 mdash mdashDDSM 8646 mdash mdash

Bruno et al [4] DDSM Curvelet + LBP SVM 85 085 mdashPL 94 094 mdash

da Rocha et al [40] DDSM LBP SVM 8831 088 mdashKanadam andChereddy [3] MIAS Sparse ROI SVM 9742 mdash mdash

Pereira et al [18] DDSM Wavelet and Wiener filter Multiple thresholding waveletand GA mdash mdash 137

Liu and Zeng [29] DDSMFFDM GLCM CLBP and geometric features SVM mdash mdash 148

De Sampaio et al[39] DDSM LBP DBSCAN 9826 019

Zyout et al [30] DDSM Second order statistics of waveletcoefficients (SOSWC) SVM 968 097 0018

MIAS 952 966 0029

Casti et al [31]DDSM

Differential features Fisher linear discriminantanalysis (FLDA) mdash mdash

168MIAS 212FFDM 082

Proposed method

MIAS

LBP based on sparse curvelet subbandcoefficients ANN

9857 098 001DDSM 9870 098 003TMCHScanner1 9830 098 005

TMCHScanner2 100 1 0

14 Journal of Healthcare Engineering

Similarly Table 5 shows the reduction in FPs as 085to 001 FPimage for MIAS 481 to 003 FPimage forDDSM and 232 to 000 FPimage for TMCH using sparsecurvelet coeiexclcient-based LBP features e results show

the eyenectiveness of sparse curvelet coeiexclcient-based LBPand ANN From Table 6 the best value of AUC 099 isobtained in benign versus malignant classication forMIAS AUC 098 in benign versus malignant in case of

Table 1 Result of SOM segmentation

Datasetused

Result of SOM clustering and threshold TPR (true-positiverate)

TPlesions

FPPI (false-positiveper image)FPimages

MassSegmented nonmass (FP) Total () images

Segmented (TP) LostMIAS 108 7 273 322 (108115) 094 (273322) 085DDSM 140 10 1203 250 (140150) 093 (1203250) 481TMCH 172 8 837 360 (172180) 095 (837360) 232

Table 2 Reduction in curvelet coeiexclcients for sample mammograms from MIAS and DDSM dataset

SrNo

MIAS DDSM

ROI Size

Total number ofcurvelet

coeiexclcients fromsubbands

Total number ofselected curveletcoeiexclcients from

subbands

reductionin curveletcoeiexclcients

ROI Size

Total number ofcurvelet

coeiexclcients fromsubbands

Total number ofselected curveletcoeiexclcients from

subbands

reductionin curveletcoeiexclcients

1 124times138 103911 77133 2577 192times187 216729 168333 22332 179times138 150123 126142 1597 294times 291 518267 366680 29253 51times 116 36815 33421 922 145times 207 181663 148765 18114 83times 83 42449 36653 1365 169times168 171873 154752 9965 84times 76 39115 35815 844 182times 248 272517 219822 19346 74times 83 37767 34448 879 213times 349 449783 301359 33007 53times 64 20969 18621 1120 578times 412 1433195 662072 53808 70times 44 18899 16610 1211 215times 219 286429 242935 15189 80times 66 32409 29454 912 420times 428 1082461 501209 537010 69times 86 36783 33552 878 203times 307 375871 266829 290111 59times116 42019 38442 851 226times 262 357209 276763 225212 81times 101 50637 46122 892 159times194 187563 148741 207013 41times 84 21427 18907 1176 718times 686 2961127 762561 742514 69times141 60641 52427 1354 409times 550 1352439 904829 331015 60times 62 22925 20475 1069 524times 375 1184671 766522 353016 96times101 59647 54235 907 311times 275 517433 397737 231317 136times139 113727 66352 4156 319times 320 614129 412606 328118 55times 94 31359 28305 974 313times 447 842855 510482 394319 157times140 132373 107000 1917 291times 517 907903 637412 297920 156times130 123007 90834 2615 370times 837 1864889 898236 5183

Average 58850 48247 14 Average 788950 437432 32

000

2000

4000

6000

8000

10000

DDSMLBP DDSM sparse_LBP

MIASLBP MIAS sparse_LBP

Normal_Abnormal

KNNANNSVM

Figure 13 Average classication rate for MIAS and DDSM dataset

10 Journal of Healthcare Engineering

DDSM AUC 094 in normal versus malignant in case ofTMCH Scanner1 and AUC 096 in normal versusmalignant classification in TMCH Scanner2 using ANNand curvelet subband-based LBP features e worstperformance of AUC 053 for MIAS is obtained with theproposed algorithm using KNN classifier as shown inTable 7 Similarly from Table 7 the best value ofAUC 098 is obtained in TMCH Scanner1 AUC 1 isobtained in TMCH Scanner2 database for normal versusmalignant classification AUC 098 in benign versusmalignant classification is attained in MIAS database andAUC 098 is achieved for normal versus malignant

classification in DDSM database using ANN classifier forsparse curvelet subband-based LBP features

However from Table 7 it should be noted that the per-formance of proposed algorithm is the best using ANN clas-sifier Figure 14 represents automated CAD system for breastcancer diagnosis with sample mammograms

Table 8 provides comparative study of methods developedfor breast tissue classification e proposed method providesbest results in terms of AUC and reduction of number of FPsas 085 to 001 FPimage for MIAS 481 to 003 FPimage forDDSM and 232 to 000 FPimage for TMCH e earlierreported work uses the fixed patch size-based approach which

Table 3 Reduction in curvelet coefficients for sample mammograms from TMCH Scanner1 and Scanner2 dataset

Srno

TMCH Scanner 1 ldquoGE Medical Senograph Systemrdquo TMCH Scanner 2 ldquoHologic Selenia Systemrdquo

ROI size

Total number ofcurvelet

coefficients fromsubbands

Total number ofselected curveletcoefficients from

subbands

reductionin curveletcoefficients

ROI size

Total number ofcurvelet

coefficients fromsubbands

Total number ofselected curveletcoefficients from

subbands

reductionin curveletcoefficients

1 459times 412 1140617 794885 3031 291times 278 489243 317014 35202 548times 513 1693403 1117873 3399 545times 246 809627 531604 34343 415times 303 758323 445585 4124 560times 483 1630583 1068439 34474 645times 495 1951443 1145580 4129 782times 510 2400137 1275073 46875 437times 691 1816651 985120 4577 87times141 75773 67871 10436 812times 500 2439937 1287065 4725 311times 185 348565 240546 30997 468times 379 1066333 710242 3340 262times 348 550303 282821 48618 673times 582 2355589 1730915 2652 610times 440 1614515 842876 47799 250times 201 304513 235670 2261 949times 391 2227209 1359338 389710 525times 488 1544691 1161942 2478 365times 385 846523 604473 285911 488times 779 2287547 1611385 2956 393times 247 585063 490474 161712 1434times 966 8326581 3742277 5506 341times 301 618111 410542 335813 348times 421 881701 616501 3008 523times 702 2206097 800057 637314 460times 530 1467227 799885 4548 370times 284 632539 423727 330115 398times 450 1078441 828064 2322 344times 202 418955 302427 278116 247times 272 404401 344822 1473 264times188 299997 231926 226917 411times 305 757657 405919 4642 233times 247 346983 267686 228518 286times 344 593155 448566 2438 370times 291 650701 486125 252919 417times 207 523477 429755 1790 680times 483 1979543 1052793 468220 463times 458 1275021 857608 3274 202times 266 324295 215042 3369

Average 1633335 984983 33 Average 952738 563543 34

Table 4 Number of ROIs resulted in FP reduction using curvelet-based LBP (without sparse) amp ANN classification at training andvalidation stage

Class Datasetused

Benignmalignant mass Nonmassbenign mass Total()

images

TPR (true-positiverate)TPlesions

FPPI (false-positiveper image) FP

imagesPreviousstage

Selected(TP)

