Research Article Detection and Classification of...

Research ArticleDetection and Classification of Cancer fromMicroscopic Biopsy Images Using Clinically Significant andBiologically Interpretable Features

Rajesh Kumar Rajeev Srivastava and Subodh Srivastava

Department of Computer Science and Engineering Indian Institute of Technology (Banaras Hindu University) Varanasi 221005 India

Correspondence should be addressed to Rajesh Kumar rajeshrscse12iitbhuacin

Received 11 May 2015 Revised 3 July 2015 Accepted 12 July 2015

Academic Editor Ying Zhuge

Copyright copy 2015 Rajesh Kumar et al This is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

A framework for automated detection and classification of cancer from microscopic biopsy images using clinically significant andbiologically interpretable features is proposed and examined The various stages involved in the proposed methodology includeenhancement of microscopic images segmentation of background cells features extraction and finally the classification Anappropriate and efficient method is employed in each of the design steps of the proposed framework after making a comparativeanalysis of commonly used method in each category For highlighting the details of the tissue and structures the contrast limitedadaptive histogram equalization approach is used For the segmentation of background cells 119896-means segmentation algorithm isused because it performs better in comparison to other commonly used segmentation methods In feature extraction phase it isproposed to extract various biologically interpretable and clinically significant shapes as well as morphology based features fromthe segmented imagesThese include gray level texture features color based features color gray level texture features Lawrsquos TextureEnergy based features Tamurarsquos features andwavelet features Finally the119870-nearest neighborhoodmethod is used for classificationof images into normal and cancerous categories because it is performing better in comparison to other commonly used methodsfor this application The performance of the proposed framework is evaluated using well-known parameters for four fundamentaltissues (connective epithelial muscular and nervous) of randomly selected 1000 microscopic biopsy images

1 Introduction

Cancer detection has always been a major issue for thepathologists and medical practitioners for diagnosis andtreatment planningThemanual identification of cancer frommicroscopic biopsy images is subjective in nature and mayvary from expert to expert depending on their expertiseand other factors which include lack of specific and accuratequantitative measures to classify the biopsy images as normalor cancerous one The automated identification of cancerouscells from microscopic biopsy images helps in alleviatingthe abovementioned issues and provides better results if thebiologically interpretable and clinically significant featurebased approaches are used for the identification of disease

About 32of Indian population gets cancer at some pointduring their life time Cancer is one of the common diseasesin India which has responsibility to maximum mortality

with about 03 million deaths per year [1] The chances ofgetting affected by this disease are accelerated due to changein habits in the people such as increase in use of tobaccodeterioration of dietary habits lack of activities and manymore The possibility of cure from cancer is increased due torecent combined advancement in medicine and engineeringThe chances of curing from cancer are primarily in itsdetection and diagnosis The selection of the treatment ofcancer totally depends on its level of malignancy Medicalprofessionals use several techniques for detection of cancerThese techniques may include various imaging modalitiessuch as X-ray Computer Tomography (CT) Scan PositronEmission Tomography (PET) Ultrasound and MagneticResonance Imaging (MRI) and pathological tests such asurine test and blood test

For accurate detection of cancer pathologists usehistopathology biopsy images that is the examination of

Hindawi Publishing CorporationJournal of Medical EngineeringVolume 2015 Article ID 457906 14 pageshttpdxdoiorg1011552015457906

2 Journal of Medical Engineering

microscopic tissue structure of the patient Thus biopsyimage analysis is a vital technique for cancer detection [2 3]Histopathology is the study of symptoms and indications ofthe disease using the microscopic biopsy images To visualizevarious parts of the tissue under a microscope the sectionsare dyed with one or more staining components The maingoal of staining is to reveal the components at cellular leveland counterstains are used to provide color visibility andcontrast Hematoxylin-Eosin (HampE) is staining componentthat has been used by pathologists for over few decadesHematoxylin stains cell nuclei which are blue in color whileEosin stains cytoplasm and connective tissues which are ofpink color The histology [4] is related to the study of cellsin terms of structure function and interpretations of thetissue and cells Microscopic biopsies are most commonlyused for both disease screenings because of the less invasivenatures The characteristic of microscopic biopsy images haspresence of isolated cells and cell clusters The microscopicbiopsy images are easier to analyze specimens compared tohistopathology due to absence of noncomplicated structures[5] The accurate manual identification of cancer frommicroscopic biopsy images has always been a major issue bythe pathologists and medical practitioners observing cell ortissue structure under the microscope

In histopathology the cancer detection process normallyconsists of categorizing the image biopsy into cancerousone or noncancerous one [6] In microscopic biopsy imageanalysis doctors and pathologists observe many of theabnormalities and categorize the sample based on variouscharacteristics of the cell nuclei such as color shape size andproportion to cytoplasmHigh resolutionmicroscopic biopsyprovides reliable information for differentiating abnormaland normal tissues The difference between normal andcancerous cells is shown in Table 1 [7]

For the detection and diagnosis of cancer from micro-scopic biopsy images the histopathologists normally lookat the specific features in the cells and tissue structuresThe various common features used for the detection anddiagnosis of cancer from the microscopic biopsy imagesinclude shape and size of cells shape and size of cell nucleiand distribution of the cells The brief descriptions of thesefeatures are given as follows

(A) Shape and Size of the Cells It has been observed that theoverall shape and size of cells in the tissues aremostly normalThe cellular structures of the cancerous cells might be eitherlarger or shorter thannormal cellsThenormal cells have evenshapes and functionality Cancer cells usually do not functionin a useful way and their shapes are often not even

(B) Size and Shape of the Cellrsquos Nucleus The shape and sizeof the nucleus of a cancer cell are often not normal Thenucleus is decentralized in the cancer cells The image ofthe cell looks like an omelet in which the central yolk isthe nucleus and the surrounding white is the cytoplasmThe nuclei of cancer cells are larger than the normal cellsand deviated from the centre of the mass The nucleusof cancer cell is darker The segmentation step mainlyfocuses on separation of regions of interests (cells) from

background tissues as well as separation of nuclei from cyto-plasm

(C) Distribution of the Cells in Tissue The function of eachtissue depends on the distribution and arrangements of thenormal cells The numbers of healthy cells per unit area areless in the cancerous tissues These adjectives of microscopicbiopsy images have been included in shape and morphologybased features texture features color based features ColorGray Level Cooccurrence Matrix (GLCM) Lawrsquos TextureEnergy (LTE) Tamurarsquos features and wavelet features whichare more biologically interpretable and clinically significant

The main aim of this paper is to design and develop aframework and a software tool for automated detection andclassification of cancer frommicroscopic biopsy images usingthe abovementioned clinically significant and biologicallyinterpretable features This paper focuses on selecting anappropriate method for each design stage of the frameworkafter making a comparative analysis of the various commonlyusedmethods in each categoryThe various stages involved inthe proposed methodology include enhancement of micro-scopic images segmentation of background cells featuresextraction and finally the classification

The rest of the paper has been structured as followsSection 2 describes the related works Section 3 presents themethods and models Section 4 describes the results anddiscussions and finally Section 5 draws the conclusion of thework presented in this paper

2 Related Works

In recent years few works have been reported in the lit-erature for the design and development of tools for auto-mated cancer detection from microscopic biopsy imagesKumar and Srivastava [9] presented detailed reviews onthe computer aided diagnosis (CAD) for cancer detectionfrom microscopic biopsy images Demir and Yener [10] alsopresented a method for automatic diagnosis of biopsy imageThey presented a cellular level diagnosis system using imageprocessing techniques Bhattacharjee et al [11] presented areview on computer aided diagnosis system to detect cancerfrom microscopic biopsy images using image processingtechniques

Bergmeir et al [12] proposed a model to extract thetexture features by using local histograms and GLCM Thequasisupervised learning algorithm operates on two datasetsthe first one having normal tissues labeled only indirectlyand the second one containing an unlabeled collection ofmixed samples of both normal and cancer tissues Thismethod was applied on the dataset of 22080 vectors withreduced dimensionality 119 from 132 The regions havingthe cancerous tissues were accurately identified having truepositive rate 88 and false positive rate 19 respectively byusing manual ground truth dataset

Mouelhi et al [13] used Haralickrsquos textures features [14]histogram of oriented gradients (HOG) and color compo-nent based statistical moments (CCSM) features selectionand extraction approaches to classify the cancerous cells frommicroscopic biopsy images The various features used in this

Journal of Medical Engineering 3

Table 1 Difference between normal and cancerous cells [7]

Normal cells Cancerous cells Description ofcancerous cells

Large and variablyshaped nuclei

Many dividing cellsand disorganizedarrangements

Variation in size andshape of nuclei

Loss of normalfeature (shape and

morphology)

paper are contrast correlation energy homogeneity GLCMtexture features [14] RGB gray level and HSV

Huang and Lai [15] presented a methodology for seg-mentation and classification techniques for histology imagesbased on texture features and by using SVM the maximumclassification accuracy obtained is 928

Landini et al [16] presented a method for morphologiccharacterization of cell neighborhoods in neoplastic and pre-neoplastic tissue of microscopic biopsy images In this paperauthors presented watershed transforms to compute the celland nuclei area and other parameters The distance measureof the neighborhood value has been used for calculating theneighborhood complexity with reference to the v-cells Thebest classification which has been obtained by 119870NN classifieris 83 for dysplastic and neoplastic classes and 58 of correctclassification

Sinha and Ramkrishan [17] extracted some features ofmicroscopic biopsy images which include eccentricity arearatio compactness average values of color componentsenergy entropy correlation and area of cells and nucleusThe classification accuracy obtained by Bayesian 119870-nearestneighbor neural networks and support vector machine was823 7060 941 and 941 respectively

Kasmin et al [18] extracted the features of microscopicbiopsy images including area perimeter convex area soliditymajor axis length orientation filled area eccentricity ratioof cell and nucleus area circularity and mean intensity ofcytoplasm The 119870NN and neural network classifier are usedfor classification accuracy 86 and 92 respectively

