A Robust Deep Neural Network Based Breast Cancer Detection And Classification

Added on - Nov 2020

Trusted by 2+ million users,
1000+ happy students everyday
Showing pages 1 to 4 of 16 pages
A Robust Deep Neural Network Based Breast CancerDetection And ClassificationAbstractThe exponential rise in breast cancer cases acrossthe globe has alarmed academia-industries to achieve certain moreefficient and robust Breast Cancer Computer Aided Diagnosis (BC-CAD) system for breast cancer detection. A number of techniqueshave been developed with focus on case (i.e., data type) centricsegmentation, feature extraction and classification of breast cancerHistopathological images. However, rising complexity and accuracyoften demands more robust solution. Recently, ConvolutionalNeural Network (CNN) has emerged as one of the most efficienttechniques for medical data analysis and various imageclassification problems. In this paper a highly robust and efficientBC-CAD solution has been proposed. Our proposed systemincorporates pre-processing, enhanced adaptive learning basedGaussian Mixture Model (GMM), connected component analysisbased region of interest localization, AlexNet-DNN based featureextraction.The Principle Component Analysis (PCA) and LinearDiscriminant Analysis (LDA) based on feature selection which isused as dimensional reduction. One of the advantages of theproposed method is that none of the current dimensional reductionalgorithms employed with SVM to perform breast cancer detectionand classification. The overall results obtained signify that theAlexNet-DNN based features at fully connected layer; FC6 inconjunction with LDA dimensional reduction and SVM basedclassification outperforms other state-of-art techniques for breastcancer detection. The proposed BC-CAD system has beenperformed over real world data BreakHis having significantdiversity and complexity, and therefore we suggest it to be used forother real-world applications.The proposed method achieved 96.15for AlexNet-FC6 and 96.20 for AlexNet-FC7 in term of evaluationmeasures.KeywordsBreast Cancer Detection; Computer AidedDiagnosis; Convolutional Neural Network; AlexNet-DNN; LinearDiscriminant Analysis.INTRODUCTIONCancer is a type of disease involving the development andgrowth of abnormal cell that invades or spreads to other parts of ahuman body. In last few decades, cancer has emerged as one ofthe deadliest heath diseases claiming huge death rate. There aredifferent types of cancers such as breast cancer, blood cancer,melanoma or skin cancer...etc. However, in last few years, breastcancer has emerged as the second most common cancer type afterskin cancer in women. Literatures reveal that almost 50% of thebreast cancer cases are found in the developing countries, whileapproximate 58% of deaths take place in less developedcountries. It is estimated that around the world over 5,08,000women died in 2011 due to breast cancer [1]. Various casestudies [1] have revealed that the earlier cancer identification andits diagnosis can play vital role in providing success diagnosis ortreatment. A similar survey [2] also indicated high pace rise indeaths caused due to breast cancerin2012, which is evenincreasing with vast pace [2]. The study revealed that the deathrate can reach up to 27 million till 2030 [3]. The exponential risein breast cancer has alarmed academia-industries to developcertain vision based approach for earlier breast cancer detectionand diagnosis. The rise in advanced computing techniques, visionbased computations and decision process...etc. has given rise to anew dimension having immense potential to meet majordemands. Non-deniably such development often gives a hope forbetter human life, health and decision process. In fact, thedevelopment of science and technology often intends to enablehuman life secure, healthy and productive. On the other hand,gigantically rising population across world requires optimalhealthcare solutions. In fact, the traditional manual diagnosisprocesses are either confined or less productive to meet thedemands and enable optimal healthcare solution. This as a resulthas motivated academia-industries to achieve more efficientcomputer aided diagnosis (CAD) solutions for earlier health-diagnosis. The development of Histopathological analysis andmolecular biology has made vision based breast cancer CADsystem more efficient [4]. Unlike traditional manual microscopicanalysis based approaches vision based CAD system can be ofparamount significance for breast cancer detection andclassification [5].In last few years, efforts have been made to exploit X-raymammography and ultrasound technique to assist breast cancerdetection