Lost(FN)

Previousstage

Selected(TN)

Lost(FP)

Normal vsabnormal

MIAS 108lowast 2 216 203 13 273 257 16 315 (203216) 094 (16315) 005DDSM 140lowast 4 560 465 95 1203 1095 108 240 (465560) 083 (108240) 045

Benign vsmalignant

MIAS 49 49 0 59 57 2 108 (4949) 100 (2108) 002DDSM 46lowast 2 92 91 1 94 91 3 140 (9192) 099 (3140) 002

Normal vsmalignant

MIAS 49lowast 4196 184 12 273 254 19 256 (184196) 094 (19256) 007DDSM 46lowast 4184 180 4 1203 1143 60 146 (180184) 098 (60146) 041TMCHScanner1 107lowast 4 428 416 12 605 551 54 217 (416428) 097 (54217) 025

TMCHScanner2 65lowast 4 260 255 5 232 214 18 135 (255260) 098 (18135) 013

lowastAugmentation of image

Journal of Healthcare Engineering 11

limits the automatic CAD system scope whereas proposedsystem provides complete solution to CAD system right fromautomatic tumor patch segmentation to reduction in FPs andfinal representation of mammogram with TP marked on itIt will drastically reduce the radiologist work by locationtumor directly on mammogram

5 Conclusion

A fully automatic CAD system which can accuratelylocate the tumor on a mammogram and reduces FPshas been proposed e developed CAD system consistsof preprocessing SOM clustering ROI extraction

Table 6 Performance evaluation of curvelet-based LBP descriptor algorithm

Dataset Classification Normal-malignant Normal-abnormal Benign-malignantClassifier Sensitivity Specificity AUC Sensitivity Specificity AUC Sensitivity Specificity AUC

MIASANN 094 093 094 094 094 094 100 097 099SVM 085 085 085 083 086 085 088 084 086KNN 067 063 065 058 057 058 062 068 063

DDSMANN 098 095 095 083 091 085 099 097 098SVM 097 088 092 071 091 083 094 089 092KNN 096 064 087 067 090 080 087 073 079

TMCH Scanner1ANN 097 091 094 mdash mdash mdash mdash mdash mdashSVM 096 091 094 mdash mdash mdash mdash mdash mdashKNN 098 082 089 mdash mdash mdash mdash mdash mdash

TMCH Scanner2ANN 098 092 096 mdash mdash mdash mdash mdash mdashSVM 097 090 094 mdash mdash mdash mdash mdash mdashKNN 092 083 088 mdash mdash mdash mdash mdash mdash

Table 7 Performance evaluation of proposed algorithm

Dataset Classification Normal-malignant Normal-abnormal Benign-malignantClassifier Sensitivity Specificity AUC Sensitivity Specificity AUC Sensitivity Specificity AUC

MIASANN 098 095 096 093 097 095 097 100 098SVM 088 083 085 085 082 084 084 092 087KNN 055 051 053 055 063 056 061 067 061

DDSMANN 099 097 098 092 096 093 097 095 096SVM 099 092 096 089 096 092 094 092 093KNN 098 073 092 074 090 082 089 077 083

TMCH Scanner1ANN 099 098 098 mdash mdash mdash mdash mdash mdashSVM 098 096 097 mdash mdash mdash mdash mdash mdashKNN 099 092 095 mdash mdash mdash mdash mdash mdash

TMCH Scanner2ANN 100 100 100 mdash mdash mdash mdash mdash mdashSVM 100 098 099 mdash mdash mdash mdash mdash mdashKNN 096 092 094 mdash mdash mdash mdash mdash mdash

Table 5 Number of ROIs resulted in FP reduction using sparse curvelet coefficient-based LBP amp ANN classification at training andvalidation stage

Class Datasetused

Benignmalignant mass Nonmassbenign massTotal ()images

TPR (true-positiverate)TPlesions

FPPI (false-positiveper image) FP

imagesPreviousstage

Selected(TP)

Lost(FN)

Previousstage

Selected(TN)

Lost(FP)

Normal vsabnormal

MIAS 108lowast 2 216 201 15 273 265 8 315 (201216) 093 (8315) 002DDSM 140lowast 4 560 516 44 1203 1155 48 240 (516560) 092 (48240) 02

Benign vsmalignant

MIAS 49 48 1 59 59 1 108 (4849) 098 (1108) 001DDSM 46lowast 2 92 89 3 94 89 5 140 (8992) 097 (5140) 003

Normal vsmalignant

MIAS 49lowast 4196 192 4 273 259 14 256 (192196) 098 (14256) 005DDSM 46lowast 4184 182 2 1203 1167 36 146 (182184) 099 (36146) 025TMCHScanner1 107lowast 4 428 424 4 605 593 12 217 (424428) 099 (12217) 005

TMCHScanner2 65lowast 4 260 260 0 232 232 0 135 (260260) 100 (0135) 0

lowastAugmentation of image

12 Journal of Healthcare Engineering

sparse LBP feature computation based on sparse Cur-velet coefficients and finally FP reduction using ANNclassifier

e proposed algorithm presents a novel concept ofextraction of curvelet coefficients according to irregularshape of mass is called as sparse curvelet coefficients andcomputation of LBP e analysis proves that the FPs arereduced significantly from 085 to 001 FPimage for MIAS

481 to 003 FPimage for DDSM and 232 to 000 FPimagefor TMCH e ANN classifier showed best results asAUC 098 and accuracy 9857 for MIAS in benign-malignant classification AUC 098 and accuracy 9870for DDSM in normal-malignant classification AUC 098and accuracy 9830 for TMCH Scanner1 and AUC 1and accuracy 100 for TMCH Scanner2 in normal-malignant classification as compared with SVM and KNN

(a) (b) (c) (d) (e) (f )

Figure 14 Representation of fully automatic CAD system for breast cancer using (a) samplemammograms fromMIAS DDSM and TMCHdatasets (b) preprocessed mammograms (c) clustered image (d) TP and FP marked on mammogram (e) TP marked by thresholding (f )TP marked by using LBP descriptor based on sparse curvelet coefficients

Journal of Healthcare Engineering 13

classifier e performance of LBP features and LBP featuresbased on sparse curvelet coefficients are nearly same whichshow that the proposed algorithm is suitable for cancer breasttissue diagnosis

In future the reduced curvelet coefficients can be used toextract local ternary patterns and other local descriptor andlocal directional patterns etc e present work deals withmammogram with single mass this can be further extendedfor multiple mass models with multiple LBP features basedon sparse curvelet coefficients

Data Availability

In this research we have used two publicly available datasetsMIAS and DDSM ese datasets can be found here in [17]and [19] e third database is collected from the localhospital Tata Memorial Cancer Hospital Mumbai whichcan be found at httpeurekasveriacin or available fromthe corresponding author upon request

Conflicts of Interest

e authors declare that they have no conflicts of interest

Acknowledgments

e TMCH database for this work was given by Departmentof Radiodiagnosis Tata Memorial Cancer Hospital Mumbai

References

[1] S Malvia S A Bagadi U S Dubey and S Saxena ldquoEpi-demiology of breast cancer in Indian womenrdquo Asia-PacificJournal of Clinical Oncology vol 13 no 4 pp 289ndash295 2017

[2] A Gupta K Shridhar and P Dhillon ldquoA review of breastcancer awareness among women in India cancer literate orawareness deficitrdquo European Journal of Cancer vol 51no 14 pp 2058ndash2066 2015

[3] K P Kanadam and S R Chereddy ldquoMammogram classifi-cation using sparse-ROI a novel representation to arbitraryshaped massesrdquo Expert Systems with Applications vol 57pp 204ndash213 2016

[4] D O T Bruno M Z do Nascimento R P Ramos et al ldquoLBPoperators on curvelet coefficients as an algorithm to describetexture in breast cancer tissuesrdquo Expert Systems with Appli-cations vol 55 pp 329ndash340 2016