In this paper a framework for automated detection andclassification of cancer frommicroscopic biopsy images usingclinically significant and biologically interpretable featuresis proposed and examined For segmentation of imagescolour 119896-means based method is used The various hybridfeatures which are extracted from the segmented imagesinclude shape and morphological features GLCM texturefeatures Tamura features Lawrsquos Texture Energy based fea-tures histogram of oriented gradients wavelet features andcolor features For classification purposes 119896-nearest neighborbased method is proposed to be used The efficacy of otherclassifiers such as SVM random forest and fuzzy 119896-meansis also examined For testing purposes 2828 microscopicbiopsy images available from histology database [8] are usedFrom the obtained results it was observed that the proposedmethod is performing better in comparison to othermethodsdiscussed as above The overall summary and comparison of


the proposed method and other methods are presented inTable 6 in Section 4 of results and analysis

3 Methods and Models

The detection and classification of cancer from microscopicbiopsy images are a challenging task because an imageusually contains many clusters and overlapping objectsThe various stages involved in the proposed methodologyinclude enhancement of microscopic images segmentationof background cells features extraction and finally theclassification For the enhancement of themicroscopic biopsyimages the contrast limited adaptive histogram equalization[19 20] approach is used and for the segmentation ofbackground cells 119896-means segmentation algorithm is usedIn feature extraction phase various biologically interpretableand clinically significant shape and morphology based fea-tures are extracted from the segmented images which includegray level texture features color based features color graylevel texture features Lawrsquos Texture Energy (LTE) basedfeatures Tamurarsquos features and wavelet features Finally the119870-nearest neighborhood (119870NN) fuzzy 119870NN and supportvector machine (SVM) based classifiers are examined forclassifying the normal and cancerous biopsy images Theseapproaches are tested on four fundamental tissues (connec-tive epithelial muscular and nervous) of randomly selected1000microscopic biopsy images Finally the performances ofthe classifiers are evaluated using well known parameters andfrom results and analysis it is observed that the fuzzy 119870NNbased classifier is performing better for the selected featuresset The flowchart for the proposed work is given in Figure 1

31 Enhancements The main purpose of the preprocessingis to remove a specific degradation such as noise reductionand contrast enhancement of region of interests The biopsyimages acquired from microscope may be defective anddeficient in some respect such as poor contrast and unevenstaining and they need to be improved through process ofimage enhancement which increases the contrast betweenthe foreground (objects of interest) and background [21]Thecontrast limited adaptive histogram equalization (CLAHE)[20] approach is used for enhancement ofmicroscopic biopsyimages Figure 2 shows the original and enhanced imageusing contrast limited adaptive histogram equalization

32 Segmentation Several segmentation methods have beenadapted for cytoplasm cell and nuclei segmentation [22]frommicroscopic biopsy images like threshold based region-based and clustering based algorithms However the selec-tions of segmentationmethods depend on the type of the fea-tures to be preserved and extracted For the segmentation ofROI (region of interest) the ground truth (GT) of the imagesis manually cropped and created from histology dataset [8]The 119896-means clustering based segmentation algorithms areused because of the preservation of the desired informationFrom the obtained results through experimentation it isobserved that the clustering based algorithms specifically 119896-means based method are the best suited for microscopic

Noncancerous

Preprocessing(enhancement and restoration)

Segmentation(segmentation of ROI and background)

Feature extraction(texture shape LTE wavelet HOG

color based features etc)

Classification

Cancerous

Microscopic biopsy image

Figure 1 Model of automated cancer detection from microscopicbiopsy images

biopsy images Figure 3 shows the original and 119896-meanssegmented microscopic biopsy image For testing and exper-imentation purpose twenty (20) microscopic biopsy imagesavailable from histology dataset [8] were used These imageswere randomly selected for segmentation The ground truth(GT) images are manually created by cropping the region ofinterest (ROI) The visual results of texture based segmenta-tion FCM segmentation 119870-means segmentation and colorbased segmentation [20 23ndash26] are presented in Figures 3(a)to 3(d)Thus from the visual results obtained and reported inFigures 3(a) to 3(d) it is observed that the 119896-means clusteringbased segmentation method performs better in most of thecases as compared to other segmentation approaches underconsideration for microscopic biopsy image segmentation

Finally the ROI segmented image of microscopic biopsyis compared to ground truth images for the quantitativeevaluation of various segmentation approaches for all 20sample images from histology dataset The performanceof the various segmentation approaches such as 119870-means[27] fuzzy 119888-means [28] texture based segmentation [29]and color based segmentation [30] was evaluated in termsof various popular parameters of segmentation measuresThese parameters include accuracy sensitivity specificityfalse positive rate (FPR) probability random index (RI)global consistency error (GCE) and variance of information(VOI)

The brief description of few of these performance mea-sures used in this paper is as follows

(i) Probability Random Index (PRI) Probability random indexis the nonparametric measure of goodness of segmentationalgorithms Random index between test (119878) and ground truth(119866) is estimated by summing the number of pixel pairs with


(a) (b)

Figure 2 The original (a) and enhanced microscopic biopsy image with CLAHE (b)

Table 2 Quantitative evaluation of segmentation methods on the basis of average values of various performance metrics for a set of 20microscopic images [8]

Accuracy Sensitivity Specificity FPR PRI GCE VOIColor 119896-means 0987799 0707025 0989218 0010782 0975985 0009205 0115479119896-means 0990444 0748991 0994933 0005067 0981119 0012839 010818FCM 0987008 0614717 0998235 0001765 0974447 0015902 0136348Texture based 097144 0306398 0990445 0009555 0944609 0029276 0250797

same label and number of pixel pairs having different labelsin both 119878 and 119866 and then dividing it by total number of pixelpairs Given a set of ground truth segmentations 119866119896 the PRIis estimated using (1) such that 119888119894119895 is an event that describes apixel pair (119894 119895) having same or different label in the test image119878test

PRI (119878test 119866119896)

=1

(1198732)sum

forall119894119895amp119894lt119895[119888119894119895119901119894119895 + (1 minus 119888119894119895) (1 minus 119901119894119895)]

(1)

(ii) Variance of Information (VOI) The variation of infor-mation is a measure of the distance between two clusters(partitions of elements) [31] Clustering with clusters isrepresented by a random variable 119883 119883 = 1 119896 suchthat 119875119894 = |119883119894|119899 119894 isin 119883 and 119899 = sum

119894119883119894 is the variation of

information between two clusters 119883 and 119884Thus VOI(119883 119884) is represented using

VOI (119883 119884) = 119867 (119883) = 119867 (119884) minus 2119868 (119883 119884) (2)

where 119867(119883) is entropy of 119883 and 119868(119883 119884) is mutual informa-tion between 119883 and 119884 VOI(119883 119884) measures how much thecluster assignment for an item in clustering 119883 reduces theuncertainty about the itemrsquos cluster in clustering 119884

(iii) Global Consistency Error (GCE) The GCE is estimatedas follows suppose segments 119904119894 and 119892119895 contain a pixel say 119901119896such that 119904 isin 119878119892 isin 119866where 119878 denotes the set of segments thatare generated by the segmentation algorithm being evaluated

and 119866 denotes the set of reference segments To begin witha measure of local refinement error is estimated using (3)and then it is used to compute local and global consistencyerrors where 119899 denotes the set of difference operation and119877(119909 119910) represents the set of pixels corresponding to region119909 that includes pixel 119910 Using (3) [31] the global consistencyerror (GCE) is computed using (4) where 119899 denotes the totalnumber of pixels of the image GCE quantify the amount oferror in segmentation (0 signifies no error and 1 indicates noagreement)

119864 (119904119894 119892119895 119901119896) =

10038161003816100381610038161003816119877 (119904119894 119901119896) 119877 (119892119895 119901119896)

100381610038161003816100381610038161003816100381610038161003816119877 (119904119894 119901119896)

1003816100381610038161003816

(3)

GCE (119878 119866) =1

119899minsum

119894

119864 (119878 119866 119901119894) sum

119894

119864 (119878 119866 119901119894) (4)

Table 2 and Figure 4 show the comparison of varioussegmentation algorithms on the basis of average accuracysensitivity specificity FPR PRI GCE and VOI for 20 sampleimages taken from histology dataset [8] From Table 2 andFigure 4 it is observed that 119896-means color 119896-means fuzzy119888-means and texture based methods are performing betterat par in terms of accuracy specificity and PRI segmenta-tion measures but except for 119896-means based segmentationmethods other methods are not performing better in termsof other important parameters Only the 119870-means basedsegmentation algorithm is associated with larger value ofaccuracy sensitivity specificity and random index (RI) andsmaller value of FPR GCE and VOI in comparison to othermethods and hence it is better in comparison to others


KM segmented blue nuclei

Original image Ground truth image ROI segmented image



Original image Ground truth image Cropped new segmented image

(a)

(A) (B)

(b)

(c)

(d)

Figure 3 Original (A) and segmented microscopic biopsy image with 119870-means segmentation approach (B) (a) Original ground truth andROI segmented by texture based segmentation (b) Original ground truth and ROI segmented by FCM segmentation (c) Original groundtruth and ROI segmented by 119896-means segmentation (d) Original ground truth and ROI segmented by color based segmentation


0

02

04

06

08

1

12

Color k-meansk-means

FCMTexture based

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

FPR

PRI

GCE VO

I

Figure 4 Comparisons of various segmentation methods on thebasis of average accuracy sensitivity specificity FPR PRI GCE andVOI for 20 sample images from histology dataset [8]

Hence 119896-means based segmentation is the only methodwhich performs better in terms of all parameters and that iswhy it is chosen as the segmentation method in the proposedframework for cancer detection from microscopic biopsyimages

33 Feature Extraction After segmentation of image featuresare extracted from the regions of interest to detect andgrade potential cancers Feature extraction is one of theimportant steps in the analysis of biopsy imagesThe featuresare extracted at tissue level and cell level of microscopicbiopsy images for better predictions To better capture theshape information we use both region-based and contour-based methods to extract anticircularity area irregularityand contour irregularity of nuclei as the three shape featuresto reflect the irregularity of nuclei in biopsy images Thecellular level feature focuses on quantifying the propertiesof individual cells without considering spatial dependencybetween them In biopsy images for a single cell the shapeand morphological textural histogram of oriented gradientsand wavelet features are extracted The tissue level featuresquantify the distribution of the cells across the tissue for thatit primarily makes use of either the spatial dependency of thecells or the gray level dependency of the pixels

