A Hybrid CNN-SVM Threshold Segmentation Approach for Tumor Detection and Classification of MRI Brain Images

Contents lists available atScienceDirect

IRBM

www.elsevier.com/locate/irbm

Original Article

A Hybrid CNN-SVM Threshold Segmentation Approach for Tumor Detection and Classification of MRI Brain Images

M.O. Khairandish

^a

, M. Sharma

^b

, V. Jain

^c

, J.M. Chatterjee

^d

, N.Z. Jhanjhi

^e,∗

aDepartmentofComputerScienceandEngineering,ChandigarhUniversity,Mohali,Punjab,India bSchoolofComputingScienceandEngineering,GalgotiasUniversity,GreaterNoida,U.P.,India cDepartmentofComputerScienceandEngineering,ShardaUniversity,GreaterNoida,U.P.,India dDepartmentofInformationTechnology,LordBuddhaEducationFoundation,Kathmandu,Nepal eSchoolofComputerScienceandEngineering(SCE),Taylor’sUniversity,Malaysia

h i g h l i g h t s g ra p h i c a l a b s t r a c t

•Theproposedhybridmodel,withconsid- erationofbothCNNand SVMmodelad- vantages,showssignificantimprovement.

•^{AHybridmethodonbrainMRIimagesto} detectandclassifythetumorhasbeenim- plemented.

•^{The overallaccuracyofthe hybridCNN-} SVMobtained98.4959%.

a rt i c l e i n f o a b s t r a c t

Articlehistory:

Received26April2020

Receivedinrevisedform1June2021 Accepted7June2021

Availableonlinexxxx Keywords:

MRI

Braintumoranalysis ML

DL CNN&SVM

Objective:Inthisresearchpaper,thebrainMRIimagesaregoingtoclassifybyconsideringtheexcellence ofCNNonapublicdatasettoclassifyBenignandMalignanttumors.

MaterialsandMethods:Deep learning (DL) methods due to good performance in the last few years have becomemore popular forImage classification. Convolution NeuralNetwork (CNN), withseveral methods,canextractfeatureswithoutusinghandcraftedmodels,andeventually,showbetteraccuracyof classification.TheproposedhybridmodelcombinedCNNandsupportvectormachine(SVM)intermsof classificationandwiththreshold-basedsegmentationintermsofdetection.

Result:The findings of previousstudies are based ondifferent models with their accuracy as Rough ExtremeLearningMachine(RELM)-94.233%, Deep CNN (DCNN)-95%, DeepNeural Network(DNN)and Discrete Wavelet Autoencoder (DWA)-96%, k-nearest neighbors (kNN)-96.6%, CNN-97.5%. The overall accuracyofthehybridCNN-SVMisobtainedas98.4959%.

Conclusion:Intoday’sworld,braincancerisoneofthemostdangerousdiseaseswiththehighestdeath rate,detectionandclassificationofbraintumorsduetoabnormalgrowthofcells,shapes,orientation,and thelocationisachallengeabletask inmedicalimaging.Magneticresonanceimaging(MRI)isatypical methodofmedical imaging for braintumoranalysis. Conventional machine learning(ML) techniques categorizebraincancer basedonsomehandicraftproperty withtheradiologistspecialist choice.That canleadtofailureintheexecutionandalsodecreasetheeffectivenessofanAlgorithm. Withabrieflook

*

Correspondingauthor.

E-mailaddress:noorzaman.jhanjhi@taylors.edu.my(N.Z. Jhanjhi).

https://doi.org/10.1016/j.irbm.2021.06.003

1959-0318/©^{2021AGBM.PublishedbyElsevierMassonSAS.Allrightsreserved.}

disease [2]. The braindisease is not the sameas anyother part of the body, but it can be caused by the anomalous growth of cells thatultimatelydestroysthebrainstructureandcausesbrain cancer[3].However,aspertheWorldHealthOrganization(WHO) report[4],itisestimatedthatabout9.6million aroundtheworld diedthatdiagnosedcancerin2018.Aswellasaround30%to50%

of those diagnosed withprimary cancer. Between manytypes of cancer,braincancerisa deadlyone. Inthismanner,thestatistics show[5],thataround17,760adultsdiedthatcausedbraintumors in2019.Duetothediresituationandabnormalgrowthofcancer and thecomplexity ofbrain structure, timely diagnosisis neces- sary.Magneticresonanceimaging(MRI)iswidelyusefulfortumor analysis,withhigh-qualitybrainimages.Inbrainimaging,theMRI technique moresignificant itprovides a unique wayto have the bestfitvisualizationofmaximumbothspatialandcontrastdeter- mination[6].

In theprocess ofan MRimage, that’s oftenusedto show the differentstateofdisease,asshowninFig.1,the(A)showssuspi- ciousinconsistency atthebottom rightregion.Meanwhile (B, C), showstheWorsestate andthetumor takesWiderspace. (D)un- like the previous stages,it showsthat thetumor is growing and destroyingtheneighboringcell.