[5] M Hussain ldquoFalse-positive reduction in mammographyusing multiscale spatial Weber law descriptor and supportvector machinesrdquo Neural Computing and Applicationsvol 25 no 1 pp 83ndash93 2014

[6] M M Pawar and S N Talbar ldquoGenetic fuzzy system (GFS)based wavelet co-occurrence feature selection in mammo-gram classification for breast cancer diagnosisrdquo Perspectives inScience vol 8 pp 247ndash250 2016

[7] C Muramatsu T Hara T Endo and H Fujita ldquoBreast massclassification on mammograms using radial local ternarypatternsrdquo Computers in Biology and Medicine vol 72pp 43ndash53 2016

[8] M M Eltoukhy I Faye and B B Samir ldquoA statistical basedfeature extraction method for breast cancer diagnosis indigital mammogram using multiresolution representationrdquo

Table 8 Comparison of classification accuracy AUC and FPimage values from different approaches in breast cancer diagnosis

Author Database Method Classifier Result AUC FPimageEltoukhy et al [33]

MIAS Biggest curvelet coefficients as a featurevector

Euclidean classifier 9407 mdash mdashEltoukhy et al [42] 9859 mdash mdashEltoukhy et al [8] SVM 973 mdash mdash

Dhahbi et al [34] Mini-MIAS Curvelet moments KNN 9127 mdash mdashDDSM 8646 mdash mdash

Bruno et al [4] DDSM Curvelet + LBP SVM 85 085 mdashPL 94 094 mdash

da Rocha et al [40] DDSM LBP SVM 8831 088 mdashKanadam andChereddy [3] MIAS Sparse ROI SVM 9742 mdash mdash

Pereira et al [18] DDSM Wavelet and Wiener filter Multiple thresholding waveletand GA mdash mdash 137

Liu and Zeng [29] DDSMFFDM GLCM CLBP and geometric features SVM mdash mdash 148

De Sampaio et al[39] DDSM LBP DBSCAN 9826 019

Zyout et al [30] DDSM Second order statistics of waveletcoefficients (SOSWC) SVM 968 097 0018

MIAS 952 966 0029

Casti et al [31]DDSM

Differential features Fisher linear discriminantanalysis (FLDA) mdash mdash

168MIAS 212FFDM 082

Proposed method

MIAS

LBP based on sparse curvelet subbandcoefficients ANN

9857 098 001DDSM 9870 098 003TMCHScanner1 9830 098 005

TMCHScanner2 100 1 0

14 Journal of Healthcare Engineering

DDSM AUC 094 in normal versus malignant in case ofTMCH Scanner1 and AUC 096 in normal versusmalignant classification in TMCH Scanner2 using ANNand curvelet subband-based LBP features e worstperformance of AUC 053 for MIAS is obtained with theproposed algorithm using KNN classifier as shown inTable 7 Similarly from Table 7 the best value ofAUC 098 is obtained in TMCH Scanner1 AUC 1 isobtained in TMCH Scanner2 database for normal versusmalignant classification AUC 098 in benign versusmalignant classification is attained in MIAS database andAUC 098 is achieved for normal versus malignant

classification in DDSM database using ANN classifier forsparse curvelet subband-based LBP features

However from Table 7 it should be noted that the per-formance of proposed algorithm is the best using ANN clas-sifier Figure 14 represents automated CAD system for breastcancer diagnosis with sample mammograms

Table 8 provides comparative study of methods developedfor breast tissue classification e proposed method providesbest results in terms of AUC and reduction of number of FPsas 085 to 001 FPimage for MIAS 481 to 003 FPimage forDDSM and 232 to 000 FPimage for TMCH e earlierreported work uses the fixed patch size-based approach which

Table 3 Reduction in curvelet coefficients for sample mammograms from TMCH Scanner1 and Scanner2 dataset

Srno

TMCH Scanner 1 ldquoGE Medical Senograph Systemrdquo TMCH Scanner 2 ldquoHologic Selenia Systemrdquo

ROI size

Total number ofcurvelet

coefficients fromsubbands

Total number ofselected curveletcoefficients from

subbands

reductionin curveletcoefficients

ROI size

Total number ofcurvelet

coefficients fromsubbands

Total number ofselected curveletcoefficients from

subbands

reductionin curveletcoefficients

1 459times 412 1140617 794885 3031 291times 278 489243 317014 35202 548times 513 1693403 1117873 3399 545times 246 809627 531604 34343 415times 303 758323 445585 4124 560times 483 1630583 1068439 34474 645times 495 1951443 1145580 4129 782times 510 2400137 1275073 46875 437times 691 1816651 985120 4577 87times141 75773 67871 10436 812times 500 2439937 1287065 4725 311times 185 348565 240546 30997 468times 379 1066333 710242 3340 262times 348 550303 282821 48618 673times 582 2355589 1730915 2652 610times 440 1614515 842876 47799 250times 201 304513 235670 2261 949times 391 2227209 1359338 389710 525times 488 1544691 1161942 2478 365times 385 846523 604473 285911 488times 779 2287547 1611385 2956 393times 247 585063 490474 161712 1434times 966 8326581 3742277 5506 341times 301 618111 410542 335813 348times 421 881701 616501 3008 523times 702 2206097 800057 637314 460times 530 1467227 799885 4548 370times 284 632539 423727 330115 398times 450 1078441 828064 2322 344times 202 418955 302427 278116 247times 272 404401 344822 1473 264times188 299997 231926 226917 411times 305 757657 405919 4642 233times 247 346983 267686 228518 286times 344 593155 448566 2438 370times 291 650701 486125 252919 417times 207 523477 429755 1790 680times 483 1979543 1052793 468220 463times 458 1275021 857608 3274 202times 266 324295 215042 3369

Average 1633335 984983 33 Average 952738 563543 34

Table 4 Number of ROIs resulted in FP reduction using curvelet-based LBP (without sparse) amp ANN classification at training andvalidation stage

Class Datasetused

Benignmalignant mass Nonmassbenign mass Total()

images

TPR (true-positiverate)TPlesions

FPPI (false-positiveper image) FP

imagesPreviousstage

Selected(TP)

Lost(FN)

Previousstage

Selected(TN)

Lost(FP)

Normal vsabnormal

MIAS 108lowast 2 216 203 13 273 257 16 315 (203216) 094 (16315) 005DDSM 140lowast 4 560 465 95 1203 1095 108 240 (465560) 083 (108240) 045

Benign vsmalignant

MIAS 49 49 0 59 57 2 108 (4949) 100 (2108) 002DDSM 46lowast 2 92 91 1 94 91 3 140 (9192) 099 (3140) 002

Normal vsmalignant

MIAS 49lowast 4196 184 12 273 254 19 256 (184196) 094 (19256) 007DDSM 46lowast 4184 180 4 1203 1143 60 146 (180184) 098 (60146) 041TMCHScanner1 107lowast 4 428 416 12 605 551 54 217 (416428) 097 (54217) 025

TMCHScanner2 65lowast 4 260 255 5 232 214 18 135 (255260) 098 (18135) 013

lowastAugmentation of image

Journal of Healthcare Engineering 11

limits the automatic CAD system scope whereas proposedsystem provides complete solution to CAD system right fromautomatic tumor patch segmentation to reduction in FPs andfinal representation of mammogram with TP marked on itIt will drastically reduce the radiologist work by locationtumor directly on mammogram

5 Conclusion

A fully automatic CAD system which can accuratelylocate the tumor on a mammogram and reduces FPshas been proposed e developed CAD system consistsof preprocessing SOM clustering ROI extraction

Table 6 Performance evaluation of curvelet-based LBP descriptor algorithm

Dataset Classification Normal-malignant Normal-abnormal Benign-malignantClassifier Sensitivity Specificity AUC Sensitivity Specificity AUC Sensitivity Specificity AUC