Based on these characteristics some important shape andmorphological based features are explained as follows

(i) Nucleus Area (A) The nucleus area can be represented bynucleus region containing total number of pixels it is shownin

119860 =

119899

sum

119894=1

119898

sum

119895=1

119861 (119894 119895) (5)

where 119860 is nucleus area and 119861 is segmented image of 119894 rowsand 119895 columns

(ii) Brightness of Nucleus The average value of intensity of thepixels belonging to the nucleus region is known as nucleusbrightness

(iii) Nucleus Longest Diameter (NLD) The largest circlersquosdiameter circumscribing the nucleus region is known asnucleus longest diameter it is shown in

NLD = radic(1199091 minus 1199092)2

+ (1199101 minus 1199102)2 (6)

where 1199091 1199101 and 1199092 1199102 are end points on major axis

(iv) Nucleus Shortest Diameter (NSD) This is representedthrough smallest circlersquos diameter circumscribing the nucleusregion It is represented in

NSD = radic(1199092 minus 1199091)2

+ (1199102 minus 1199101)2 (7)

where 1199091 1199101 and 1199092 1199102 are end points on minor axis

(v) Nucleus Elongation This is represented by the ratio ofthe shortest diameter to the longest diameter of the nucleusregion shown in

Nucleus elongation =NLD

Perimeter (8)

(vi) Nucleus Perimeter (P) The length of the perimeter of thenucleus region is represented using

119875 = Even count + radic2 (odd count) unit (9)

(vii) Nucleus Roundness (120574) The ratio of the nucleus area tothe area of the circle corresponding to the nucleus longestdiameter is known as nucleus compactness shown in

120574 =119860

119875=

4120587 times Area1198752

(10)

(viii) Solidity Solidity is ratio of actual cellnucleus area toconvex hull area shown in

Solidity =Area

Convex Area (11)

(ix) EccentricityThe ratio ofmajor axis length andminor axislength is known as eccentricity and defined in

Eccentricity =Length of mejor AxisLength of minor Axis

(12)

(x) Compactness Compactness is the ratio of area and squareof the perimeter It is formulated as

Compactness =Area

Perimeter2 (13)


There are seven sets of features used for computing thefeature vector of microscopic biopsy images explained asfollows

(i) Texture Features (F1ndashF22) [32ndash34] Autocorrelation con-trast correlation cluster prominence cluster shade differ-ence variance dissimilarity energy entropy homogeneitymaximum probability sum of squares sum average sumvariance sum entropy difference entropy information mea-sure of correlation 1 information measure of correlation2 inverse difference (INV) inverse difference normalized(INN) and inverse difference moment normalized are majortexture features which can be calculated using equations ofthe texture features

(ii) Morphology and Shape Feature (F23ndashF32) In papers [3536] authors describe the shape and morphology featuresTheconsidered shape and morphological features in this paperare area perimeter major axis length minor axis lengthequivalent diameter orientation convex area filled areasolidity and eccentricity

(iii) Histogram of Oriented Gradient (HOG) (F33ndashF68) His-togram of oriented gradient is one of the good features set todeify the objects [32] In our observation it will be includedfor better and accurate identification of objects present inmicroscopic biopsy images

(iv) Wavelet Features (F69ndash100) Wavelets are small wavewhich is used to transform the signals for effective processing[3] The wavelets are useful in multiresolution analysis ofmicroscopic biopsy images because they are fast and givebetter compression as compared to other transforms TheFourier transform converts a signal into a continuous seriesof sine waves but the wavelet precedes it in both timeand frequency This accounts for the efficiency of wavelettransforms [37] Daubechies wavelets have been used becausethey have fractal structures and they are useful in the caseof microscopic biopsy images In this paper mean entropyenergy contrast homogeneity and sumofwavelet coefficientsare taken into consideration

(v) Color Features (F101ndashF106) The components of thesemodels are hue saturation and value (HSV) [34] Thisis represented by the six sided pyramids the vertical axisbehaves as brightness the horizontal distance from the axisrepresents the saturation and the angle represents the hueHere mean and standard deviation of HSV components aretaken as features

(vi) Tamurarsquos Features (F107ndashF109) Tamurarsquos features arecomputed on the basis of three fundamental texture featurescontrast coarseness and directionality [3] Contrast is themeasure of variety of the texture patternTherefore the largerblocks that make up the image have a larger contrast It isaffected by the use of varying black and white intensities[32] Coarseness is the measure of granularity of an image[32] thus coarseness can be represented using average sizeof regions that have the same intensity [38] Directionality is

Table 3 The distribution of various features extracted from imagesand their ranges

Name of features Number of features(range F1ndashF115)

Texture features 22 (F1ndashF22)Morphology and shape feature 10 (F23ndashF32)Histogram of oriented gradient (HOG) 36 (F33ndashF68)Wavelet features 32 (F69ndash100)Color features 6 (F101ndashF106)Tamurarsquos features 3 (F107ndashF109)Lawrsquos Texture Energy 16 (F110ndashF115)

the measure of directions of the grey values within the image[32]

(vii) Lawrsquos Texture Energy (LTE) (F110ndashF115) These featuresare texture description features which mainly used averagegray level edges spots ripples and wave to generate vectorsof the masks Lawrsquos mask is represented by the features ofan image without using frequency domain [39] Laws sig-nificantly determined that several masks of appropriate sizeswere very instructive for discriminating between differentkinds of texture features present in the microscopic biopsyimages Thus its classified samples are based on expectedvalues of variance-like squaremeasures of these convolutionscalled texture energy measures The LTE mask method isbased on texture energy transforms applied to the imageclassification used to estimate the energy within the passregion of filters [40]

Table 3 provides the distribution of name of the featuretype and the number of features selected for the classificationof microscopic biopsy images

34 Classification The classification of microscopic biopsyimages is themost challenging task for automatic detection ofcancer frommicroscopic biopsy images Classification mightprovide the answer whether microscopic biopsy is benignor malignant For classification purposes many classifiershave been used Some commonly used classificationmethodsare artificial neural networks (ANN) Bayesian classifica-tion 119870-nearest neighbor classifiers support vector machine(SVM) and random forest (RF) Supervised machine learn-ing approaches are used for the classification of microscopicbiopsy images There are various steps involved in thesupervised learning approaches First step is to prepare thedata (feature set) the second step is to choose an appropriatealgorithm the third step is to fit a model the fourth stepis to train the fitted model and then the final step is touse fitted model for predictionThe 119870-nearest neighborhood(119870NN) fuzzy 119870NN and support vector machine (SVM) andrandom forest classifiers are used for classifying the normaland cancerous biopsy images

4 Results and Discussions

The proposed methodologies were implemented with MAT-LAB 2013b on dataset of digitized at 5x magnification on


PC with 34GHz Intel Core i7 processor 2 GB RAM andwindows 7 platform

For the testing and experimentation purposes a totalof 2828 histology images from the histology image dataset(histologyDS2828) and annotations are taken froma subset ofimages related to above database [8]The image distributionsbased on the fundamental tissue structures in the histologydataset include Connective-484 Epithelial-804 Muscular-514 and Nervous-1026 microscopic biopsy images withmagnifications 25x 5x 10x 20x and 40x Although themethod ismagnification independent in this work the resultsare provided on samples digitized at 5x magnification Thefeatures extracted from microscopic biopsy images must bebiologically interpretable and clinically significant for betterdiagnosis of cancer Table 4 provides the brief description ofdataset used for identification of cancer from microscopicbiopsy images

The proposed methodology for detection and diagnosisof cancer detection from microscopic biopsy images consistsof the stages of images enhancement segmentation featureextraction and classification

The contrast limited adaptive histogram equalization(CLAHE) is used for enhancement of microscopic biopsyimages because it has ability to better highlight the regionsof interests in the images as tested through experimentation

To better preserve the desired information inmicroscopicbiopsy images during segmentation process the variousclustering and texture based segmentation approaches wereexamined For microscopic biopsy images it is required todiscover as much as possible the nuclei information in orderto make reliable and accurate detection and diagnosis basedon cells and nuclei parameters From results and analysispresented in Section 4 119896-means segmentation algorithm [40]was used for segmenting the microscopic biopsy images asit performs better in comparison to other methods Duringsegmentation process of 119896-means clustering method thenumber of clusters 119896 was set to 119896 = 3 Further to find thecenter of the clusters squared Euclidean distance measuresare used as similarity measures

In feature extraction phase various biologically inter-pretable and clinically significant shape and morphologybased features were extracted from the segmented imageswhich include gray level texture features (F1ndashF22) shapeand morphology based features (F23ndashF32) histogram oforiented gradients (F33ndashF68) wavelet features (F69ndashF100)color based features (F101ndashF106) Tamurarsquos features (F107ndashF119) and Lawrsquos Texture Energy (F110ndashF115) based featuresFinally a 2D matrix of 2828 times 115 feature matrix was formedusing all the feature sets where 2828 are the number ofmicroscopic images in the dataset and 115 are the totalnumber of features extracted

Randomly selected 1000 datasamples were used fortesting various classification algorithms The 10-fold crossvalidation approach was used to partition the data in trainingand testing setsThus 900 datasamples were used for trainingpurposes and 100 datasamples were used for testing pur-poses The 119870-nearest neighbor (119870NN) is a simple classifierin which a feature vector is assigned For 119870NN classificationthe numbers of nearest neighbor (119896) were set to 5 and

Table 4 Image distribution of fundamental tissues dataset of 2828histology images [8]

Fundamental tissue Number of imagesConnective 484Epithelial 804Muscular 514Nervous 1026Total 2828