Detection of brain tumors[7] is considered as an image seg- mentationproblemtolabelwhetherthereisatumorintheimage or not. To answer this problem, different ML algorithms and as wellasimageprocessingmethodsareimplementedbyresearchers on MRI images [8]. The main issue in image segmentation is to cluster similar feature vectors of an image. So, accurate extrac- tionoffeaturesisthekeyfactortosegmentan image[9].Among therecentstudiesbasedonimageprocessing, regainbasedmeth- ods, andthresholding approachesarecommon[10].Forexample, Sherlin and Murugan [11] applied a hybrid model of K-nearest neighbor (KNN) andthe FuzzyC-Mean method,andtry with2D medical images of the brain to define tumor cells and cluster them in the earliest stages from MRI images. In another work, Ashima etal.[12] usedcombined watershedandneural network (NN) segmentation and alsoself-organizing maps (SOM),to have segment braintumors. Abdalla M. et al. [41] proposed a swarm optimization model based on ant lion optimization (ALO), grey wolfoptimization(GWO),andartificialbeecolony(ABC),itapplied thesealgorithms on Computedtomography (CT)andMRI to seg- ment theliver Disease.The experimental resultshowedthatALO obtained 94.49%, respectively GWO obtained 94.08, and ABC ob- tained93.73%accuracy.Inasimilarwork,ShereenSaidetal.[42]

proposedasegmentationapproachbasedontheMoth-flameopti- mization(MFO)techniqueforthearrangementofabdominalcases of liver MRI images. In the same manner, it uses the Structural Similarity Index (SSI) to find the best fit output & obtained the experimental overall accuracy of95.66%.As per different studies, whiletheNNandthreshold-basedarecommonfortumorsegmen- tation, itisalsopossibleto havecombinedwithKNN toobtain a betteroutcome[13].Brain tumorscanbedetectedusingMLmeth- ods that required extensive processing. According to a different

Fig. 1.Growthstatusofbraintumorfromtheprimaryuptotheworsestate,(A) showstheinitialdetectionwithlessenhancement,(B,C)showstheacutestateof growthofatumor,and(D)showthattakingtodestroytheneighborcellsandtry spreadtoanotherregion[40].

factorofML,itmainlyneedsforhumanperceptionandfeedback.

Inthemeantime,extensivestudiesbasedonMLundertakentoac- celerate thisprocess. Prabhjot andAmardeep[14] haveused the fireflyalgorithmtoprovideanewwaytodetectthetumor,inthe Firefliesmethodsome chemical reaction,thatnamed biolumines- centis happening and produces light in their body which make probable to detect and confine tumor in an accurate way from brainMRIimages.InthestudythatVarunaShreeandKumar[15]

havedone,they useddiscretewavelettransformation(DWT)with theextractionofgraylevelco-occurrencematrix(GLCM)features thatarebasedonmorphologicprocesseswithprobabilisticNNfor classification of normal and abnormal tumor images. In another work,Bjoern H.Menze etal.[16] reportedthe resultofa multi- modalbraintumorsegmentationbenchmarkthatisgoingtoapply onlowandhighgrades of65multi-contractMRI patients.Mean- while,itusessimulationsoftwareandconvolutionhumanmethod toshowthedifficultyofsegmentation,andthescorerangeisbe- tween(74%- 85%).Andtheresultshowsthatalgorithmsbasedon regionsandreaching thatpoint,differentmethodscan takeplace intoprate.

DLisaNN-based architecture [17], inanotherviewpoint, itis oneoftheMLtechniquesthatmakesitpossibletoaddmanyhid- denlayersamonginputandoutputlayers[18].Itcanbeappliedin differentdomainssuch asspeechidentification[19],objectrecog- nition[20,21],imageclassification[22].TheDLarchitecturewidely canbe usedwithCNN[23],that easilywe canmanage theinput layers,hiddenlayers,andoutput layers[24]. However,thetypical

Fig. 2.CNN general architecture of image analysis.

CNNmethodcandefinewhetheranimagehasanobjectbutwith- outmentioningthelocation[25].

DLisaflexibleandhighlyefficienttechniquetosolvetheprob- lems in medical imaging analysis, including lung cancer, breast cancer, and braincancer detection [26]. With DLtechniques, au- tomaticsegmentationcanbedoneeffectivelyonalargeamountof MRIdatasets[27].Particularly,theCNNalgorithmforanautomatic brain MRI image segmentation has produced effectiveresults by considering severalfeatures [28]. As per recentstudies andsolu- tions thatprovidedtothisproblem,Hossam, etal.[29],proposed amulti-classificationofbraintumorMRIimagesbasedonDNN,it tries withtwo Publicly available datasets to make a newmodel, that firstly define the tumor type, andin thesecond stepdefine thegliomagradesofdetectimage.Inanotherstudyonbraintumor identification,AbduGumael,etal.[30] ahybridtechniqueregular- izedextremelearningmachine(RELM)usedtodevelopaneffective braintumorclassifiermodel.Theexperimentaloutput ofthepro- posedworkshows94.233%overallaccuracy.