MIASANN 094 093 094 094 094 094 100 097 099SVM 085 085 085 083 086 085 088 084 086KNN 067 063 065 058 057 058 062 068 063

DDSMANN 098 095 095 083 091 085 099 097 098SVM 097 088 092 071 091 083 094 089 092KNN 096 064 087 067 090 080 087 073 079

TMCH Scanner1ANN 097 091 094 mdash mdash mdash mdash mdash mdashSVM 096 091 094 mdash mdash mdash mdash mdash mdashKNN 098 082 089 mdash mdash mdash mdash mdash mdash

TMCH Scanner2ANN 098 092 096 mdash mdash mdash mdash mdash mdashSVM 097 090 094 mdash mdash mdash mdash mdash mdashKNN 092 083 088 mdash mdash mdash mdash mdash mdash

Table 7 Performance evaluation of proposed algorithm

Dataset Classification Normal-malignant Normal-abnormal Benign-malignantClassifier Sensitivity Specificity AUC Sensitivity Specificity AUC Sensitivity Specificity AUC

MIASANN 098 095 096 093 097 095 097 100 098SVM 088 083 085 085 082 084 084 092 087KNN 055 051 053 055 063 056 061 067 061

DDSMANN 099 097 098 092 096 093 097 095 096SVM 099 092 096 089 096 092 094 092 093KNN 098 073 092 074 090 082 089 077 083

TMCH Scanner1ANN 099 098 098 mdash mdash mdash mdash mdash mdashSVM 098 096 097 mdash mdash mdash mdash mdash mdashKNN 099 092 095 mdash mdash mdash mdash mdash mdash

TMCH Scanner2ANN 100 100 100 mdash mdash mdash mdash mdash mdashSVM 100 098 099 mdash mdash mdash mdash mdash mdashKNN 096 092 094 mdash mdash mdash mdash mdash mdash

Table 5 Number of ROIs resulted in FP reduction using sparse curvelet coefficient-based LBP amp ANN classification at training andvalidation stage

Class Datasetused

Benignmalignant mass Nonmassbenign massTotal ()images

TPR (true-positiverate)TPlesions

FPPI (false-positiveper image) FP

imagesPreviousstage

Selected(TP)

Lost(FN)

Previousstage

Selected(TN)

Lost(FP)

Normal vsabnormal

MIAS 108lowast 2 216 201 15 273 265 8 315 (201216) 093 (8315) 002DDSM 140lowast 4 560 516 44 1203 1155 48 240 (516560) 092 (48240) 02

Benign vsmalignant

MIAS 49 48 1 59 59 1 108 (4849) 098 (1108) 001DDSM 46lowast 2 92 89 3 94 89 5 140 (8992) 097 (5140) 003

Normal vsmalignant

MIAS 49lowast 4196 192 4 273 259 14 256 (192196) 098 (14256) 005DDSM 46lowast 4184 182 2 1203 1167 36 146 (182184) 099 (36146) 025TMCHScanner1 107lowast 4 428 424 4 605 593 12 217 (424428) 099 (12217) 005

TMCHScanner2 65lowast 4 260 260 0 232 232 0 135 (260260) 100 (0135) 0

lowastAugmentation of image

12 Journal of Healthcare Engineering

sparse LBP feature computation based on sparse Cur-velet coefficients and finally FP reduction using ANNclassifier

e proposed algorithm presents a novel concept ofextraction of curvelet coefficients according to irregularshape of mass is called as sparse curvelet coefficients andcomputation of LBP e analysis proves that the FPs arereduced significantly from 085 to 001 FPimage for MIAS

481 to 003 FPimage for DDSM and 232 to 000 FPimagefor TMCH e ANN classifier showed best results asAUC 098 and accuracy 9857 for MIAS in benign-malignant classification AUC 098 and accuracy 9870for DDSM in normal-malignant classification AUC 098and accuracy 9830 for TMCH Scanner1 and AUC 1and accuracy 100 for TMCH Scanner2 in normal-malignant classification as compared with SVM and KNN

(a) (b) (c) (d) (e) (f )

Figure 14 Representation of fully automatic CAD system for breast cancer using (a) samplemammograms fromMIAS DDSM and TMCHdatasets (b) preprocessed mammograms (c) clustered image (d) TP and FP marked on mammogram (e) TP marked by thresholding (f )TP marked by using LBP descriptor based on sparse curvelet coefficients

Journal of Healthcare Engineering 13

classifier e performance of LBP features and LBP featuresbased on sparse curvelet coefficients are nearly same whichshow that the proposed algorithm is suitable for cancer breasttissue diagnosis

In future the reduced curvelet coefficients can be used toextract local ternary patterns and other local descriptor andlocal directional patterns etc e present work deals withmammogram with single mass this can be further extendedfor multiple mass models with multiple LBP features basedon sparse curvelet coefficients

Data Availability

In this research we have used two publicly available datasetsMIAS and DDSM ese datasets can be found here in [17]and [19] e third database is collected from the localhospital Tata Memorial Cancer Hospital Mumbai whichcan be found at httpeurekasveriacin or available fromthe corresponding author upon request

Conflicts of Interest

e authors declare that they have no conflicts of interest

Acknowledgments

e TMCH database for this work was given by Departmentof Radiodiagnosis Tata Memorial Cancer Hospital Mumbai

References

[1] S Malvia S A Bagadi U S Dubey and S Saxena ldquoEpi-demiology of breast cancer in Indian womenrdquo Asia-PacificJournal of Clinical Oncology vol 13 no 4 pp 289ndash295 2017

[2] A Gupta K Shridhar and P Dhillon ldquoA review of breastcancer awareness among women in India cancer literate orawareness deficitrdquo European Journal of Cancer vol 51no 14 pp 2058ndash2066 2015

[3] K P Kanadam and S R Chereddy ldquoMammogram classifi-cation using sparse-ROI a novel representation to arbitraryshaped massesrdquo Expert Systems with Applications vol 57pp 204ndash213 2016

[4] D O T Bruno M Z do Nascimento R P Ramos et al ldquoLBPoperators on curvelet coefficients as an algorithm to describetexture in breast cancer tissuesrdquo Expert Systems with Appli-cations vol 55 pp 329ndash340 2016

[5] M Hussain ldquoFalse-positive reduction in mammographyusing multiscale spatial Weber law descriptor and supportvector machinesrdquo Neural Computing and Applicationsvol 25 no 1 pp 83ndash93 2014

[6] M M Pawar and S N Talbar ldquoGenetic fuzzy system (GFS)based wavelet co-occurrence feature selection in mammo-gram classification for breast cancer diagnosisrdquo Perspectives inScience vol 8 pp 247ndash250 2016

[7] C Muramatsu T Hara T Endo and H Fujita ldquoBreast massclassification on mammograms using radial local ternarypatternsrdquo Computers in Biology and Medicine vol 72pp 43ndash53 2016

[8] M M Eltoukhy I Faye and B B Samir ldquoA statistical basedfeature extraction method for breast cancer diagnosis indigital mammogram using multiresolution representationrdquo

Table 8 Comparison of classification accuracy AUC and FPimage values from different approaches in breast cancer diagnosis

Author Database Method Classifier Result AUC FPimageEltoukhy et al [33]

MIAS Biggest curvelet coefficients as a featurevector

Euclidean classifier 9407 mdash mdashEltoukhy et al [42] 9859 mdash mdashEltoukhy et al [8] SVM 973 mdash mdash

Dhahbi et al [34] Mini-MIAS Curvelet moments KNN 9127 mdash mdashDDSM 8646 mdash mdash

Bruno et al [4] DDSM Curvelet + LBP SVM 85 085 mdashPL 94 094 mdash

da Rocha et al [40] DDSM LBP SVM 8831 088 mdashKanadam andChereddy [3] MIAS Sparse ROI SVM 9742 mdash mdash

Pereira et al [18] DDSM Wavelet and Wiener filter Multiple thresholding waveletand GA mdash mdash 137