Euclidean distance matrix and the ldquonearestrdquo rule to decidehow to classify the sample were used The proposed methodwas also tested by using support vector machine (SVM)based classifier for linear kernel function with 10-fold crossvalidationmethods In SVM classificationmodel the kernelrsquosparameters and soft margin parameter 119862 play vital rolein classification process the best combination of 119862 and 120574

was selected by a grid search with exponentially growingsequences of 119862 and 120574 Each combination of parameterchoices was checked using cross validations (10-fold) and theparameters with best cross validation accuracy were selectedFor SVMrsquos linear kernel function quadratic programming(QP) optimization parameter was used to find separatinghyperplane In the case of random forest the value by defaultis 500 trees and mtry = 10

The performance of classifiers was calculated using con-fusion matrix of size 2 times 2 matrix and the value of TPTN FP and FN was calculated The performance parametersaccuracy sensitivity and specificity were calculated using(14)ndash(19)

The fundamental definitions of these performance mea-sures could be illustrated as follows

Accuracy The classification accuracy of a technique dependsupon the number of correctly classified samples (ie truenegative and true positive) [40] and is calculated as follows

Accuracy =TP + TN

119873times 100 (14)

where 119873 is the total number of samples present in themicroscopic biopsy images

Sensitivity Sensitivity is a measure of the proportion ofpositive samples which are correctly classified [41] It can becalculated using

Sensitivity =TP

TP + FN (15)

where the value of sensitivity ranges between 0 and 1 where0 and 1 respectively mean worst and best classification

Specificity Specificity is a measure of the proportion ofnegative samples that are correctly classified [42] The valueof sensitivity is calculated using

Specificity =TN

TN + FP (16)


Table 5 Comparative performances of various classifiers for the chosen features for various tissue types

Accuracy Specificity Sensitivity BCR 119865-measure MCC Accuracy Specificity Sensitivity BCR 119865-

measure MCC

Connective tissues Epithelial tissuesRF 0907245 0993668 0493996 0743832 0647373 0642137 0849306 0966243 0555332 0760788 0675868 0609494SVM 089245 0888438 0948297 0918756 0538314 055879 0796998 07851 0898525 0842279 0472804 04587FYZZY119870NN 0787879 0867476 0370074 0618789 0356613 0231013 0665834 076465 0407057 0585984 0401181 017053

119870NN 0921909 0940164 0819922 0880263 0759395 0717455 0884727 0916446 0801733 0859435 0795319 071626Muscular tissues Nervous tissues

RF 0889878 0995023 0193145 0594084 0313309 037318 0843102 092827 0723262 0825766 0792403 0676888SVM 0884379 0886718 0786303 083681 0263764 0320547 0769545 0723056 0946068 0834923 0630126 0552038FUZZY119870NN 0614958 0672503 0535894 0604364 0538571 0208941 0808453 0882722 0242776 0562835 0225886 011837

119870NN 0897321 0923277 0650761 0787092 0543009 049783 0861763 0880866 0835733 0858482 0834116 0716492

Its value ranges between 0 and 1 where 0 and 1 respectivelymean worst and best classification

Balanced Classification Rate (BCR) The geometric mean ofsensitivity and specificity is considered as balance classifica-tion rate [43 44] It is represented by

BCR = radicSensitivity times Specificity (17)

F-Measure 119865-measure is a harmonic mean of precision andrecall It is defined by using

Precision =TP

TP + FP

Recall =TP

TP + FN

119865-measure = 2 timesPrecision times RecallPrecision + Recall

(18)

The value of 119865-measure ranges between 0 and 1 where 0means the worst classification and 1 means the best classifi-cation

Matthewsrsquos Correlation Coefficient (MCC) MCC is a measureof the eminence of binary class classifications [43] It can becalculated using the following formula

MCC

=TP times TN minus FP times FN

radic((TP + FN) (TP + FP) (TN + FN) (TN + FP))(19)

Its value ranges between minus1 and +1 where minus1 +1 and 0respectively correspond to worst best at random prediction

Discussions of Results Table 5 shows classification results ofthe proposed framework for four different tissues of micro-scopic biopsy images containing cancer and noncancer cases

tested using four popular classifiers like 119896-nearest neighborSVM fuzzy 119870NN and random forest

From Table 5 and Figure 5(a) the following observationsare made for sample test cases containing connective tissues

(i) For the identification of cancer from biopsy imagesof connective tissues in the case of 119870NN the averagevalue of accuracy specificity sensitivity BCR 119865-measure and MCC is 0921909 0940164 08199220880263 0759395 and 0717455 respectively

(ii) For the identification of cancer from biopsy of con-nective tissues in the case of SVM the average valueof accuracy specificity sensitivity BCR 119865-measureand MCC is 089245 0888438 0948297 09187560538314 and 055879 respectively

(iii) For the identification of cancer from biopsy of con-nective tissues in the case of fuzzy 119870NN the averagevalue of accuracy specificity sensitivity BCR 119865-measure and MCC is 0787879 0867476 03700740618789 0356613 and 0231013 respectively

(iv) For the identification of cancer from biopsy of con-nective tissues in the case of random forest classifierthe average value of accuracy specificity sensitivityBCR 119865-measure and MCC is 0907245 09936680493996 0743832 0647373 and 0642137 respec-tively

From Table 5 and Figure 5(b) the following observationsare made for sample test cases containing epithelial tissues

(i) For the identification of cancer from biopsy images ofepithelial tissues in the case of119870NN the average valueof accuracy specificity sensitivity BCR 119865-measureand MCC is 0884727 0916446 0801733 08594350795319 and 071626 respectively

(ii) For the identification of cancer from biopsy of epithe-lial tissues in the case of SVM the average valueof accuracy specificity sensitivity BCR 119865-measureand MCC is 0796998 07851 0898525 08422790472804 and 04587 respectively


0

02

04

06

08

1

12

RFSVM

Fuzzy KNNKNN

Connective tissues

BCR

F-m

easu

re

MCC

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

(a)

RFSVM

Fuzzy KNNKNN

0

02

04

06

08

1

12Epithelial tissue

BCR

F-m

easu

re

MCC

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

(b)

RFSVM

Fuzzy KNNKNN

0

02

04

06

08

1

12Muscular tissues

BCR

F-m

easu

re

MCC

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

(c)

RFSVM

Fuzzy KNNKNN

0

02

04

06

08

1Nervous tissues

BCR

F-m

easu

re

MCC

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

(d)

Figure 5 Performance analysis of classifiers with four fundamental tissues connective tissue as (a) epithelial tissue as (b) muscular tissueas (c) and nervous tissue as (d)

(iii) For the identification of cancer from biopsy of epithe-lial tissues in the case of fuzzy119870NN the average valueof accuracy specificity sensitivity BCR 119865-measureand MCC is 0665834 076465 0407057 05859840401181 and 017053 respectively

(iv) For the identification of cancer from biopsy of epithe-lial tissues in the case of random forest classifierthe average value of accuracy specificity sensitivity

BCR 119865-measure and MCC is 0849306 09662430555332 0760788 0675868 and 0609494 respec-tively

From Table 5 and Figure 5(c) the following observationsare made for sample test cases containing muscular tissues

(i) For the identification of cancer from biopsy images ofmuscular tissues in the case of119870NN the average valueof accuracy specificity sensitivity BCR 119865-measure


Table 6 The comparison of the proposed method with other standard methods

Authors (year) Feature set used Methods of classification Parameters used () Dataset used

Huang and Lai(2010) [15] Texture features Support vector machine

(SVM) Accuracy = 9281000 times 1000 4000 times

3000 and 275 times 275HCC biopsy images

Di Cataldo et al(2010) [45]

Texture andmorphology

Support vector machine(SVM) Accuracy = 9177 Digitized histology lung

cancer IHC tissue imagesHe et al (2008)[46]

Shape morphologyand texture

Artificial neural network(ANN) and SVM Accuracy = 9000 Digitized histology

imagesMookiah et al(2011) [47]


Error backpropagationneural network (BPNN)

Accuracy = 9643 sensitivity= 9231 and specificity = 82

83 normal and 29 OSFimages

Krishnan et al(2011) [48] HOG LBP and LTE LDA Accuracy = 82 Normal-83

OSFWD-29

Krishnan et al(2011) [48] HOG LBP and LTE Support vector machine

(SVM) Accuracy = 8838

Histology imagesNormal-90OSFWD-42OSFD-26

Caicedo et al(2009) [8] Bag of features Support vector machine

(SVM)Sensitivity = 92Specificity = 88 2828 histology images

Sinha andRamkrishan(2003) [17]

Texture and statisticalfeatures 119870NN Accuracy = 706 Blood cells histology

images

The proposedapproach

Texture shape andmorphology HOGwavelet colorTamurarsquos featureand LTE

KNN

Average accuracy = 9219sensitivity = 9401specificity = 8199 BCR =8802 F-measure = 7594MCC = 7174

2828 histology images

and MCC is 0897321 0923277 0650761 07870920543009 and 049783 respectively

(ii) For the identification of cancer from biopsy of mus-cular tissues in the case of SVM the average valueof accuracy specificity sensitivity BCR 119865-measureand MCC is 0884379 0886718 0786303 0836810263764 and 0320547 respectively

(iii) For the identification of cancer frombiopsy ofmuscu-lar tissues in the case of fuzzy 119870NN the average valueof accuracy specificity sensitivity BCR 119865-measureand MCC is 0614958 0672503 0535894 06043640538571 and 0208941 respectively

(iv) For the identification of cancer from biopsy of mus-cular tissues in the case of random forest classifierthe accuracy specificity sensitivity BCR 119865-measureand MCC are 0889878 0995023 0193145 05940840313309 and 037318 respectively

From Table 5 and Figure 5(d) the following observationsare made for sample test cases containing nervous tissues

(i) For the identification of cancer from biopsy images ofnervous tissues in the case of 119870NN the average valueof accuracy specificity sensitivity BCR 119865-measureand MCC is 0861763 0880866 0835733 08584820834116 and 0716492 respectively