Son etal. [34] focuson thelatestadvancementoverlooks into regardingMLforbigdatasystematic&variousstrategiesinterms of present-day figuring conditions for various cultural applica- tions.Chatterjee[35] attemptedtoprovideasignificantconception amongsttheIoTinBDstrategynearbyitsdifferentproblems,chal- lenges&triedtoprovideapossibleplanbyMLstrategy.Jhanzhiet al. [36] aposition&wormholeattacklocationsystemisproposed byutilizingtheMLmethod.Chatterjee[37] hasgivenashortsur- vey of how ML can be used in Bioinformatics. Jain& Chatterjee [38] givesakindofsummaryofpresent&risingMLstandardsfor healthcare informatics& mirrors the assorted variety, intricacy&

theprofundity&expansivenessofthismulti-disciplinaryregion.

It isdebatablethat wehave gonethrough manychallengesin theprocess ofanalyzingandexaminingbraintumors.Theauthor intheearlystagesthatarethebasementofresearchdidnothave access to high-quality images to analyze and diagnose brain tu- mors, this research issue had a problem, which in many cases disruptedtheauthor’sthoughts,butwiththeevaluationandpro- visionoftheregulardatabase,thisproblemwassolved.Inanother step,inordertoachieveauniquemodelandproposeanewmodel, theauthorusesexisting researchtoexaminethebestmodelsand designthe formatofthisresearch,which itselfhastakenalotof timeintheresearchstructure,andit’srememberabletowritethat adapting a modelandputtingit together,it hasbeenone of the challengesintheresearch process,thathassuccessfullyovercome thischallenge. The authorafter so manyreviewsalsofound that the proposed method could be used inother areasby providing detailedinformationandimprovementontheconcerntopic.

Inthiswork, MRIbrainimages arestudied usinga threshold- based segmentation method to detect the tumor regain in the image. Also, the detected image with the tumor is going to be classifiedusing hybrid CNNand SVM intoone ofthe considered types:BenignandMalignant.The hybridCNN-SVMmethod com- binedthe advantages ofbothCNN andSVM,whichare the most effectiveandpopularmethodsforimageclassification.Theperfor- manceofworkevaluatedbyparameterssuchasPositivePredictive Value (PPV), False Predictive Value (FPV), and accuracy. The im- plementationandcodingpartaredonewithPythonprogramming language (Version3.7.6)andalso forthispropose theOpenCV li- brary,TensorFlowlibrary,andTkinterformakingaproperGUIhas beenused.Themaincontributioninthisresearchareasfollows:

• ^{ThehybridproposedmodelcombinedCNNandSVMinterms} of classification and with threshold-based segmentation in termsofdetection.

• ^{The overall accuracy of the hybrid CNN-SVM obtained} 98.4959%.

• ^{The proposed hybrid model with} consideration of both CNN andSVMmodeladvantagesshowssignificantimprovement.

• ^{AHybridmethod,onbrainMRIimages,todetectandclassify} thetumorhasbeenimplementedusingtheBRATSdatabasein thisresearchstudy.

The rest of the article is structured as follows: the proposed hybridmethodofCNN&SVMiselaboratedinsection2.Section3 describedthemethodologyoftheproposedwork.Section4elabo- ratedontheutilizeddataset.Besides,theexperimentalresultsand comparisonarementionedinsection5.Conclusionofthepresent workisdiscussedinsections6.

2. ProposedhybridmethodofCNNandSVM

Theproposedhybrid(CNN-SVM)modelisdesignedtocombine both CNN& SVM advantages, aspresented in Fig.2, the general architecture ofCNNwithlayers (Convolution layer,Pooling Layer, flattenlayer,andFullyconnectedlayer)aredemonstrated[31].The hybrid model of CNN-SVM will be discussed at the end of this section.

2.1. Convolutionalneuralnetwork(CNN)

CNNisatypicaltypeofartificialneuralnetwork(ANN).Ingen- eral,thistypeofnetworkconsistsoftheinputlayer,hiddenlayer&

outputlayer[1,6,23,24].Typically,inthistypeofnetwork,theout- putofonelayer isusedasaninput forthenext layer.Asshown

Fig. 3.The structure of hybrid CNN & SVM.

inFig.2theCNNmodelisusedforimageanalysisthatconsistsof fourstages:convolutionlayer,Poolinglayer,Flattenlayer,andFully connectedlayer,howeverwithasamplelookwithotherarchitec- turesyoucancometoknowthatthenumberoflayers,aswellas typeoflayers,wouldbedifferent[23,24].

Basically, in theCNN model,the convolution layer is the first layerto extractthefeaturefroman image;withinthis, theprop- ertyofeachpixelandrelationofneighborpixelsisgoingtoextract with some mathematical operation. Afterextracting features, the pooling layer going to consider the most significant information by avoidinguseless data,this processing calledsubsampling and triestoreducethesizeofthedatamapwithoutlosingimporting dataanditcanbe(Maxpooling,Avgpooling,SumPooling),when thepoolinglayerappliedandtheall-importantfeatureismapped, sometimesthefeaturemappingmakesoverfitting,theflattenlayer 2D arrays to 1D arrays before applying a fully connected layer.