Liu and Zeng [29] DDSMFFDM GLCM CLBP and geometric features SVM mdash mdash 148

De Sampaio et al[39] DDSM LBP DBSCAN 9826 019

Zyout et al [30] DDSM Second order statistics of waveletcoefficients (SOSWC) SVM 968 097 0018

MIAS 952 966 0029

Casti et al [31]DDSM

Differential features Fisher linear discriminantanalysis (FLDA) mdash mdash

168MIAS 212FFDM 082

Proposed method

MIAS

LBP based on sparse curvelet subbandcoefficients ANN

9857 098 001DDSM 9870 098 003TMCHScanner1 9830 098 005

TMCHScanner2 100 1 0

14 Journal of Healthcare Engineering

limits the automatic CAD system scope whereas proposedsystem provides complete solution to CAD system right fromautomatic tumor patch segmentation to reduction in FPs andfinal representation of mammogram with TP marked on itIt will drastically reduce the radiologist work by locationtumor directly on mammogram

5 Conclusion

A fully automatic CAD system which can accuratelylocate the tumor on a mammogram and reduces FPshas been proposed e developed CAD system consistsof preprocessing SOM clustering ROI extraction

Table 6 Performance evaluation of curvelet-based LBP descriptor algorithm

Dataset Classification Normal-malignant Normal-abnormal Benign-malignantClassifier Sensitivity Specificity AUC Sensitivity Specificity AUC Sensitivity Specificity AUC

MIASANN 094 093 094 094 094 094 100 097 099SVM 085 085 085 083 086 085 088 084 086KNN 067 063 065 058 057 058 062 068 063

DDSMANN 098 095 095 083 091 085 099 097 098SVM 097 088 092 071 091 083 094 089 092KNN 096 064 087 067 090 080 087 073 079

TMCH Scanner1ANN 097 091 094 mdash mdash mdash mdash mdash mdashSVM 096 091 094 mdash mdash mdash mdash mdash mdashKNN 098 082 089 mdash mdash mdash mdash mdash mdash

TMCH Scanner2ANN 098 092 096 mdash mdash mdash mdash mdash mdashSVM 097 090 094 mdash mdash mdash mdash mdash mdashKNN 092 083 088 mdash mdash mdash mdash mdash mdash

Table 7 Performance evaluation of proposed algorithm

Dataset Classification Normal-malignant Normal-abnormal Benign-malignantClassifier Sensitivity Specificity AUC Sensitivity Specificity AUC Sensitivity Specificity AUC

MIASANN 098 095 096 093 097 095 097 100 098SVM 088 083 085 085 082 084 084 092 087KNN 055 051 053 055 063 056 061 067 061

DDSMANN 099 097 098 092 096 093 097 095 096SVM 099 092 096 089 096 092 094 092 093KNN 098 073 092 074 090 082 089 077 083

TMCH Scanner1ANN 099 098 098 mdash mdash mdash mdash mdash mdashSVM 098 096 097 mdash mdash mdash mdash mdash mdashKNN 099 092 095 mdash mdash mdash mdash mdash mdash

TMCH Scanner2ANN 100 100 100 mdash mdash mdash mdash mdash mdashSVM 100 098 099 mdash mdash mdash mdash mdash mdashKNN 096 092 094 mdash mdash mdash mdash mdash mdash

Table 5 Number of ROIs resulted in FP reduction using sparse curvelet coefficient-based LBP amp ANN classification at training andvalidation stage

Class Datasetused

Benignmalignant mass Nonmassbenign massTotal ()images

TPR (true-positiverate)TPlesions

FPPI (false-positiveper image) FP

imagesPreviousstage

Selected(TP)

Lost(FN)

Previousstage

Selected(TN)

Lost(FP)

Normal vsabnormal

MIAS 108lowast 2 216 201 15 273 265 8 315 (201216) 093 (8315) 002DDSM 140lowast 4 560 516 44 1203 1155 48 240 (516560) 092 (48240) 02

Benign vsmalignant

MIAS 49 48 1 59 59 1 108 (4849) 098 (1108) 001DDSM 46lowast 2 92 89 3 94 89 5 140 (8992) 097 (5140) 003

Normal vsmalignant

MIAS 49lowast 4196 192 4 273 259 14 256 (192196) 098 (14256) 005DDSM 46lowast 4184 182 2 1203 1167 36 146 (182184) 099 (36146) 025TMCHScanner1 107lowast 4 428 424 4 605 593 12 217 (424428) 099 (12217) 005

TMCHScanner2 65lowast 4 260 260 0 232 232 0 135 (260260) 100 (0135) 0

lowastAugmentation of image

12 Journal of Healthcare Engineering

sparse LBP feature computation based on sparse Cur-velet coefficients and finally FP reduction using ANNclassifier

e proposed algorithm presents a novel concept ofextraction of curvelet coefficients according to irregularshape of mass is called as sparse curvelet coefficients andcomputation of LBP e analysis proves that the FPs arereduced significantly from 085 to 001 FPimage for MIAS

481 to 003 FPimage for DDSM and 232 to 000 FPimagefor TMCH e ANN classifier showed best results asAUC 098 and accuracy 9857 for MIAS in benign-malignant classification AUC 098 and accuracy 9870for DDSM in normal-malignant classification AUC 098and accuracy 9830 for TMCH Scanner1 and AUC 1and accuracy 100 for TMCH Scanner2 in normal-malignant classification as compared with SVM and KNN

(a) (b) (c) (d) (e) (f )

Figure 14 Representation of fully automatic CAD system for breast cancer using (a) samplemammograms fromMIAS DDSM and TMCHdatasets (b) preprocessed mammograms (c) clustered image (d) TP and FP marked on mammogram (e) TP marked by thresholding (f )TP marked by using LBP descriptor based on sparse curvelet coefficients

Journal of Healthcare Engineering 13

classifier e performance of LBP features and LBP featuresbased on sparse curvelet coefficients are nearly same whichshow that the proposed algorithm is suitable for cancer breasttissue diagnosis

In future the reduced curvelet coefficients can be used toextract local ternary patterns and other local descriptor andlocal directional patterns etc e present work deals withmammogram with single mass this can be further extendedfor multiple mass models with multiple LBP features basedon sparse curvelet coefficients

Data Availability

In this research we have used two publicly available datasetsMIAS and DDSM ese datasets can be found here in [17]and [19] e third database is collected from the localhospital Tata Memorial Cancer Hospital Mumbai whichcan be found at httpeurekasveriacin or available fromthe corresponding author upon request

Conflicts of Interest

e authors declare that they have no conflicts of interest

Acknowledgments

e TMCH database for this work was given by Departmentof Radiodiagnosis Tata Memorial Cancer Hospital Mumbai

References

[1] S Malvia S A Bagadi U S Dubey and S Saxena ldquoEpi-demiology of breast cancer in Indian womenrdquo Asia-PacificJournal of Clinical Oncology vol 13 no 4 pp 289ndash295 2017

[2] A Gupta K Shridhar and P Dhillon ldquoA review of breastcancer awareness among women in India cancer literate orawareness deficitrdquo European Journal of Cancer vol 51no 14 pp 2058ndash2066 2015

[3] K P Kanadam and S R Chereddy ldquoMammogram classifi-cation using sparse-ROI a novel representation to arbitraryshaped massesrdquo Expert Systems with Applications vol 57pp 204ndash213 2016

[4] D O T Bruno M Z do Nascimento R P Ramos et al ldquoLBPoperators on curvelet coefficients as an algorithm to describetexture in breast cancer tissuesrdquo Expert Systems with Appli-cations vol 55 pp 329ndash340 2016

[5] M Hussain ldquoFalse-positive reduction in mammographyusing multiscale spatial Weber law descriptor and supportvector machinesrdquo Neural Computing and Applicationsvol 25 no 1 pp 83ndash93 2014