(ii) For the identification of cancer from biopsy of ner-vous tissues in the case of SVM the average value

of accuracy specificity sensitivity BCR 119865-measureand MCC is 0769545 0723056 0946068 08349230630126 and 0552038 respectively

(iii) For the identification of cancer from biopsy of ner-vous tissues in the case of fuzzy 119870NN the accuracyspecificity sensitivity BCR 119865-measure and MCCare 0808453 0882722 0242776 0562835 0225886and 011837 respectively

(iv) For the identification of cancer from biopsy of ner-vous tissues in the case of random forest classifier theaverage value of accuracy specificity sensitivity BCR119865-measure and MCC is 0843102 092827 07232620825766 0792403 and 0676888 respectively

From the above discussions for all four categories of testcases it is observed that the 119870NN is performing better incomparison to other classifiers for the identification of cancerfrom biopsy images of nervous tissues

From all above observations it is concluded that the119870NN classifier is producing better results in comparison toother methods for the case of biopsy images of connectivetissues The maximum values of the accuracy sensitivity andspecificity are 09552 09615 and 09543 respectively The 119896-nearest neighbor classifier is also performing better for allcases as well as that was discussed above Table 6 gives acomparative analysis of the proposed framework with otherstandard methods available in the literature From Table 6it can be observed that the proposed method is performingbetter in comparison to all other methods


5 Conclusions

An automated detection and classification procedure waspresented for detection of cancer from microscopic biopsyimages using clinically significant and biologically inter-pretable set of features The proposed analysis was basedon tissues level microscopic observations of cell and nucleifor cancer detection and classification For enhancement ofmicroscopic biopsy images contrast limited adaptive his-togram equalization based method was used For segmen-tation of images 119896-means clustering method was used Aftersegmentation of images a total of 115 hybrid sets of featureswere extracted for 2828 sample histology images taken fromhistology database [8] After feature extraction 1000 sampleswere selected randomly for classification purposes Out of1000 samples of 115 features 900 samples were selected fortraining purposes and 100 samples were selected for testingpurposes The various classification approaches tested were119870-nearest neighborhood (119870NN) fuzzy119870NN support vectormachine (SVM) and random forest based classifiers FromTable 5 we are in position to conclude that 119870NN is the bestsuited classification algorithm for detection of noncancerousand cancerous microscopic biopsy images containing all fourfundamental tissues SVM provides average results for allthe tissues types but not better than 119870NN Fuzzy 119870NN iscomparatively a less good classifier RF classifier provides verylow sensitivity and down accuracy rate as compared to 119870NNclassifier for this dataset Hence from experimental results itwas observed that 119870NN classifier is performing better for allcategories of test cases present in the selected test data Thesecategories of test data are connective tissues epithelial tissuesmuscular tissues andnervous tissues Among all categories oftest cases further it was observed that the proposed methodis performing better for connective tissues type sampletest cases The performance measures for connective tissuesdataset in terms of the average accuracy specificity sensi-tivity BCR 119865-measure and MCC are 0921909 09401640819922 0880263 0759395 and 0717455 respectively

Conflict of Interests

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

References

[1] I AliWAWani andK Saleem ldquoCancer scenario in Indiawithfuture perspectivesrdquo Cancer Therapy vol 8 pp 56ndash70 2011

[2] A Tabesh M Teverovskiy H-Y Pang et al ldquoMultifeatureprostate cancer diagnosis and gleason grading of histologicalimagesrdquo IEEE Transactions on Medical Imaging vol 26 no 10pp 1366ndash1378 2007

[3] A Madabhushi ldquoDigital pathology image analysis opportuni-ties and challengesrdquo Imaging in Medicine vol 1 no 1 pp 7ndash102009

[4] A N Esgiar R N G Naguib B S Sharif M K Bennettand A Murray ldquoFractal analysis in the detection of coloniccancer imagesrdquo IEEE Transactions on Information Technology inBiomedicine vol 6 no 1 pp 54ndash58 2002

[5] L Yang O Tuzel P Meer and D J Foran ldquoAutomatic imageanalysis of histopathology specimens using concave vertexgraphrdquo in Medical Image Computing and Computer-AssistedInterventionmdashMICCAI 2008 pp 833ndash841 Springer BerlinGermany 2008

[6] R C Gonzalez Digital Image Processing Pearson EducationIndia 2009

[7] S Liao M W K Law and A C S Chung ldquoDominant localbinary patterns for texture classificationrdquo IEEE Transactions onImage Processing vol 18 no 5 pp 1107ndash1118 2009

[8] J C Caicedo A Cruz and F A Gonzalez ldquoHistopathologyimage classification using bag of features and kernel functionsrdquoinArtificial Intelligence in Medicine vol 5651 of Lecture Notes inComputer Science pp 126ndash135 Springer Berlin Germany 2009

[9] R Kumar and R Srivastava ldquoSome observations on the per-formance of segmentation algorithms for microscopic biopsyimagesrdquo in Proceedings of the International Conference onModeling and Simulation of Diffusive Processes and Applica-tions (ICMSDPA rsquo14) pp 16ndash22 Department of MathematicsBanaras Hindu University Varanasi India October 2014

[10] C Demir and B Yener ldquoAutomated cancer diagnosis basedon histopathological images a systematic surveyrdquo Tech RepRensselaer Polytechnic Institute New York NY USA 2005

[11] S Bhattacharjee J Mukherjee S Nag I K Maitra and SK Bandyopadhyay ldquoReview on histopathological slide analysisusing digital microscopyrdquo International Journal of AdvancedScience and Technology vol 62 pp 65ndash96 2014

[12] C Bergmeir M G Silvente and J M Benıtez ldquoSegmentationof cervical cell nuclei in high-resolution microscopic imagesa new algorithm and a web-based software frameworkrdquo Com-puter Methods and Programs in Biomedicine vol 107 no 3 pp497ndash512 2012

[13] A Mouelhi M Sayadi F Fnaiech K Mrad and K BRomdhane ldquoAutomatic image segmentation of nuclear stainedbreast tissue sections using color active contour model and animproved watershed methodrdquo Biomedical Signal Processing andControl vol 8 no 5 pp 421ndash436 2013

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

[15] P-W Huang and Y-H Lai ldquoEffective segmentation and classifi-cation for HCC biopsy imagesrdquo Pattern Recognition vol 43 no4 pp 1550ndash1563 2010

[16] G Landini D A Randell T P Breckon and J W Han ldquoMor-phologic characterization of cell neighborhoods in neoplasticand preneoplastic epitheliumrdquo Analytical and QuantitativeCytology and Histology vol 32 no 1 pp 30ndash38 2010

[17] N Sinha and A G Ramkrishan ldquoAutomation of differentialblood countrdquo in Proceedings of the Conference on ConvergentTechnologies for Asia-Pacific Region (TINCON rsquo03) pp 547ndash551Bangalore India 2003

[18] F Kasmin A S Prabuwono and A Abdullah ldquoDetectionof leukemia in human blood sample based on microscopicimages a studyrdquo Journal of Theoretical amp Applied InformationTechnology vol 46 no 2 2012

[19] R Srivastava J R P Gupta and H Parthasarathy ldquoEnhance-ment and restoration of microscopic images corrupted withpoissonrsquos noise using a nonlinear partial differential equation-based filterrdquo Defence Science Journal vol 61 no 5 pp 452ndash4612011

[20] E D Pisano S Zong BMHemminger et al ldquoContrast limitedadaptive histogram equalization image processing to improve


the detection of simulated spiculations in densemammogramsrdquoJournal of Digital Imaging vol 11 no 4 pp 193ndash200 1998

[21] P Perona and J Malik ldquoScale-space and edge detection usinganisotropic diffusionrdquo IEEE Transactions on Pattern Analysisand Machine Intelligence vol 12 no 7 pp 629ndash639 1990

[22] Y Al-Kofahi W Lassoued W Lee and B Roysam ldquoImprovedautomatic detection and segmentation of cell nuclei inhistopathology imagesrdquo IEEE Transactions on Biomedical Engi-neering vol 57 no 4 pp 841ndash852 2010

[23] D L PhamC Xu and J L Prince ldquoCurrentmethods inmedicalimage segmentationrdquoAnnual Review of Biomedical Engineeringvol 2 no 1 pp 315ndash337 2000

[24] R Eid G Landini and O P Unit ldquoOral epithelial dysplasiacan quantifiable morphological features help in the gradingdilemmardquo in Proceedings of the 1st ImageJ User and DeveloperConference Luxembourg City Luxembourg 2006

[25] N Bonnet ldquoSome trends in microscope image processingrdquoMicron vol 35 no 8 pp 635ndash653 2004

[26] M M R Krishnan C Chakraborty R R Paul and A KRay ldquoHybrid segmentation characterization and classificationof basal cell nuclei from histopathological images of normaloral mucosa and oral submucous fibrosisrdquo Expert Systems withApplications vol 39 no 1 pp 1062ndash1077 2012

[27] H P Ng S H Ong K W C Foong P S Goh and WL Nowinski ldquoMedical image segmentation using k-meansclustering and improved watershed algorithmrdquo in Proceedingsof the 7th IEEE Southwest Symposium on Image Analysis andInterpretation pp 61ndash65 IEEE March 2006

[28] K-S Chuang H-L Tzeng S Chen J Wu and T-J ChenldquoFuzzy c-means clustering with spatial information for imagesegmentationrdquo Computerized Medical Imaging and Graphicsvol 30 no 1 pp 9ndash15 2006

[29] N R Pal and S K Pal ldquoA review on image segmentationtechniquesrdquo Pattern Recognition vol 26 no 9 pp 1277ndash12941993

[30] M-N Wu C-C Lin and C-C Chang ldquoBrain tumor detec-tion using color-based K-means clustering segmentationrdquo inProceedings of the 3rd International Conference on IntelligentInformation Hiding and Multimedia Signal Processing (IIHMSPrsquo07) pp 245ndash248 IEEE November 2007

[31] S Srivastava N Sharma S K Singh and R Srivastava ldquoAcombined approach for the enhancement and segmentationof mammograms using modified fuzzy C-means method inwavelet domainrdquo Journal of Medical Physics vol 39 no 3 pp169ndash183 2014