and diagnosis; however, mammography usuallyexhibits poor sensitivity to the minute size cancers. In addition, itis insignificant especially for augmented breast or dense breastconditions [60]. Similarly, poor specificity and the complexity inwhole breast-imaging process make ultrasound method confinedand ineffective. As an enhanced solution, authors proposedMagnetic Resonance Imagining (MRI) systems that possessbetter soft tissue resolution capacity and enhanced sensitivity forbreast cancer detection [6].The MRI was method used to avoidany ionization radiation and hence it can be suitable for patientswith implants. The key usefulness of MRI method is its ability ofquantification of tumor volume, multi-focal and multi-centricbreast cancer detection [7]. Vision based CAD systems have been1
found efficient to perform more time efficient and accurate breastcancer detection to assist early diagnosis. Exploring in depth ofthe CAD systems, it can be found that its efficacy primarilydepends on the efficiency of the Region Of Interest (ROI)segmentation, feature extraction, feature mapping andclassification techniques. There are a number of techniques, suchas wavelet transform [8][9][10], Gabor transform [11], appliedfor feature extraction while classification is done using standardsupport vector machine (SVM) or other artificial intelligenceapproaches [9][12][13]. However, most of the traditionalapproaches are confined due to limited feature extractionefficiency and classification for lesion of different shape, size,and orientation density. The low information availability tooconfines the efficacy of the traditional feature extraction andcancer classification models. In addition, there is an inevitableneed to develop efficient approaches to deal with highdimensional features, feature selection and classification orprediction. Considering large size unannotated data (i.e., breastcancer image data), the use of CNN as feature extraction tool canensure better breast cancer CAD (BC-CAD) solution.In this paper, a highly robust and efficient CNN based DeepNeural Network (DNN) model has been developed for breastcancer detection and classification. To enable an optimalsolution, our proposed model encompasses various enhancementsincluding pre-processing, ROI identification, Caffe-enet-AlexNetDNN based feature extraction, PCA and LDA based featureselection and SVM based classification. To examine theeffectiveness of the proposed DNN model, a parallel BC-CADmodel using Spatial Invariant Fourier Transform (SIFT) featureextraction has been developed. In addition, to perform breastcancer prediction, a two class classification model is developed;where the overall result affirms that the AlexNet-DNN with LDAfeature extraction outperforms other state-of-art techniques forBC-CAD purpose.The other sections of the presented manuscript are divided asfollows: Section two presents the related work, which is followedby the discussion of the proposed AlexNet DNN based breastcancer detection and classification in Section three. Section fourpresents the results and discussion, while the overall researchconclusion is presented in Section five.RELATED WORKThis section briefs some of the key literatures pertaining tothe breast cancer detection and related technologies.Rehmanet al.[12] derived a diverse feature based breastcancer detection model, where they applied phylogenetic trees,various statistical features and local binary patterns forgenerating certain distinctive and discriminative features toperform classification. Authors applied radial basis functionbased SVM for two class classification: cancerous and non-cancerous. Pramaniket al. [8] assessed the breast thermogram toperform cancer detection, where they at first applied InitialFeature point Image (IFI) to perform feature extraction for eachsegmented breast thermogram. They applied DWT to performfeature extraction, which was followed by feature classificationusing feed-forward Artificial Neural Network (ANN). A similareffort was made in [9], where authors applied wavelet andcounterlet transformation for feature extraction and SVMclassifier based classification of the mammograms. Yousefiet al.[11] used multi-channel Gabor wavelet filter for featureextraction over mammograms. Authors applied Gabor filterdesign to perform feature extraction, which was then followed bytwo class classifications using Bayesian classifier. Caorsi et al.