[6] M M Pawar and S N Talbar ldquoGenetic fuzzy system (GFS)based wavelet co-occurrence feature selection in mammo-gram classification for breast cancer diagnosisrdquo Perspectives inScience vol 8 pp 247ndash250 2016

[7] C Muramatsu T Hara T Endo and H Fujita ldquoBreast massclassification on mammograms using radial local ternarypatternsrdquo Computers in Biology and Medicine vol 72pp 43ndash53 2016

[8] M M Eltoukhy I Faye and B B Samir ldquoA statistical basedfeature extraction method for breast cancer diagnosis indigital mammogram using multiresolution representationrdquo

Table 8 Comparison of classification accuracy AUC and FPimage values from different approaches in breast cancer diagnosis

Author Database Method Classifier Result AUC FPimageEltoukhy et al [33]

MIAS Biggest curvelet coefficients as a featurevector

Euclidean classifier 9407 mdash mdashEltoukhy et al [42] 9859 mdash mdashEltoukhy et al [8] SVM 973 mdash mdash

Dhahbi et al [34] Mini-MIAS Curvelet moments KNN 9127 mdash mdashDDSM 8646 mdash mdash

Bruno et al [4] DDSM Curvelet + LBP SVM 85 085 mdashPL 94 094 mdash

da Rocha et al [40] DDSM LBP SVM 8831 088 mdashKanadam andChereddy [3] MIAS Sparse ROI SVM 9742 mdash mdash

Pereira et al [18] DDSM Wavelet and Wiener filter Multiple thresholding waveletand GA mdash mdash 137

Liu and Zeng [29] DDSMFFDM GLCM CLBP and geometric features SVM mdash mdash 148

De Sampaio et al[39] DDSM LBP DBSCAN 9826 019

Zyout et al [30] DDSM Second order statistics of waveletcoefficients (SOSWC) SVM 968 097 0018

MIAS 952 966 0029

Casti et al [31]DDSM

Differential features Fisher linear discriminantanalysis (FLDA) mdash mdash

168MIAS 212FFDM 082

Proposed method

MIAS

LBP based on sparse curvelet subbandcoefficients ANN

9857 098 001DDSM 9870 098 003TMCHScanner1 9830 098 005

TMCHScanner2 100 1 0

14 Journal of Healthcare Engineering

sparse LBP feature computation based on sparse Cur-velet coefficients and finally FP reduction using ANNclassifier

e proposed algorithm presents a novel concept ofextraction of curvelet coefficients according to irregularshape of mass is called as sparse curvelet coefficients andcomputation of LBP e analysis proves that the FPs arereduced significantly from 085 to 001 FPimage for MIAS

481 to 003 FPimage for DDSM and 232 to 000 FPimagefor TMCH e ANN classifier showed best results asAUC 098 and accuracy 9857 for MIAS in benign-malignant classification AUC 098 and accuracy 9870for DDSM in normal-malignant classification AUC 098and accuracy 9830 for TMCH Scanner1 and AUC 1and accuracy 100 for TMCH Scanner2 in normal-malignant classification as compared with SVM and KNN

(a) (b) (c) (d) (e) (f )

Figure 14 Representation of fully automatic CAD system for breast cancer using (a) samplemammograms fromMIAS DDSM and TMCHdatasets (b) preprocessed mammograms (c) clustered image (d) TP and FP marked on mammogram (e) TP marked by thresholding (f )TP marked by using LBP descriptor based on sparse curvelet coefficients

Journal of Healthcare Engineering 13

classifier e performance of LBP features and LBP featuresbased on sparse curvelet coefficients are nearly same whichshow that the proposed algorithm is suitable for cancer breasttissue diagnosis

In future the reduced curvelet coefficients can be used toextract local ternary patterns and other local descriptor andlocal directional patterns etc e present work deals withmammogram with single mass this can be further extendedfor multiple mass models with multiple LBP features basedon sparse curvelet coefficients

Data Availability

In this research we have used two publicly available datasetsMIAS and DDSM ese datasets can be found here in [17]and [19] e third database is collected from the localhospital Tata Memorial Cancer Hospital Mumbai whichcan be found at httpeurekasveriacin or available fromthe corresponding author upon request

Conflicts of Interest

e authors declare that they have no conflicts of interest

Acknowledgments

e TMCH database for this work was given by Departmentof Radiodiagnosis Tata Memorial Cancer Hospital Mumbai

References

[1] S Malvia S A Bagadi U S Dubey and S Saxena ldquoEpi-demiology of breast cancer in Indian womenrdquo Asia-PacificJournal of Clinical Oncology vol 13 no 4 pp 289ndash295 2017

[2] A Gupta K Shridhar and P Dhillon ldquoA review of breastcancer awareness among women in India cancer literate orawareness deficitrdquo European Journal of Cancer vol 51no 14 pp 2058ndash2066 2015

[3] K P Kanadam and S R Chereddy ldquoMammogram classifi-cation using sparse-ROI a novel representation to arbitraryshaped massesrdquo Expert Systems with Applications vol 57pp 204ndash213 2016

[4] D O T Bruno M Z do Nascimento R P Ramos et al ldquoLBPoperators on curvelet coefficients as an algorithm to describetexture in breast cancer tissuesrdquo Expert Systems with Appli-cations vol 55 pp 329ndash340 2016

[5] M Hussain ldquoFalse-positive reduction in mammographyusing multiscale spatial Weber law descriptor and supportvector machinesrdquo Neural Computing and Applicationsvol 25 no 1 pp 83ndash93 2014

[6] M M Pawar and S N Talbar ldquoGenetic fuzzy system (GFS)based wavelet co-occurrence feature selection in mammo-gram classification for breast cancer diagnosisrdquo Perspectives inScience vol 8 pp 247ndash250 2016

[7] C Muramatsu T Hara T Endo and H Fujita ldquoBreast massclassification on mammograms using radial local ternarypatternsrdquo Computers in Biology and Medicine vol 72pp 43ndash53 2016

[8] M M Eltoukhy I Faye and B B Samir ldquoA statistical basedfeature extraction method for breast cancer diagnosis indigital mammogram using multiresolution representationrdquo

Table 8 Comparison of classification accuracy AUC and FPimage values from different approaches in breast cancer diagnosis

Author Database Method Classifier Result AUC FPimageEltoukhy et al [33]

MIAS Biggest curvelet coefficients as a featurevector

Euclidean classifier 9407 mdash mdashEltoukhy et al [42] 9859 mdash mdashEltoukhy et al [8] SVM 973 mdash mdash

Dhahbi et al [34] Mini-MIAS Curvelet moments KNN 9127 mdash mdashDDSM 8646 mdash mdash

Bruno et al [4] DDSM Curvelet + LBP SVM 85 085 mdashPL 94 094 mdash

da Rocha et al [40] DDSM LBP SVM 8831 088 mdashKanadam andChereddy [3] MIAS Sparse ROI SVM 9742 mdash mdash

Pereira et al [18] DDSM Wavelet and Wiener filter Multiple thresholding waveletand GA mdash mdash 137

Liu and Zeng [29] DDSMFFDM GLCM CLBP and geometric features SVM mdash mdash 148

De Sampaio et al[39] DDSM LBP DBSCAN 9826 019

Zyout et al [30] DDSM Second order statistics of waveletcoefficients (SOSWC) SVM 968 097 0018

MIAS 952 966 0029

Casti et al [31]DDSM

Differential features Fisher linear discriminantanalysis (FLDA) mdash mdash

168MIAS 212FFDM 082

Proposed method

MIAS

LBP based on sparse curvelet subbandcoefficients ANN

9857 098 001DDSM 9870 098 003TMCHScanner1 9830 098 005

TMCHScanner2 100 1 0

14 Journal of Healthcare Engineering

classifier e performance of LBP features and LBP featuresbased on sparse curvelet coefficients are nearly same whichshow that the proposed algorithm is suitable for cancer breasttissue diagnosis