[32] J Kong O Sertel H Shimada K L Boyer J H Saltz and MN Gurcan ldquoComputer-aided evaluation of neuroblastoma onwhole-slide histology images classifying grade of neuroblasticdifferentiationrdquo Pattern Recognition vol 42 no 6 pp 1080ndash1092 2009

[33] C G Loukas and A Linney ldquoA survey on histological imageanalysis-based assessment of three major biological factorsinfluencing radiotherapy proliferation hypoxia and vascula-turerdquo Computer Methods and Programs in Biomedicine vol 74no 3 pp 183ndash199 2004

[34] N Orlov L Shamir T Macura J Johnston D M Eckley andI G Goldberg ldquoWND-CHARM multi-purpose image classifi-cation using compound image transformsrdquo Pattern RecognitionLetters vol 29 no 11 pp 1684ndash1693 2008

[35] J Diamond N H Anderson P H Bartels R Montironi andP W Hamilton ldquoThe use of morphological characteristics and

texture analysis in the identification of tissue composition inprostatic neoplasiardquo Human Pathology vol 35 no 9 pp 1121ndash1131 2004

[36] S Doyle M Hwang K Shah AMadabhushi M Feldman andJ Tomaszeweski ldquoAutomated grading of prostate cancer usingarchitectural and textural image featuresrdquo in Proceedings of the4th IEEE International Symposium on Biomedical Imaging FromNano to Macro (ISBI rsquo07) pp 1284ndash1287 April 2007

[37] R O Duda and P E Hart Pattern Classification and SceneAnalysis vol 3 Wiley New York NY USA 1973

[38] A K Jain Fundamentals of Digital Image Processing vol 3Prentice-Hall Englewood Cliffs NJ USA 1989

[39] M M R Krishnan V Venkatraghavan U R Acharya et alldquoAutomated oral cancer identification using histopathologicalimages a hybrid feature extraction paradigmrdquo Micron vol 43no 2-3 pp 352ndash364 2012

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

[41] L Wei Y Yang and R M Nishikawa ldquoMicrocalcificationclassification assisted by content-based image retrieval forbreast cancer diagnosisrdquo Pattern Recognition vol 42 no 6 pp1126ndash1132 2009

[42] G Lalli D Kalamani and N Manikandaprabu ldquoA perspectivepattern recognition using retinal nerve fibers with hybridfeature setrdquo Life Science Journal vol 10 no 2 pp 2725ndash27302013

[43] Y Yang L Wei and R M Nishikawa ldquoMicrocalcification clas-sification assisted by content-based image retrieval for breastcancer diagnosisrdquo in Proceedings of the 14th IEEE InternationalConference on Image Processing (ICIP rsquo07) vol 5 pp 1ndash4September 2007

[44] L Hadjiiski P Filev H-P Chan et al ldquoComputerized detectionand classification of malignant and benign microcalcificationson full field digital mammogramsrdquo in Digital Mammography9th International Workshop IWDM 2008 Tucson AZ USAJuly 20ndash23 2008 Proceedings E A Krupinski Ed vol 5116of Lecture Notes in Computer Science pp 336ndash342 SpringerBerlin Germany 2008

[45] S Di Cataldo E Ficarra A Acquaviva and E Macii ldquoAuto-mated segmentation of tissue images for computerized IHCanalysisrdquo Computer Methods and Programs in Biomedicine vol100 no 1 pp 1ndash15 2010

[46] L He Z Peng B Everding et al ldquoA comparative study ofdeformable contour methods on medical image segmentationrdquoImage and Vision Computing vol 26 no 2 pp 141ndash163 2008

[47] M R Mookiah P Shah C Chakraborty and A K RayldquoBrownian motion curve-based textural classification and itsapplication in cancer diagnosisrdquo Analytical and QuantitativeCytology and Histology vol 33 no 3 pp 158ndash168 2011

[48] M M R Krishnan C Chakraborty R R Paul and A KRay ldquoQuantitative analysis of sub-epithelial connective tissuecell population of oral submucous fibrosis using support vectormachinerdquo Journal of Medical Imaging and Health Informaticsvol 1 no 1 pp 4ndash12 2011

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014


Active and Passive Electronic Components

Control Scienceand Engineering

Journal of



RotatingMachinery


Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design



Shock and Vibration


Civil EngineeringAdvances in

Acoustics and VibrationAdvances in



Electrical and Computer Engineering

Journal of

Advances inOptoElectronics


Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of


Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014


Chemical EngineeringInternational Journal of Antennas and

Propagation




Navigation and Observation



DistributedSensor Networks



microscopic tissue structure of the patient Thus biopsyimage analysis is a vital technique for cancer detection [2 3]Histopathology is the study of symptoms and indications ofthe disease using the microscopic biopsy images To visualizevarious parts of the tissue under a microscope the sectionsare dyed with one or more staining components The maingoal of staining is to reveal the components at cellular leveland counterstains are used to provide color visibility andcontrast Hematoxylin-Eosin (HampE) is staining componentthat has been used by pathologists for over few decadesHematoxylin stains cell nuclei which are blue in color whileEosin stains cytoplasm and connective tissues which are ofpink color The histology [4] is related to the study of cellsin terms of structure function and interpretations of thetissue and cells Microscopic biopsies are most commonlyused for both disease screenings because of the less invasivenatures The characteristic of microscopic biopsy images haspresence of isolated cells and cell clusters The microscopicbiopsy images are easier to analyze specimens compared tohistopathology due to absence of noncomplicated structures[5] The accurate manual identification of cancer frommicroscopic biopsy images has always been a major issue bythe pathologists and medical practitioners observing cell ortissue structure under the microscope

In histopathology the cancer detection process normallyconsists of categorizing the image biopsy into cancerousone or noncancerous one [6] In microscopic biopsy imageanalysis doctors and pathologists observe many of theabnormalities and categorize the sample based on variouscharacteristics of the cell nuclei such as color shape size andproportion to cytoplasmHigh resolutionmicroscopic biopsyprovides reliable information for differentiating abnormaland normal tissues The difference between normal andcancerous cells is shown in Table 1 [7]

For the detection and diagnosis of cancer from micro-scopic biopsy images the histopathologists normally lookat the specific features in the cells and tissue structuresThe various common features used for the detection anddiagnosis of cancer from the microscopic biopsy imagesinclude shape and size of cells shape and size of cell nucleiand distribution of the cells The brief descriptions of thesefeatures are given as follows

(A) Shape and Size of the Cells It has been observed that theoverall shape and size of cells in the tissues aremostly normalThe cellular structures of the cancerous cells might be eitherlarger or shorter thannormal cellsThenormal cells have evenshapes and functionality Cancer cells usually do not functionin a useful way and their shapes are often not even

(B) Size and Shape of the Cellrsquos Nucleus The shape and sizeof the nucleus of a cancer cell are often not normal Thenucleus is decentralized in the cancer cells The image ofthe cell looks like an omelet in which the central yolk isthe nucleus and the surrounding white is the cytoplasmThe nuclei of cancer cells are larger than the normal cellsand deviated from the centre of the mass The nucleusof cancer cell is darker The segmentation step mainlyfocuses on separation of regions of interests (cells) from

background tissues as well as separation of nuclei from cyto-plasm

(C) Distribution of the Cells in Tissue The function of eachtissue depends on the distribution and arrangements of thenormal cells The numbers of healthy cells per unit area areless in the cancerous tissues These adjectives of microscopicbiopsy images have been included in shape and morphologybased features texture features color based features ColorGray Level Cooccurrence Matrix (GLCM) Lawrsquos TextureEnergy (LTE) Tamurarsquos features and wavelet features whichare more biologically interpretable and clinically significant

The main aim of this paper is to design and develop aframework and a software tool for automated detection andclassification of cancer frommicroscopic biopsy images usingthe abovementioned clinically significant and biologicallyinterpretable features This paper focuses on selecting anappropriate method for each design stage of the frameworkafter making a comparative analysis of the various commonlyusedmethods in each categoryThe various stages involved inthe proposed methodology include enhancement of micro-scopic images segmentation of background cells featuresextraction and finally the classification

The rest of the paper has been structured as followsSection 2 describes the related works Section 3 presents themethods and models Section 4 describes the results anddiscussions and finally Section 5 draws the conclusion of thework presented in this paper

2 Related Works

In recent years few works have been reported in the lit-erature for the design and development of tools for auto-mated cancer detection from microscopic biopsy imagesKumar and Srivastava [9] presented detailed reviews onthe computer aided diagnosis (CAD) for cancer detectionfrom microscopic biopsy images Demir and Yener [10] alsopresented a method for automatic diagnosis of biopsy imageThey presented a cellular level diagnosis system using imageprocessing techniques Bhattacharjee et al [11] presented areview on computer aided diagnosis system to detect cancerfrom microscopic biopsy images using image processingtechniques

Bergmeir et al [12] proposed a model to extract thetexture features by using local histograms and GLCM Thequasisupervised learning algorithm operates on two datasetsthe first one having normal tissues labeled only indirectlyand the second one containing an unlabeled collection ofmixed samples of both normal and cancer tissues Thismethod was applied on the dataset of 22080 vectors withreduced dimensionality 119 from 132 The regions havingthe cancerous tissues were accurately identified having truepositive rate 88 and false positive rate 19 respectively byusing manual ground truth dataset

Mouelhi et al [13] used Haralickrsquos textures features [14]histogram of oriented gradients (HOG) and color compo-nent based statistical moments (CCSM) features selectionand extraction approaches to classify the cancerous cells frommicroscopic biopsy images The various features used in this








morphology)













Noncancerous





Classification

Cancerous








(a) (b)





PRI (119878test 119866119896)

=1

(1198732)sum


(1)




VOI (119883 119884) = 119867 (119883) = 119867 (119884) minus 2119868 (119883 119884) (2)




119864 (119904119894 119892119895 119901119896) =

10038161003816100381610038161003816119877 (119904119894 119901119896) 119877 (119892119895 119901119896)