Typically, fully connected is the last layer of a network, and all network haveproperconnection,theoutput offullyconnectedis thefinaloutput[32].

2.2. Supportvectormachine(SVM)

It is a supervised ML that includes the linear and non-linear data and could be utilized for both classification or regression problems. Withinthisresearch, the SVM classifier is used in the lastlayerofthefullyconnectedlayerofCNNtoimprovetheflexi- bilityandeffectivelytofitthedatelengthtoturnthekernel[33].

2.3. HybridCNN-SVMmodel

The structure of the combined CNN-SVM model is going to design by replacing the last layer of CNN with the SVM model;

therefore, the output ofthe fullyconnected layer of CNN is go- ing to process asan input for SVM tomake better improvement on the classification. The main reason for combining the advan- tagesof CNN andSVM isthat theCNN modelwith rapidaccess to add hidden layers that have led to the process of extracting features, aswell as increasing accuracyand performance, hasal- ways been of paramount importance. However, SVM also, with itsuniquecharacteristicsinextractingfeatures,hashighefficiency and speed amongother algorithms. Having thiscombinationcan give us accessto thebest results inbrain tumor extractionwith MRIimages.Fig.3demonstratesthestructureofthehybridmodel of CNNand SVM. Inthis manner, the CNNmodel applies differ- ent convolutionandsubsampling andsimultaneouslyconvolution

layeronbasisof28×^{28 featuremap withhaving5}×^{5 convolu-} tion,andinthesamewayfeaturemap14×^{14 withhaving2}ײ convolution.Thatby applyingthisfunction aimsto speedup the extractionofinformationbasedonthetrainandtestprocess.After this,theSVMmodeltakestheoutputofafullyconnectedlayeras inputandrespectivelytrainsthefeaturevectors,theclassification, anddecisionmakinginthebetterway.

Inthis case, inaddition to thefact that the useof diagnostic knowledgecanhelpreduce mortalityfrombraintumors,thisisa bigstepintermsofscientificprogressandcanacceleratethediag- nosisoftumorsinthevery earlystages.Theauthorbelievesthat theproposed modelenables itby using theprinciples ofprecise layering andadjustingfiltersin imagesandfinally displayingthe tumorintheimage.Ontheotherhand,theauthorhasfoundthat combiningmodelssignificantlyintheprocessofinfluenceandre- sultsaremuchmoreeffectivethanusingasinglemodelalone.

3. Methodology

Fig. 4presents the methodology of the proposed hybrid CNN and SVM model and its stages. The architecture of the hybrid modelconcludedinfivegeneralsteps:

I. InputtheMRIImage

TheMRI image ofa brainistakento segmentthe tumorre- gion,withinthissteptheimagegoing totakefromaspecific directorytoloadontheGUIplatformandpasstothenextpart ofthesystem.

II. Pre-ProcessingonMRIBrainImage

Inthisstep,theprimaryfunctionsgoingtoimplementonthe MRIimagestomakesurethatthesystemcanreadtheproper input, and make a better condition for image analysis, the stepsareasfollows:

i. ResizetheImage: Withconsideration of input image the size ofeach image is differentin highandwidth, in this stage going to define afixed size like(32×^{32), to study} thewholedatasetinthesamecondition.

ii. RemovetheSkull:Duetothesignificantimportanceofthe brain, within this step, the skull around the brain with some functionsisusedtoremovethebackgroundandex- tracttheskullofthebrain.

iii. ImageFiltration:Atthisstage,themedianfilterisapplied toremovethenoise,whichcanmakeabetterpossibilityto

Fig. 4.DifferentstateofBrainMRIimage,usingthresholdsegmentation.(a)Originalimage(b)Testlevel1ofthreshold(c)Testlevel2ofthresholdthatdetecttheobjecton theimage.

detectfeaturesandalsofortrainingandtestingconditions it’susefulonMRIimages.

III. FeatureExtraction

Featureextractionisone oftheimportantfactorsthat canaf- fect the result. According to thisstudy, the researchers used differentalgorithms tofind out thefeature fromthe images.

Nabil Neggaz et al.[43] implement the Henry gases solubil- ityoptimization(HGSO)algorithmtoselectthefeatureon12 different datasets. The main point of this model that makes possible toevaluate the differentblock size ofimage feature to find the best factorto find a similar region to the query image. Meanwhile, in another work, Essam et al. [44] used the Harris hawks Optimization (HHO) and also in the same manner for better comparison it uses KNN and SVM in the feature classification process andfeature evaluation, andfor- tunately, this comparison shows good improvement. And in anotherstudyAbdel-Azimetal.[45] proposedanovelmethod based on binary optimizationtechniques that try by consid- eringthehyperbolic angleoffitness equation inan image to findthebestfeatureinanefficientmanner.