In future the reduced curvelet coefficients can be used toextract local ternary patterns and other local descriptor andlocal directional patterns etc e present work deals withmammogram with single mass this can be further extendedfor multiple mass models with multiple LBP features basedon sparse curvelet coefficients

Data Availability

In this research we have used two publicly available datasetsMIAS and DDSM ese datasets can be found here in [17]and [19] e third database is collected from the localhospital Tata Memorial Cancer Hospital Mumbai whichcan be found at httpeurekasveriacin or available fromthe corresponding author upon request

Conflicts of Interest

e authors declare that they have no conflicts of interest

Acknowledgments

e TMCH database for this work was given by Departmentof Radiodiagnosis Tata Memorial Cancer Hospital Mumbai

References

[1] S Malvia S A Bagadi U S Dubey and S Saxena ldquoEpi-demiology of breast cancer in Indian womenrdquo Asia-PacificJournal of Clinical Oncology vol 13 no 4 pp 289ndash295 2017

[2] A Gupta K Shridhar and P Dhillon ldquoA review of breastcancer awareness among women in India cancer literate orawareness deficitrdquo European Journal of Cancer vol 51no 14 pp 2058ndash2066 2015

[3] K P Kanadam and S R Chereddy ldquoMammogram classifi-cation using sparse-ROI a novel representation to arbitraryshaped massesrdquo Expert Systems with Applications vol 57pp 204ndash213 2016

[4] D O T Bruno M Z do Nascimento R P Ramos et al ldquoLBPoperators on curvelet coefficients as an algorithm to describetexture in breast cancer tissuesrdquo Expert Systems with Appli-cations vol 55 pp 329ndash340 2016

[5] M Hussain ldquoFalse-positive reduction in mammographyusing multiscale spatial Weber law descriptor and supportvector machinesrdquo Neural Computing and Applicationsvol 25 no 1 pp 83ndash93 2014

[6] M M Pawar and S N Talbar ldquoGenetic fuzzy system (GFS)based wavelet co-occurrence feature selection in mammo-gram classification for breast cancer diagnosisrdquo Perspectives inScience vol 8 pp 247ndash250 2016

[7] C Muramatsu T Hara T Endo and H Fujita ldquoBreast massclassification on mammograms using radial local ternarypatternsrdquo Computers in Biology and Medicine vol 72pp 43ndash53 2016

[8] M M Eltoukhy I Faye and B B Samir ldquoA statistical basedfeature extraction method for breast cancer diagnosis indigital mammogram using multiresolution representationrdquo

Table 8 Comparison of classification accuracy AUC and FPimage values from different approaches in breast cancer diagnosis

Author Database Method Classifier Result AUC FPimageEltoukhy et al [33]

MIAS Biggest curvelet coefficients as a featurevector

Euclidean classifier 9407 mdash mdashEltoukhy et al [42] 9859 mdash mdashEltoukhy et al [8] SVM 973 mdash mdash

Dhahbi et al [34] Mini-MIAS Curvelet moments KNN 9127 mdash mdashDDSM 8646 mdash mdash

Bruno et al [4] DDSM Curvelet + LBP SVM 85 085 mdashPL 94 094 mdash

da Rocha et al [40] DDSM LBP SVM 8831 088 mdashKanadam andChereddy [3] MIAS Sparse ROI SVM 9742 mdash mdash

Pereira et al [18] DDSM Wavelet and Wiener filter Multiple thresholding waveletand GA mdash mdash 137

Liu and Zeng [29] DDSMFFDM GLCM CLBP and geometric features SVM mdash mdash 148

De Sampaio et al[39] DDSM LBP DBSCAN 9826 019

Zyout et al [30] DDSM Second order statistics of waveletcoefficients (SOSWC) SVM 968 097 0018

MIAS 952 966 0029

Casti et al [31]DDSM

Differential features Fisher linear discriminantanalysis (FLDA) mdash mdash

168MIAS 212FFDM 082

Proposed method

MIAS

LBP based on sparse curvelet subbandcoefficients ANN

9857 098 001DDSM 9870 098 003TMCHScanner1 9830 098 005

TMCHScanner2 100 1 0

14 Journal of Healthcare Engineering

Computers in Biology andMedicine vol 42 no 1 pp 123ndash1282012

[9] Y Li H Chen Y Yang L Cheng and L Cao ldquoA bilateralanalysis scheme for false positive reduction in mammogrammass detectionrdquo Computers in Biology and Medicine vol 57pp 84ndash95 2015

[10] A Gandhamal S Talbar S Gajre A F M Hani andD Kumar ldquoLocal gray level S-curve transformationmdashageneralized contrast enhancement technique for medicalimagesrdquo Computers in Biology and Medicine vol 83pp 120ndash133 2017

[11] S Anand and S Gayathri ldquoMammogram image enhancementby two-stage adaptive histogram equalizationrdquo Optik-International Journal for Light and Electron Optics vol 126no 21 pp 3150ndash3152 2015

[12] M M Pawar and S N Talbar ldquoLocal entropy maximizationbased image fusion for contrast enhancement of mammo-gramrdquo Journal of King Saud University-Computer and In-formation Sciences 2018

[13] K Ganesan U R Acharya K C Chua L C Min andK T Abraham ldquoPectoral muscle segmentation a reviewrdquoComputer Methods and Programs in Biomedicine vol 110no 1 pp 48ndash57 2013

[14] I K Maitra S Nag and S K Bandyopadhyay ldquoTechnique forpreprocessing of digital mammogramrdquo Computer Methodsand Programs in Biomedicine vol 107 no 2 pp 175ndash1882012

[15] 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

[16] P Gorgel A Sertbas and O N Ucan ldquoMammographicalmass detection and classification using local seed regiongrowingndashspherical wavelet transform (lsrgndashswt) hybridschemerdquo Computers in Biology and Medicine vol 43 no 6pp 765ndash774 2013

[17] J Suckling J Parker D Dance et al 6e MammographicImage Analysis Society Digital Mammogram Database In-ternational Congress Series Exerpta Medica England UK1994

[18] D C Pereira R P Ramos and M Z do NascimentoldquoSegmentation and detection of breast cancer in mammo-grams combining wavelet analysis and genetic algorithmrdquoComputer Methods and Programs in Biomedicine vol 114no 1 pp 88ndash101 2014

[19] M Heath K Bowyer D Kopans R Moore andP Kegelmeyer Jr ldquoe digital database for screeningmammographyrdquo in Proceedings of the 5th InternationalWorkshop on Digital Mammography M J Yaffe EdMedical Physics Publishing 2001 ISBN 1-930524-00-5

[20] R Rouhi M Jafari S Kasaei and P Keshavarzian ldquoBenignand malignant breast tumors classification based on regiongrowing and CNN segmentationrdquo Expert Systems with Ap-plications vol 42 no 3 pp 990ndash1002 2015

[21] T Berber A Alpkocak P Balci and O Dicle ldquoBreast masscontour segmentation algorithm in digital mammogramsrdquoComputer Methods and Programs in Biomedicine vol 110no 2 pp 150ndash159 2013

[22] R Rouhi and M Jafari ldquoClassification of benign and ma-lignant breast tumors based on hybrid level set segmentationrdquoExpert Systems with Applications vol 46 pp 45ndash59 2016

[23] L A Salazar-Licea J C Pedraza-Ortega A Pastrana-Palmaand M A Aceves-Fernandez ldquoLocation of mammogramsROIrsquos and reduction of false-positiverdquo ComputerMethods andPrograms in Biomedicine vol 143 pp 97ndash111 2017

[24] T Kohonen ldquoe self-organizing maprdquo Neurocomputingvol 21 no 1 pp 1ndash6 1998

[25] A Demirhan and I Guler ldquoCombining stationary wavelettransform and self-organizing maps for brain MR imagesegmentationrdquo Engineering Applications of Artificial In-telligence vol 24 no 2 pp 358ndash367 2011