[14] proposed an ANN based radar data processing model forbreast cancer detection where they emphasized not only on thedetection of the cancerous tumour but also its geometric featuressuch as depth and width. A similar effort was made in [15],where authors proposed a data-driven matched field model toassist microwave breast cancer detection. George et al. [16]developed robust and intelligent breast cancer detection andclassification model using cytological images. Authors applieddifferent classifiers including back-propagation based multilayerperceptron, Probabilistic Neural Network (PNN), SVM andlearning vector quantization to perform breast cancerclassification. Considering mitosis as a cancer signifier,Tashketal.[17] developed an automatic BC-CAD model, where at firstthey applied 2-D anisotropic diffusion for noise removal andimage morphological process. To achieve pixel-wise featuresfrom the ROI (i.e., mitosis region), authors derived a statisticalfeature extraction model. In addition, to alleviate the issue ofmisclassification of mitosis and non-mitosis objects, authorsdeveloped an object-wise Completed Local Binary Pattern(CLBP) that enabled efficient textural features extraction. Thepredominant novelty of the CLBP model was that it is robustagainst positional variation, color changes ...etc. Authors appliedSVM to perform feature classification. A space processing modelwas developed in [18] that enabled time reversal imagingapproach to perform breast cancer detection. To performmalignant lesion detection, authors applied FDTD breast modelthat comprises dense breast tissues with differing fibro-glandulartissue composition. To achieve better accuracy, Naeemabadi etal. [13] applied LS and SMO techniques rather applyingtraditional SVM for classification.Li et al.[19] developed anMFSVM-FKNN ensemble classifier for breast cancer detection,where authors applied the concept of mixture membershipfunction. Fuzzy classifier based breast cancer detection andclassification model was proposed in [20]. A bio-inspiredimmunological scheme was proposed for mammographic massclassification that classifies malignant tumors from the benignones.Gaike et al.[21] at first focused on retrieving the higherorder (particularly third order) features that was later processedfor clustering to perform malignant breast cancer detection.Alietal. [22] developed an automatic segmentation model using theconcept of the data acquisition protocol parameter over the imagestatistics of DMR-IR database. A similar effort was made in [23],where authors developed a two phase BC-CAD system. In theirapproach at first they applied Neutrosophic Sets (NS) andoptimized Fast Fuzzy C-Mean (F-FCM) algorithm for ROIsegmentation, which was followed by classification to performtwo class classifications, i.e. normal and abnormal tissue.2
Ismahan et al.[24] used a mathematical morphology concept toperform masses detection over digitized mammograms. Textureanalysis based ROI detection for thermal imaging based BC-CAD was performed in [25]. Considering the significance offeature extraction for mammograms classification, Sanae et al.[10] exploited comprehensive statistical Block-Based features,which were extracted from sub-bands of the discrete wavelettransformation. Once mapping the extraction features, SVM wasapplied to perform classification. Unlike traditional approaches,Maken et al.[26] examined the efficacy of Multiple InstanceLearning (MIL) algorithm for mammograms classification,where they developed MIL using tile-based spatio-temporalfeatures.Rosaet al. [27] andHatipogluet al. [28] intended toexploit MIL in conjunction with SVM for mammogramsclassification. Non-deniably, MIL approaches have performedbetter, particularly for unannotated data; however the likelihoodof better performance with deep features cannot be ignored.To enable more efficient feature extraction and accurateperformance, DNN has gained significant attention. CNN hasgained widespread recognition to extract features from thecomplex image data, such as mammograms, MRI medical data oreven histopathological datasets. CNN techniques which areinspired biologically by the organization of human visual cortexpossess robust efficiency due to its ability to learn featuresinvariant to translation, rotation and shifting is their greatadvantage.Spanholet al. [29] applied CNN approach to classifybreast cancer Histopathological images. Their proposed modelperformed learning CNN by means of training patches generatedthrough varied approaches. Authors obtained the bestclassification accuracy of 89.6% for images with 40xenlargement. To achieve this result, authors considered thepatches of the size 64x64 pixels.Hatipogluet al. [28] used CNNto classify cellular and non-cellular structures in breast cancerhistopathological images. Authors applied neural network toperform classification, where they achieved the best classificationaccuracy of 86.88%. Recently,Wanget al. [30] used GooglenetCNN model to perform Histopathological image classification,where they achieved the accuracy of 98.4% patch classification.To detect the invasive ductal carcinoma tissue in histologicalimages for BC-CAD, Cruz-Roa et al.