Therefore, with consideration of proposed methodology, as showninFig.3,andFig.5,theCNNmodelinthefirststage, by having different Convolution layer andpooling layer that is prepared with a different set of filter size is made possi- bletoextract thefeaturesofthe inputimage atthe moment ofthequery. Andalsoasillustratesin Fig.3,theCNNmodel consistsoffour convolution layerthat by implementingwith differentfiltersize,andsubsequentlyused theSVMmodelin thelastlayerofCNNandtrytoselectthebestfitdatatoim- provetheaccuracyoftheproposedmodel,andeventuallythe proposed modelobtain great improvementin comparison of existingwork.

IV. ImageSegmentation

Withconsiderationoftheimportanceofimage segmentation, andalso itscritical phaseon theextraction ofobjectsin the fieldofimageprocessing,patternrecognition.Basically,itpar- titions theinput imageintothemultiplesegmentsthatmake it easier to find the best fit data to detect and extract the needed area. The presentresearch studyuses the threshold- based segmentation model,by experimentingwith theinput image indifferentthresholdsandMaxvalue wecanenhance themeaningfulinformation.Inthemainpoint,thethreshold- based segmentation in the first phase divides the grayscale image into twoblack andwhiteblocks.As you seein Fig.4, having different threshold levels achieved the best result in brainMRIimages.

V. ClassificationofBrainMRIImages

The images afterproper feature extractionandsegmentation needtoclassify,basedonthisstep,firstly,themodelisgoing to train withSVM andCNN, respectively, going to test them

Fig. 5.Proposed Methodology of Hybrid CNN-SVM.

withSVM,CNN,andalsoHybridSVM+ ^{CNN,withPPV,FPV,} Accuracy parameter. And in the last stage going to make a comparativeanalysistoshowtheresultbasedontheparame- terthatisdefinedtoevaluatetheperformanceofthework.

4. Explanationofstatisticalfunctions

Withadeepunderstandingofthemathematicalfunctionroleto findtheeffectivedateofan image,within thisresearch,byusing different statistical functionstried to detect the best features in brainMRIImages.Thebriefinformationwithgoingtodescribeas below.

A. Meanvalue:Themeanprovidestheconsistencybrightnessof animage.Ithasahighquantityofbrightpixels.Inthisman- ner,themeandefinedthesumofall pixelsanddividedthem intothetotalamountofpixels.Thepreprocessingstepwithin

=

x-axis. And Finally, to have the power for all the point that doesn’tmissanypoint,wetrytoreachby“F(x,y)”toaccess everypointonanimage.

B. Variance: It is anotherfactor of an image, that specifies the rateofallocationingraylevels.Inacase,ifthereisavariation on the gray levelamount based on mean,then the variance increasedtoo.Thevariancevalueprovidesthemeasureofeach pixelfromthemeanamount.And,italsoprovidestheaverage ofthesquarebasedonthesingle-pixelandmeanvalue.

V ariance

=

1

mxn

m

−¹

x=⁰ n−¹

y=⁰

(

f

(

x

,

y

) −

^M

)

²

Inthe samemannerthevariance equation inadditionto the meanvalueittriestoevaluatetofindthedifferencebetween the gray level based on the average amount of mean using (f(x,y)−^M)².

C. StandardDeviation: An essential factor that tries to define thedifference betweenthemeandatavalues.Andinanother viewpoint, it can define the general square of variance, as shownbelow:

S D

( σ ) =

1 mxn

m

−1 x=0

n−1

y=0

(

f

(

x

,

y

) −

^M

)

²

InSD(

α

),themainpointtoshowthedifferencebetweenthe processed image data value andthe first datavalue that are available withinthe imageforthefirst time,basedon thisit caneasilyfindthechangesthatarecanaffectbysomefilters.

D. Entropy: Itusedtoevaluatetheheterogeneity ofMRIimages orintheplace thatwe trytomeet thetumor.We mustadd thisalsothatentropyisastatisticalfactorthatusestospecify thetexturedataontheimage.

Entropy

= −

m

−¹ x=⁰

n−¹

y=⁰

f

(

x

,

y

)

log2f

(

x

,

y

)

Besides of importance the entropy equation, by using (log 2f(x,y)) it tries to fit the statistical result on the best wayandfindtheexacttexturedataontheimage.

E. Energy:Energyplays themajor ruletofindthe similarityon eachpixeloftheimage.Besides this,thegraylevelisthees- sentialpart,andmore importantly,itprovidesthe sumof(f) squareofgraypixeldataonanimage.

Energ y

=

m

−¹ x=⁰

n−¹

y=⁰ f²

(

x

,

y

)

Accordingtothedetails thataredescribed theEnergy,byus- ing of (F) square (x,f), it triesto double-check every pixel tofindthe bestsimilardataoneach imageto showthebest result.

F. Homogeneity: It provides a degree value that evaluates the closest originbasedon the one GCLMmodelof an image to anotherGCLMmodelofanotherimage.