[26] X Llado A Oliver J Freixenet R Martı and J Martı ldquoAtextural approach for mass false positive reduction inmammographyrdquo Computerized Medical Imaging andGraphics vol 33 no 6 pp 415ndash422 2009

[27] G B Junior S V da Rocha M Gattass A C Silva andA C de Paiva ldquoA mass classification using spatial diversityapproaches in mammography images for false positive re-ductionrdquo Expert Systems with Applications vol 40 no 18pp 7534ndash7543 2013

[28] N Vallez G Bueno O Deniz et al ldquoBreast density classi-fication to reduce false positives in CADe systemsrdquo ComputerMethods and Programs in Biomedicine vol 113 no 2pp 569ndash584 2014

[29] X Liu and Z Zeng ldquoA new automatic mass detection methodfor breast cancer with false positive reductionrdquo Neuro-computing vol 152 pp 388ndash402 2015

[30] I Zyout J Czajkowska and M Grzegorzek ldquoMulti-scaletextural feature extraction and particle swarm optimizationbased model selection for false positive reduction in mam-mographyrdquo Computerized Medical Imaging and Graphicsvol 46 pp 95ndash107 2015

[31] P Casti A Mencattini M Salmeri et al ldquoContour-independent detection and classification of mammographiclesionsrdquo Biomedical Signal Processing and Control vol 25pp 165ndash177 2016

[32] S Beura B Majhi and R Dash ldquoMammogram classificationusing two dimensional discrete wavelet transform and gray-level co-occurrence matrix for detection of breast cancerrdquoNeurocomputing vol 154 pp 1ndash14 2015

[33] M M Eltoukhy I Faye and B B Samir ldquoA comparison ofwavelet and curvelet for breast cancer diagnosis in digitalmammogramrdquo Computers in Biology and Medicine vol 40no 4 pp 384ndash391 2010

[34] S Dhahbi W Barhoumi and E Zagrouba ldquoBreast cancerdiagnosis in digitized mammograms using curvelet mo-mentsrdquo Computers in Biology and Medicine vol 64 pp 79ndash90 2015

[35] U Raghavendra U R Acharya H Fujita A GudigarJ H Tan and S Chokkadi ldquoApplication of Gabor wavelet andlocality sensitive discriminant analysis for automated iden-tification of breast cancer using digitized mammogram im-agesrdquo Applied Soft Computing vol 46 pp 151ndash161 2016

[36] K Ganesan U R Acharya C K Chua L C MinK T Abraham and K-H Ng ldquoComputer-aided breast cancerdetection using mammograms a reviewrdquo IEEE Reviews inBiomedical Engineering vol 6 pp 77ndash98 2013

[37] M Abdel-Nasser H A Rashwan D Puig and A MorenoldquoAnalysis of tissue abnormality and breast density in mam-mographic images using a uniform local directional patternrdquoExpert Systems with Applications vol 42 no 24 pp 9499ndash9511 2015

[38] S K Wajid and A Hussain ldquoLocal energy-based shapehistogram feature extraction technique for breast cancer di-agnosisrdquo Expert Systems with Applications vol 42 no 20pp 6990ndash6999 2015

[39] W B de Sampaio A C Silva A C de Paiva and M GattassldquoDetection of masses inmammograms with adaption to breastdensity using genetic algorithm phylogenetic trees LBP and

Journal of Healthcare Engineering 15

SVMrdquo Expert Systems with Applications vol 42 no 22pp 8911ndash8928 2015

[40] S V da Rocha G B Junior A C Silva A C de Paiva andM Gattass ldquoTexture analysis of masses malignant in mam-mograms images using a combined approach of diversityindex and local binary patterns distributionrdquo Expert Systemswith Applications vol 66 pp 7ndash19 2016

[41] T Ojala M Pietikainen and D Harwood ldquoA comparativestudy of texture measures with classification based on featureddistributionsrdquo Pattern Recognition vol 29 no 1 pp 51ndash591996

[42] M M Eltoukhy I Faye and B B Samir ldquoBreast cancerdiagnosis in digital mammogram using multiscale curvelettransformrdquo Computerized Medical Imaging and Graphicsvol 34 no 4 pp 269ndash276 2010

[43] T Ojala M Pietikainen and T Maenpaa ldquoMultiresolutiongray-scale and rotation invariant texture classification withlocal binary patternsrdquo IEEE Transactions on Pattern Analysisand Machine Intelligence vol 24 no 7 pp 971ndash987 2002

[44] E Candes L Demanet D Donoho and L Ying ldquoFast discretecurvelet transformsrdquo Multiscale Modeling amp Simulationvol 5 no 3 pp 861ndash899 2006

[45] A Dudhane G Shingadkar P Sanghavi B Jankharia andS Talbar ldquoInterstitial lung disease classification using feedforward neural networksrdquo in Proceedings of Advances in Intel-ligent Systems Research vol 137 pp 515ndash521 2017

[46] A A Dudhane and S N Talbar ldquoMulti-scale directional maskpattern for medical image classification and retrievalrdquo inProceedings of 2nd International Conference on ComputerVision amp Image Processing Indian Institute of TechnologyRoorkee Roorkee India 2018

[47] httpeurekasveriacin

16 Journal of Healthcare Engineering

SVMrdquo Expert Systems with Applications vol 42 no 22pp 8911ndash8928 2015

[40] S V da Rocha G B Junior A C Silva A C de Paiva andM Gattass ldquoTexture analysis of masses malignant in mam-mograms images using a combined approach of diversityindex and local binary patterns distributionrdquo Expert Systemswith Applications vol 66 pp 7ndash19 2016

[41] T Ojala M Pietikainen and D Harwood ldquoA comparativestudy of texture measures with classification based on featureddistributionsrdquo Pattern Recognition vol 29 no 1 pp 51ndash591996

[42] M M Eltoukhy I Faye and B B Samir ldquoBreast cancerdiagnosis in digital mammogram using multiscale curvelettransformrdquo Computerized Medical Imaging and Graphicsvol 34 no 4 pp 269ndash276 2010

[43] T Ojala M Pietikainen and T Maenpaa ldquoMultiresolutiongray-scale and rotation invariant texture classification withlocal binary patternsrdquo IEEE Transactions on Pattern Analysisand Machine Intelligence vol 24 no 7 pp 971ndash987 2002

[44] E Candes L Demanet D Donoho and L Ying ldquoFast discretecurvelet transformsrdquo Multiscale Modeling amp Simulationvol 5 no 3 pp 861ndash899 2006

[45] A Dudhane G Shingadkar P Sanghavi B Jankharia andS Talbar ldquoInterstitial lung disease classification using feedforward neural networksrdquo in Proceedings of Advances in Intel-ligent Systems Research vol 137 pp 515ndash521 2017

[46] A A Dudhane and S N Talbar ldquoMulti-scale directional maskpattern for medical image classification and retrievalrdquo inProceedings of 2nd International Conference on ComputerVision amp Image Processing Indian Institute of TechnologyRoorkee Roorkee India 2018

[47] httpeurekasveriacin

16 Journal of Healthcare Engineering

International Journal of

AerospaceEngineeringHindawiwwwhindawicom Volume 2018

RoboticsJournal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Active and Passive Electronic Components

VLSI Design

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Shock and Vibration

Hindawiwwwhindawicom Volume 2018

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawiwwwhindawicom

Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Control Scienceand Engineering

Journal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom

Journal ofEngineeringVolume 2018

SensorsJournal of

Hindawiwwwhindawicom Volume 2018

International Journal of

RotatingMachinery

Hindawiwwwhindawicom Volume 2018

Modelling ampSimulationin EngineeringHindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Navigation and Observation

International Journal of

Hindawi

wwwhindawicom Volume 2018

Advances in

Multimedia

Submit your manuscripts atwwwhindawicom