100381610038161003816100381610038161003816100381610038161003816119877 (119904119894 119901119896)

1003816100381610038161003816

(3)

GCE (119878 119866) =1

119899minsum

119894

119864 (119878 119866 119901119894) sum

119894

119864 (119878 119866 119901119894) (4)








(a)

(A) (B)

(b)

(c)

(d)



0

02

04

06

08

1

12


FCMTexture based

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

FPR

PRI

GCE VO

I






119860 =

119899

sum

119894=1

119898

sum

119895=1

119861 (119894 119895) (5)




NLD = radic(1199091 minus 1199092)2

+ (1199101 minus 1199102)2 (6)



NSD = radic(1199092 minus 1199091)2

+ (1199102 minus 1199101)2 (7)




Perimeter (8)




120574 =119860

119875=

4120587 times Area1198752

(10)


Solidity =Area

Convex Area (11)



(12)


Compactness =Area

Perimeter2 (13)

































Accuracy =TP + TN

119873times 100 (14)



Sensitivity =TP

TP + FN (15)



Specificity =TN

TN + FP (16)




measure MCC



RF 0889878 0995023 0193145 0594084 0313309 037318 0843102 092827 0723262 0825766 0792403 0676888SVM 0884379 0886718 0786303 083681 0263764 0320547 0769545 0723056 0946068 0834923 0630126 0552038FUZZY119870NN 0614958 0672503 0535894 0604364 0538571 0208941 0808453 0882722 0242776 0562835 0225886 011837

119870NN 0897321 0923277 0650761 0787092 0543009 049783 0861763 0880866 0835733 0858482 0834116 0716492





Precision =TP

TP + FP

Recall =TP

TP + FN


(18)



MCC















0

02

04

06

08

1

12

RFSVM

Fuzzy KNNKNN

Connective tissues

BCR

F-m

easu

re

MCC

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

(a)

RFSVM

Fuzzy KNNKNN

0

02

04

06

08

1

12Epithelial tissue

BCR

F-m

easu

re

MCC

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

(b)

RFSVM

Fuzzy KNNKNN

0

02

04

06

08

1

12Muscular tissues

BCR

F-m

easu

re

MCC

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

(c)

RFSVM

Fuzzy KNNKNN

0

02

04

06

08

1Nervous tissues

BCR

F-m

easu

re

MCC

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

(d)

























OSFWD-29








images



KNN
















5 Conclusions




References






















































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation
















morphology)













Noncancerous





Classification

Cancerous








(a) (b)





PRI (119878test 119866119896)

=1

(1198732)sum


(1)




VOI (119883 119884) = 119867 (119883) = 119867 (119884) minus 2119868 (119883 119884) (2)




119864 (119904119894 119892119895 119901119896) =

10038161003816100381610038161003816119877 (119904119894 119901119896) 119877 (119892119895 119901119896)

100381610038161003816100381610038161003816100381610038161003816119877 (119904119894 119901119896)

1003816100381610038161003816

(3)

GCE (119878 119866) =1

119899minsum

119894

119864 (119878 119866 119901119894) sum

119894

119864 (119878 119866 119901119894) (4)








(a)

(A) (B)

(b)

(c)

(d)



0

02

04

06

08

1

12


FCMTexture based

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

FPR

PRI

GCE VO

I






119860 =

119899

sum

119894=1

119898

sum

119895=1

119861 (119894 119895) (5)




NLD = radic(1199091 minus 1199092)2

+ (1199101 minus 1199102)2 (6)



NSD = radic(1199092 minus 1199091)2

+ (1199102 minus 1199101)2 (7)




Perimeter (8)




120574 =119860

119875=

4120587 times Area1198752

(10)


Solidity =Area

Convex Area (11)



(12)


Compactness =Area

Perimeter2 (13)

































Accuracy =TP + TN

119873times 100 (14)



Sensitivity =TP

TP + FN (15)



Specificity =TN

TN + FP (16)




measure MCC



RF 0889878 0995023 0193145 0594084 0313309 037318 0843102 092827 0723262 0825766 0792403 0676888SVM 0884379 0886718 0786303 083681 0263764 0320547 0769545 0723056 0946068 0834923 0630126 0552038FUZZY119870NN 0614958 0672503 0535894 0604364 0538571 0208941 0808453 0882722 0242776 0562835 0225886 011837

119870NN 0897321 0923277 0650761 0787092 0543009 049783 0861763 0880866 0835733 0858482 0834116 0716492





Precision =TP

TP + FP

Recall =TP

TP + FN


(18)



MCC















0

02

04

06

08

1

12

RFSVM

Fuzzy KNNKNN

Connective tissues

BCR

F-m

easu

re

MCC

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

(a)

RFSVM

Fuzzy KNNKNN

0

02

04

06

08

1

12Epithelial tissue

BCR

F-m

easu

re

MCC

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

(b)

RFSVM

Fuzzy KNNKNN

0

02

04

06

08

1

12Muscular tissues

BCR

F-m

easu

re

MCC

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

(c)

RFSVM

Fuzzy KNNKNN

0

02

04

06

08

1Nervous tissues

BCR

F-m

easu

re

MCC

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

(d)

























OSFWD-29








images



KNN
















5 Conclusions




References






















































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation















Noncancerous





Classification

Cancerous








(a) (b)





PRI (119878test 119866119896)

=1

(1198732)sum


(1)




VOI (119883 119884) = 119867 (119883) = 119867 (119884) minus 2119868 (119883 119884) (2)




119864 (119904119894 119892119895 119901119896) =

10038161003816100381610038161003816119877 (119904119894 119901119896) 119877 (119892119895 119901119896)

100381610038161003816100381610038161003816100381610038161003816119877 (119904119894 119901119896)

1003816100381610038161003816

(3)

GCE (119878 119866) =1

119899minsum

119894

119864 (119878 119866 119901119894) sum

119894

119864 (119878 119866 119901119894) (4)








(a)

(A) (B)

(b)

(c)

(d)



0

02

04

06

08

1

12


FCMTexture based

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

FPR

PRI

GCE VO

I






119860 =

119899

sum

119894=1

119898

sum

119895=1

119861 (119894 119895) (5)




NLD = radic(1199091 minus 1199092)2

+ (1199101 minus 1199102)2 (6)



NSD = radic(1199092 minus 1199091)2

+ (1199102 minus 1199101)2 (7)




Perimeter (8)




120574 =119860

119875=

4120587 times Area1198752

(10)


Solidity =Area

Convex Area (11)



(12)


Compactness =Area

Perimeter2 (13)

































Accuracy =TP + TN

119873times 100 (14)



Sensitivity =TP

TP + FN (15)



Specificity =TN

TN + FP (16)




measure MCC



RF 0889878 0995023 0193145 0594084 0313309 037318 0843102 092827 0723262 0825766 0792403 0676888SVM 0884379 0886718 0786303 083681 0263764 0320547 0769545 0723056 0946068 0834923 0630126 0552038FUZZY119870NN 0614958 0672503 0535894 0604364 0538571 0208941 0808453 0882722 0242776 0562835 0225886 011837

119870NN 0897321 0923277 0650761 0787092 0543009 049783 0861763 0880866 0835733 0858482 0834116 0716492





Precision =TP

TP + FP

Recall =TP

TP + FN


(18)



MCC















0

02

04

06

08

1

12

RFSVM

Fuzzy KNNKNN

Connective tissues

BCR

F-m

easu

re

MCC

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

(a)

RFSVM

Fuzzy KNNKNN

0

02

04

06

08

1

12Epithelial tissue

BCR

F-m

easu

re

MCC

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

(b)

RFSVM

Fuzzy KNNKNN

0

02

04

06

08

1

12Muscular tissues

BCR

F-m

easu

re

MCC

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

(c)

RFSVM

Fuzzy KNNKNN

0

02

04

06

08

1Nervous tissues

BCR

F-m

easu

re

MCC

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

(d)

























OSFWD-29








images



KNN
















5 Conclusions




References






















































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










(a) (b)





PRI (119878test 119866119896)

=1

(1198732)sum


(1)




VOI (119883 119884) = 119867 (119883) = 119867 (119884) minus 2119868 (119883 119884) (2)




119864 (119904119894 119892119895 119901119896) =

10038161003816100381610038161003816119877 (119904119894 119901119896) 119877 (119892119895 119901119896)

100381610038161003816100381610038161003816100381610038161003816119877 (119904119894 119901119896)

1003816100381610038161003816

(3)

GCE (119878 119866) =1

119899minsum

119894

119864 (119878 119866 119901119894) sum

119894

119864 (119878 119866 119901119894) (4)








(a)

(A) (B)

(b)

(c)

(d)



0

02

04

06

08

1

12


FCMTexture based

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

FPR

PRI

GCE VO

I






119860 =

119899

sum

119894=1

119898

sum

119895=1

119861 (119894 119895) (5)




NLD = radic(1199091 minus 1199092)2

+ (1199101 minus 1199102)2 (6)



NSD = radic(1199092 minus 1199091)2

+ (1199102 minus 1199101)2 (7)




Perimeter (8)




120574 =119860

119875=

4120587 times Area1198752

(10)


Solidity =Area

Convex Area (11)



(12)


Compactness =Area

Perimeter2 (13)

































Accuracy =TP + TN

119873times 100 (14)



Sensitivity =TP

TP + FN (15)



Specificity =TN

TN + FP (16)




measure MCC



RF 0889878 0995023 0193145 0594084 0313309 037318 0843102 092827 0723262 0825766 0792403 0676888SVM 0884379 0886718 0786303 083681 0263764 0320547 0769545 0723056 0946068 0834923 0630126 0552038FUZZY119870NN 0614958 0672503 0535894 0604364 0538571 0208941 0808453 0882722 0242776 0562835 0225886 011837

119870NN 0897321 0923277 0650761 0787092 0543009 049783 0861763 0880866 0835733 0858482 0834116 0716492





Precision =TP

TP + FP

Recall =TP

TP + FN


(18)