Homogeneit y

=

x,y

p

(

x

,

y

)

1

+ |

^x

−

^y

|

Withconsiderationofthematrixineveryimage,itnecessary to first evaluate the number of each matrix by using sum- mation x-axisandy-axis (x,y),within thisforbetter output it needs to find all point by (P(x,y)) and respectively and one step goes ahead andalso decrease that from the whole amountofxandy thatis(1+(x−^y)).

G. Correlation:itshowshowanypixelhascorrelatedtoanother pixelinanimage.

Correlation

=

_m−1

x=0

_n−1

y=0

(

x

,

y

)

f

(

x

,

y

) −

^{MxM y}

σ

x

σ

y

Anothermostsignificant factorthat canhelp tofindthebest possiblecorrelationbetweeneachpixelistheCorrelationfac- tor. In equation f(x,y) make way to access all x-axis and y-axis, and besides, it going to discard the average value of x and y based on the Mean factor to the findthe exact re- lationwithother pixels.In thelast stage, thelower sigmais usedtomeasurebothxand y andfindthestandardvalueto removetheerror.

H. Contrast:It showstheamountofchanges thatcanaffectthe waytomeasuretheintensityvalueamongwholeimagepixels and neighbor pixels. If there is a major change in the gray level,thenitshowsahighdegreeofdissimilarity.

Contrast

=

m

−1 x=0

n−1

y=0

(

x

,

y

)

²f

(

x

,

y

)

In which the ((x,y)²f(x,y)) shows that x and y both are by considering of x-axis and y-axis to reach the maximum amount that makes this possible to measure the intensity valueamongwholeimagepixelsandneighborpixels.

Withconsidering the statisticalfunctions, thisstudy observes the date that helps to find the best result as per the proposed hybrid model it shows the benign andmalignant min and max based on differentevaluated factors, the obtainedinformation is showninTable1.

5. Dataset

Thedatasetused inthisresearchwas takenfromBRATS2015 [39],withinthis,theBRATSareavailableinadifferentversion,that comestosolvemedicalimagingchallenge.TheBRATS2015imag- ingdatasetobtainedfromBRATS2012andBRATS2013,whichcan

Fig. 6.Illustration of Benign and Malignant Brain Image.

show an updated version of newcases withbetter performance.

It consistsoftwo type trainingwith(110 cases) andrespectively testingwith(220cases).

6. Experimentalresultsanddiscussion

The proposed modelevaluates theperformance ofwork ona publicimagedatabaseofBRATS2015andimplementsonDelllap- top with(Core i7 CPU,8 GBRam,and4 GB Nvidia GPU).In the includeddatabase,twotypesoftumorsthatarebenignandmalig- nantwithlowandhighgradesare available.Aspertheproposed methodology, afterpreprocessing,theextractedfeaturewithcon- sideration of training and testing is going to process andset in specific testsand train category as shown in Fig. 3.In thispro- cess, as shownin Fig. 3 andFig. 5, the CNN model,extract the features andwith various steps transfer to fully connected layer usingtheactivationfunction.Theextractedimagefeaturesgoingto train theCNNclassifier.Ontheother hand,theSVMmodeltakes thefullyconnectedoutput ofCNNasan inputtotheprocessand traineachimage.Atthelaststage,theclassifierisgoingtocatego- rizewhethertheinputimageismalignantorbenign,asshownin Fig.6.Eventually,theaccuracyofthehybridproposedCNN-SVMis evaluatedbyconsideringtheevaluationdefinedparameter.

6.1. Evaluationparameters

(a) Accuracy:Basedonimageclassification,theaccuracyisaper- centagethat showsthetotalamountofproperlyclassification ofpixelstothenumberofpixelsintheimage.Itevaluatesthe completeproperlypixelsinanimage.

Accuracy

=

^TP

+

TN

(

TP

+

^TN

+

^FP

+

^FN

)

(b) PPV: Inaddition toaccuracy, thetruepositive (TP) show the probabilityofpixelsthatcorrectlyclassified.

Positive Predictive Value

=

^TP

(

TP

+

^FP

)

Table 2

ComparisonofexistingmodelswithproposedHybridCNNandSVM.

The existing model along with the method and obtained accuracy

Existing works Method Accuracy

[31] RegularizedExtremeLearning Machine(RELM)

94.233%

[24] DeepConvolutionalNeuralNetwork (DCNN)

95%

[20] Deepneuralnetworks(DNN),deep waveletautoencoder(DWA)

96%

[40] K-Means,K-nearestneighbor(KNN) 96.6

[23] ConvolutionalNeuralNetworks(CNN) 97.5%

Proposed hybrid model classification accuracy

Stage SVM CNN Hybrid

Benign 61.67% 97.724% 98.5873%

Malignant 67.98% 97.8455% 98.6702%

(c) FPV: However, with underusing accuracy and TP, in image analysis,truenegative (TN) istheprobability pixelidentifica- tionthatlogicallyshowsnormalbutclassifiedasanabnormal feature.