[31] applied CNN featureextraction approach. Ertosun et al. [32] developed BC-CADmodel using Deep Learning with three distinct CNN models so asto localize masses in mammography images. To enhanceaccuracy of segmentation or ROI identification, authorsincorporated additional novelties such as cropping, translation,rotation, flipping and scaling techniques.Arevalo et al.[33]applied CNN based feature extraction followed by SVM basedclassification for BC-CAD solution. Authors achieved ReceiverOperating Characteristics (ROC) of 86%. A similar work wasperformed in [34], where authors achieved classificationaccuracy of 96.7%. Russakovsky et al. [35] and Zuiderveld et al.[40] trained CNN over ImageNet to perform breast cancerclassification. Authors Abdel-Zaheret al. [36] andWanget al.[41] performed breast cancer classification, developing aclassifier by means of the weights of a previously trained DeepBelief Network (DBN). They applied Levenberg Marquardtlearning based ANN to perform classification. Amongst thevarious researches it has been realized that in addition to the ROIidentification and feature extraction efficiency, selecting optimalfeatures is equally significant to ensure efficient performance. Toachieve this, dimensional reduction and feature selectionmeasures are of utmost significance. With these objectives, Olfatiet al. [37] applied Genetic Algorithm (GA) to enhanced PCA byselecting optimal principal componentsanalysis(GA-SPCA).They applied PCA for dimension reduction, GA for featureselection [38] and SVM for classificationwhich facilitates toachieve better performance in order to gain desired results.The developed system is much effective as it is prepared byutilizing advanced technology but it has a risk which may createmajor problems due to minor technical issues. Moreover, it issuggested that they can utilizeComputerized tomography (CT)scan andPositron emission tomography (PET) scan to detrmineactual situation of breats cancer appropriately.I.PROPOSED METHODOLOGYThis section primarily discusses the proposed researchwork and implementation model to achieve intended BC-CADsolution.This solution provide an accurate breast cancerdetection to assist early diagnosis in more efficient way. Mycontribution is all about to make highly robust and efficient CNNbased on a model named as Deep Neural Network (DNN) model.Moreover, it has been developed for breast cancer detection andclassification.In the current research work, a novel deep learningbased Breast Cancer detection and classification algorithm isdeveloped. In our work, the predominant emphasis is made ondeveloping a novel and robust automated CAD system for BreastCancer (BC) detection and classification.Our research methodapplied on multi-dimensional shape to increase the performanceof the proposed BC-CAD system, where enhancement has beenmade for major functional processes including pre-processing,ROI detection and segmentation, feature extraction, featureselection and classification. In addition, a standard breast cancerdataset named BreaKHis [39], which is the Histopathologicalimage data for breast cancer, has been considered for our study.To ensure better efficiency enriching input data or medical imagequality is often a better solution. With this objective, the inputHistopathological images are processed for noise removal andresizing. Once retrieving the suitable input images, it has beenfurther processed for ROI segmentation, where an enhancedGMM approach has been taken into consideration. Unlikegeneric GMM algorithm, we have derived an adaptive learningbased Gaussian approach that enables swift and accurate ROIidentification. This as a result can play vital role in accurate andsignificant feature retrieval for optimal BC-CAD solution.Noticeably, our applied datasets BreaKHis [39] possesses keyfeatures as marked cellular atypia, mitosis, disruption ofbasement membranes, metastasize ...etc. These features havebeen applied to perform two class classifications. Realizing thenon-deniable fact that inaccurate ROI localization andinsignificant pixel conjuncture might lead inaccurate3