MCC















0

02

04

06

08

1

12

RFSVM

Fuzzy KNNKNN

Connective tissues

BCR

F-m

easu

re

MCC

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

(a)

RFSVM

Fuzzy KNNKNN

0

02

04

06

08

1

12Epithelial tissue

BCR

F-m

easu

re

MCC

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

(b)

RFSVM

Fuzzy KNNKNN

0

02

04

06

08

1

12Muscular tissues

BCR

F-m

easu

re

MCC

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

(c)

RFSVM

Fuzzy KNNKNN

0

02

04

06

08

1Nervous tissues

BCR

F-m

easu

re

MCC

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

(d)

























OSFWD-29








images



KNN
















5 Conclusions




References






















































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation















(a)

(A) (B)

(b)

(c)

(d)



0

02

04

06

08

1

12


FCMTexture based

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

FPR

PRI

GCE VO

I






119860 =

119899

sum

119894=1

119898

sum

119895=1

119861 (119894 119895) (5)




NLD = radic(1199091 minus 1199092)2

+ (1199101 minus 1199102)2 (6)



NSD = radic(1199092 minus 1199091)2

+ (1199102 minus 1199101)2 (7)




Perimeter (8)




120574 =119860

119875=

4120587 times Area1198752

(10)


Solidity =Area

Convex Area (11)



(12)


Compactness =Area

Perimeter2 (13)

































Accuracy =TP + TN

119873times 100 (14)



Sensitivity =TP

TP + FN (15)



Specificity =TN

TN + FP (16)




measure MCC



RF 0889878 0995023 0193145 0594084 0313309 037318 0843102 092827 0723262 0825766 0792403 0676888SVM 0884379 0886718 0786303 083681 0263764 0320547 0769545 0723056 0946068 0834923 0630126 0552038FUZZY119870NN 0614958 0672503 0535894 0604364 0538571 0208941 0808453 0882722 0242776 0562835 0225886 011837

119870NN 0897321 0923277 0650761 0787092 0543009 049783 0861763 0880866 0835733 0858482 0834116 0716492





Precision =TP

TP + FP

Recall =TP

TP + FN


(18)



MCC















0

02

04

06

08

1

12

RFSVM

Fuzzy KNNKNN

Connective tissues

BCR

F-m

easu

re

MCC

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

(a)

RFSVM

Fuzzy KNNKNN

0

02

04

06

08

1

12Epithelial tissue

BCR

F-m

easu

re

MCC

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

(b)

RFSVM

Fuzzy KNNKNN

0

02

04

06

08

1

12Muscular tissues

BCR

F-m

easu

re

MCC

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

(c)

RFSVM

Fuzzy KNNKNN

0

02

04

06

08

1Nervous tissues

BCR

F-m

easu

re

MCC

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

(d)

























OSFWD-29








images



KNN
















5 Conclusions




References






















































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










0

02

04

06

08

1

12


FCMTexture based

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

FPR

PRI

GCE VO

I






119860 =

119899

sum

119894=1

119898

sum

119895=1

119861 (119894 119895) (5)




NLD = radic(1199091 minus 1199092)2

+ (1199101 minus 1199102)2 (6)



NSD = radic(1199092 minus 1199091)2

+ (1199102 minus 1199101)2 (7)




Perimeter (8)




120574 =119860

119875=

4120587 times Area1198752

(10)


Solidity =Area

Convex Area (11)



(12)


Compactness =Area

Perimeter2 (13)

































Accuracy =TP + TN

119873times 100 (14)



Sensitivity =TP

TP + FN (15)



Specificity =TN

TN + FP (16)




measure MCC



RF 0889878 0995023 0193145 0594084 0313309 037318 0843102 092827 0723262 0825766 0792403 0676888SVM 0884379 0886718 0786303 083681 0263764 0320547 0769545 0723056 0946068 0834923 0630126 0552038FUZZY119870NN 0614958 0672503 0535894 0604364 0538571 0208941 0808453 0882722 0242776 0562835 0225886 011837

119870NN 0897321 0923277 0650761 0787092 0543009 049783 0861763 0880866 0835733 0858482 0834116 0716492





Precision =TP

TP + FP

Recall =TP

TP + FN


(18)



MCC















0

02

04

06

08

1

12

RFSVM

Fuzzy KNNKNN

Connective tissues

BCR

F-m

easu

re

MCC

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

(a)

RFSVM

Fuzzy KNNKNN

0

02

04

06

08

1

12Epithelial tissue

BCR

F-m

easu

re

MCC

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

(b)

RFSVM

Fuzzy KNNKNN

0

02

04

06

08

1

12Muscular tissues

BCR

F-m

easu

re

MCC

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

(c)

RFSVM

Fuzzy KNNKNN

0

02

04

06

08

1Nervous tissues

BCR

F-m

easu

re

MCC

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

(d)

























OSFWD-29








images



KNN
















5 Conclusions




References






















































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation









































Accuracy =TP + TN

119873times 100 (14)



Sensitivity =TP

TP + FN (15)



Specificity =TN

TN + FP (16)




measure MCC



RF 0889878 0995023 0193145 0594084 0313309 037318 0843102 092827 0723262 0825766 0792403 0676888SVM 0884379 0886718 0786303 083681 0263764 0320547 0769545 0723056 0946068 0834923 0630126 0552038FUZZY119870NN 0614958 0672503 0535894 0604364 0538571 0208941 0808453 0882722 0242776 0562835 0225886 011837

119870NN 0897321 0923277 0650761 0787092 0543009 049783 0861763 0880866 0835733 0858482 0834116 0716492





Precision =TP

TP + FP

Recall =TP

TP + FN


(18)



MCC















0

02

04

06

08

1

12

RFSVM

Fuzzy KNNKNN

Connective tissues

BCR

F-m

easu

re

MCC

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

(a)

RFSVM

Fuzzy KNNKNN

0

02

04

06

08

1

12Epithelial tissue

BCR

F-m

easu

re

MCC

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

(b)

RFSVM

Fuzzy KNNKNN

0

02

04

06

08

1

12Muscular tissues

BCR

F-m

easu

re

MCC

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

(c)

RFSVM

Fuzzy KNNKNN

0

02

04

06

08

1Nervous tissues

BCR

F-m

easu

re

MCC

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

(d)

























OSFWD-29








images



KNN
















5 Conclusions




References






















































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation
























Accuracy =TP + TN

119873times 100 (14)



Sensitivity =TP

TP + FN (15)



Specificity =TN

TN + FP (16)




measure MCC



RF 0889878 0995023 0193145 0594084 0313309 037318 0843102 092827 0723262 0825766 0792403 0676888SVM 0884379 0886718 0786303 083681 0263764 0320547 0769545 0723056 0946068 0834923 0630126 0552038FUZZY119870NN 0614958 0672503 0535894 0604364 0538571 0208941 0808453 0882722 0242776 0562835 0225886 011837

119870NN 0897321 0923277 0650761 0787092 0543009 049783 0861763 0880866 0835733 0858482 0834116 0716492





Precision =TP

TP + FP

Recall =TP

TP + FN


(18)



MCC















0

02

04

06

08

1

12

RFSVM

Fuzzy KNNKNN

Connective tissues

BCR

F-m

easu

re

MCC

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

(a)

RFSVM

Fuzzy KNNKNN

0

02

04

06

08

1

12Epithelial tissue

BCR

F-m

easu

re

MCC

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

(b)

RFSVM

Fuzzy KNNKNN

0

02

04

06

08

1

12Muscular tissues

BCR

F-m

easu

re

MCC

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

(c)

RFSVM

Fuzzy KNNKNN

0

02

04

06

08

1Nervous tissues

BCR

F-m

easu

re

MCC

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

(d)

























OSFWD-29








images



KNN
















5 Conclusions




References






















































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation












measure MCC



RF 0889878 0995023 0193145 0594084 0313309 037318 0843102 092827 0723262 0825766 0792403 0676888SVM 0884379 0886718 0786303 083681 0263764 0320547 0769545 0723056 0946068 0834923 0630126 0552038FUZZY119870NN 0614958 0672503 0535894 0604364 0538571 0208941 0808453 0882722 0242776 0562835 0225886 011837

119870NN 0897321 0923277 0650761 0787092 0543009 049783 0861763 0880866 0835733 0858482 0834116 0716492





Precision =TP

TP + FP

Recall =TP

TP + FN


(18)



MCC















0

02

04

06

08

1

12

RFSVM

Fuzzy KNNKNN

Connective tissues

BCR

F-m

easu

re

MCC

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

(a)

RFSVM

Fuzzy KNNKNN

0

02

04

06

08

1

12Epithelial tissue

BCR

F-m

easu

re

MCC

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

(b)

RFSVM

Fuzzy KNNKNN

0

02

04

06

08

1

12Muscular tissues

BCR

F-m

easu

re

MCC

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

(c)

RFSVM

Fuzzy KNNKNN

0

02

04

06

08

1Nervous tissues

BCR

F-m

easu

re

MCC

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

(d)

























OSFWD-29








images



KNN
















5 Conclusions




References






















































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










0

02

04

06

08

1

12

RFSVM

Fuzzy KNNKNN

Connective tissues

BCR

F-m

easu

re

MCC

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

(a)

RFSVM

Fuzzy KNNKNN

0

02

04

06

08

1

12Epithelial tissue

BCR

F-m

easu

re

MCC

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

(b)

RFSVM

Fuzzy KNNKNN

0

02

04

06

08

1

12Muscular tissues

BCR

F-m

easu

re

MCC

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

(c)

RFSVM

Fuzzy KNNKNN

0

02

04

06

08

1Nervous tissues

BCR

F-m

easu

re

MCC

Accu

racy

Sens

itivi

ty

Spec

ifici

ty

(d)

























OSFWD-29








images



KNN
















5 Conclusions




References






















































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation



























OSFWD-29








images



KNN
















5 Conclusions




References






















































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










5 Conclusions




References






















































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation











RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation









Date post:	11-Jun-2020
Category:	Documents
Upload:	others
View:	10 times
Download:	0 times

Research Article Detection and Classification of...

Documents