False Predictive Value

=

^FP

(

FP

+

^TP

)

6.2.Resultanalysis

The proposed technique got better outcomes compared with existingmodels.InTable2,asshowntheaccuracyofexistingtech- niquescomparedwiththeproposedmodel.Meanwhile,theaccu- racyofhybridCNN-SVMinoverallcomes98.495%,inanotherview forclassificationbasedonBenigntumor withSVMgot61.67%,for CNNgot 97.724%and respectivelyon the hybrid modelobtained 98.5873%.Theclassification basedon Malignanttumor withSVM got67.98%,forCNN got97.8455%, andrespectivelyon thehybrid modelobtained98.6702%.

AsshowninFig.7,thecomparativeanalysisofthehybridpro- posedmodelshowsthatbyconsideringthestrongestpointofboth CNNandSVM,canleadustohaveimprovedmodelswithabetter

Fig. 7.ComparativeanalysisofCNN,SVMwithHybridCNN-SVMmodel.(Forinter- pretationofthecolorsinthefigure(s),thereaderisreferredtothewebversionof thisarticle.)

result onclassification. As per this study,the overall average ac- curacyisasflows,SVMwith72.5536%,respectivelytheCNNwith 97.4394%,andtheHybridmodelobtained98.4959%.However,with considerationofotherevaluationparameterssuchasPPVandFPV, wecometoknowthattheclassificationeffectsthisparameterthe sameasshowninFig.7.

If we take a closerlook atthe graph presented in Fig. 8,we will seethat theaccuracy levelofCNN andSVMseparately does notshowgoodprogress,thatthecompleteinformationmentionin Table2.Atthesametime,theaccuracydegreeofproposedhybrid CNNandSVMexpresssignificantprogressinthesameway.

However, FPV is another importantfeature forevaluating the correctperformanceofanactivity.Fig.9,showsthatCNNandSVM did not show a better result when it used individually, but the proposedmethodhasdonethisprocesswithhigheraccuracy.

Atthesametime,PPV,whichshowsthedegreeofcorrectim- ageclassification. Infact,withacloseview ofCNNandSVM,we come toknowthat the resultisnot satisfactory andneeds more

Brain tumor detection and classification due to abnormal growth of cells or portable progression throughout the body, is oneofthedeathlessdiseasesintheworld,amongother diseases.

AHybridmethod,onbrainMRIimagesto detectandclassifythe tumor, has been implemented using the BRATS database in this researchstudy.Theimplementedsystemtriestoclassifybrainim- ages as a benign and malignant tumor using supervised hybrid CNNand SVM techniques. The primary preprocessing steps have been implemented to normalize the input images, and also sig- nificant features extract from the preprocessed image by using the Maximally stable extremal regions (MSER) method and then segmentedwiththreshold-basedsegmentation technique. Thela- beled segmented features are pass as input to hybrid CNN and SVM algorithms to classify brain MRI images. The result shows the highestcorrectly classified withPPV andlowest of FPV, and overall, it achieved 98.4959% classified correctly this the hybrid model,whereas, SVMSeparately got72.5536%,andCNNobtained 97.4394%.Witha brieflookcame toknowthat theproposed hy- bridmodel provides more effectiveand improvementtechniques for classification. In future studies, for better decision making, the faster CNN with SVM andother optimization methods (bio- inspiredalgorithms)canbeusedtoshowtheoverallimprovement.

Anddueto thedetection ofthetumor,consideration ofsize and exactlocationofthetumoralsointhecase.

Fig. 8.Accuracy comparison of CNN, SVM with Hybrid CNN-SVM model.

Fig. 9.FPV comparison on CNN, SVM with Hybrid CNN-SVM model.

Fig. 10.PPV comparison on CNN, SVM with Hybrid CNN-SVM model.

Fig. 11.Comparison of Proposed work with the existing model.

Humanandanimalrights

The authors declarethat the workdescribed has not involved experimentationonhumansoranimals.

Funding

This work didnot receive anygrant fromfundingagencies in thepublic,commercial,ornot-for-profitsectors.

Authorcontributions

AllauthorsattestthattheymeetthecurrentInternationalCom- mitteeofMedicalJournalEditors(ICMJE)criteriaforAuthorship.

Declarationofcompetinginterest

Theauthorsdeclarethattheyhavenoknowncompetingfinan- cial orpersonalrelationships that couldbe viewedasinfluencing theworkreportedinthispaper.

References

[1]Yang A, Yang X, Wu W, Liu H, Zhuansun Y. Research on feature extrac- tionof tumor image based on convolutional neural network. IEEE Access 2019;7:24204–13.

[2]Bandyopadhyay SK.Pre-processingandsegmentationofbrainimagefortumor detection.JISUniversity,India2019;1:15–9.

[3]Li Y-Q,Chiu K-S,Liu X-R,Hsiao T-Y,Zhao G,Li S-J,etal.Polarization-sensitive opticalcoherencetomographyforbraintumorcharacterization.IEEEJSelTop QuantumElectron2019;25:1–7.

[4]Shakeel PM,Tobely TEE,Al-Feel H,Manogaran G,Baskar S.Neuralnetwork- basedbraintumordetectionusingwirelessinfraredimagingsensor.IEEEAc- cess2019;7:5577–88.