classification accuracy, we have applied Connected ComponentAnalysis (CCA) model over segmented region. It eliminatesthose image components that percept to be connected with thetarget ROI; however, it is insignificant toward BC-CADclassification. Once obtaining the CCA processed segmentedROI, we have processed it for feature extraction followed byfeature selection and classification to achieve anticipated BC-CAD solution. Unlike existing feature extraction models, such asGabor filter [11], Wavelet Transform [8][9][10], MIL [26-28] oreven classical CNNs [28-33][35], in our research a CNN modelis derived. In the proposed research work, an enhanced DNNmodel named AlexNet DNN has been applied to extractsignificant features from the Histopathological image datasets.Realizing the fact that high dimensional features often providemore significant information to make precise classification oranalysis; in our research work, we have developed AlexNet-DNNto extract high dimensional features with 4096 Kernels ordimensions. The proposed AlexNet-DNN model comprises fiveconvolutional layers in succession with three Fully Connected(FC) layers (FC6, FC7 and FC8). The feature extracted at thehigher layer provides more significant information that helps inachieving accurate BC-CAD performance. Considering majorclassical approaches, where the probability of over-fitting andaccuracy is often ignored, we have applied a supplementarymodel called Caffe-enet that assists AlexNet-DNN to overcomeexisting limitations and enables its (AlexNet-DNN) executionover general purpose computers without any need ofsophisticated Graphical Processing Units (GPUs). The use ofCaffe-enet in conjunction with AlexNet-DNN plays a vital role inassuring efficient feature extraction even under large scaleunannotated data. It makes a suitable environment for majorbreast cancer detection applications. Specifically, owing to highunannotated breast cancer data, performing DNN learning andfurther classification is a mammoth task. Hence to alleviate suchissues, in our proposed model AlexNet applied multilayeredDNN architecture, where at each layer we retrieve the featuresfor further mapping and classification. However, to achieve moreefficient results, we have considered 4096-dimensional featuresextracted at the Fully Connected (FC) layers, FC-6 and FC-7 ofthe AlexNet-DNN. The detailed discussion of the developedDNN model and its implementation for BC-CAD is given in thenext section. In addition, to the proposed AlexNet-DNN basedBC-CAD solution, in this research we have developed a parallelfeature extraction model, called Scale Invariant FourierTransform (SIFT). We have applied Fully-Connected (FC)layers: FC6 and FC7 features, which are 4096-dimensionalfeatures and hence amounts a huge data to process. To enhancecomputational efficiency, in this research work, we have appliedPCA and LDA techniques to perform dimension reduction orfeature selection. Once processing for the dimensional reduction,the selected features are projected to polynomial kernel basedclassifier that performs two class classifications: Malignant andBenign. To further strengthen the efficiency of our proposedwork, 10-fold cross validation approach is applied that ensuresoptimal classification accuracy for BC-CAD solution.The overall proposed research methodology andimplementation schematic is given in the next section.PROPOSED SYSTEM DESIGNIn this section, the detailed discussion of the proposedresearch work and implementation schematic is presented. Beforediscussing the proposed BC-CAD solution, introducing keyterminologies is must. Table I presents the nomenclature ofdifferent abbreviated terms.BCDataPre/PostProcessingBreast CancerregionSegmentationSIFTCaffee-AlexNetDNNFC6FC7AlexNet-FC6-LDAAlexNet-FC6-PCAAlexNet-FC7-LDAAlexNet-FC7-PCASIFT-FV--PCASIFT-FV-LDAFig. 1. Propsoed BC-CAD SystemA detailed discussion of the proposed research model isgiven, as presented in Fig. 2, as follows:.aBreast Cancer Histopathological Image DataCollection.bPre-processing.cROI Detection and Localization.dFeature Extraction.eDimensional reduction of Feature Selection,and.fBC-CAD Two-Class Classification.Which will be discussed in details:1.Breast Cancer Histopathological Image DatacollectionIn this research or study, we have used standard datasetsnamed Breast Cancer Histopathological Image Classification(BreakHis) [39] that contains a total of 9,109 Histopathological(say, microscopic) images of breast cancer tissue. The historicalperspective of the BreaKHis states that the microscopic imagesare retrieved from 82 patients. BreakHis dataset has been built inassociation with the P&D Laboratory- Pathological Anatomyand Cytopathology, Parana, Brazil (http://www.prevencaoediagnose.com.br).Unlike generic singe size or singlefeature datasets, BreaKHis data are obtained with varied4
Desklib Logo
You are reading a preview
Upload your documents to download or

Become a Desklib member to get access