[5] Board CE. Cancer.Net. American Societyof Clinical Oncology (ASCO), 2019.

[Online].Available:https://www.cancer.net/cancer-types/brain-tumor/statistics.

[Accessed2019].

[6]Akkus Z,Galimzianova A,Hoogi A,Rubin DL,Erickson BJ.Deeplearningfor brainMRIsegmentation:stateoftheartandfuturedirections.JDigitImaging 2017;30(4):449–59.

[7]Fabelo H,Ortega S,Szolna A,Bulters D,Piñeiro JF,Kabwama S,etal.In-vivo hyperspectralhumanbrainimagedatabaseforbraincancerdetection.IEEEAc- cess2019;7:39098–116.

[8]Ahmadvand A,Daliri M.BrainMRimagesegmentationmethodsandapplica- tions.OMICSJRadiol2014;3(4):e130.

[9]Popoola JJ,Godson TE,Olasoji YO,Adu MR.Studyoncapabilitiesofdifferent segmentationalgorithmsindetectingandreducingbraintumorsizeinmag- neticresonanceimagingforeffectivetelemedicineservices.EurJEngResSci 2019;4:23–9.

[10]Hu K,Gan Q, Zhang Y,Deng S,Xiao F, Huang W,et al.Brain tumor seg- mentationusingmulti-cascadedconvolutionalneuralnetworksandconditional randomfield.IEEEAccess2019;7:92615–29.

[11]Sherlin DMD.ImplementationofhybridKNN–FCMforbraintumorsegmen- tation.IJEDR2017;5:1251–4.

[12]Anand A. Brain tumor segmentation using watershed technique and self- organizingmaps.IndianJSciTechnol2017;10.

[13]Arnaud A, Forbes F, Coquery N, Collomb N,Lemasson B, Barbier EL. Fully automatic lesion localizationand characterization: application to brain tu- mors usingmultiparametricquantitative MRIdata. IEEETransMedImaging 2018;37:1678–89.

[14]Kaur P,Kaur A.AnautomatedbraintumordetectioninMRIusingfireflyopti- mizedsegmentation.IntJAdvResComputSciSoftwEng2017;7.

[15]Shree NV,Kumar TNR.IdentificationandclassificationofbraintumorMRIim- ageswithfeatureextractionusingDWTandprobabilisticneuralnetwork.Brain Inform2018;5:23–30.

[16]Menze BH,Jakab A,Bauer S,Kalpathy-Cramer J,Farahani K,Kirby J,etal.The multimodalbraintumorimagesegmentationbenchmark(BRATS).IEEETrans MedImaging2014;34:1993–2024.

[17]Loganayaki M,Geetha K.Braintumoridentificationusingmulti-atlassegmen- tation.IJRTE2019;8.

[18]Sindhia R,ShankarMahalingam SP.BraintumordetectionusingMRIbyclassi- ficationandsegmentation.SSRGIntJMedSci2019;6:12–4.

[25] Gamage PT.Identificationofbraintumorusingimageprocessingtechniques.

Faculty of Information Technology, University of Moratuwa. https://www.

researchgate.net/publication/276133543,2017.

[26]Xing F,Xie Y,Yang L.Anautomaticlearning-basedframeworkforrobustnu- cleussegmentation.IEEETransMedImaging2015;35:550–66.

[27]Dahab DA,Ghoniemy SSA,Selim GM, et al.Automatedbrain tumordetec- tionandidentificationusingimageprocessingandprobabilisticneuralnetwork techniques.IntJImageProcessVisCommun2012;1:1–8.

[28]Baraiya N,Modi H. Comparativestudy ofdifferent methods for brain tu- morextractionfromMRIimagesusingimageprocessing.IndianJSciTechnol 2016;9:1–5.

[29]Sultan HH,Salem NM,Al-Atabany W.Multi-classificationofbraintumorimages usingdeepneuralnetwork.IEEEAccess2019.

[30]Gumaei A,Hassan MM,Hassan MR,Alelaiwi A,Fortino G.Ahybridfeature extractionmethodwithregularizedextremelearningmachineforbraintumor classification.IEEEAccess2019;7:36266–73.

[31]Sec S,Mahalakshmi R.Anefficientstrategyoverbraintumouranalysisusing imageprocessingtechnique.IntJPharmTechnol2017;9:30687–96.

[32]Demirhan A,Törü M,Güler I.Segmentationoftumorandedemaalongwith healthytissuesofbrainusingwaveletsand neuralnetworks.IEEEJBiomed HealthInform2014;19:1451–8.

[39]https://www.smir.ch/BraTS/Start2015#!#download.

[40]Anitha V,Murugavalli S.Braintumourclassificationusingtwo-tier classifier withadaptivesegmentationtechnique.IETComputVis2016;10(1):9–17.

[41]Mostafa A,Houssein EH,Houseni M,Hassanien AE,Hefny H.Evaluatingswarm optimization algorithms for segmentation ofliver images. In: Advances in soft computingand machine learning inimageprocessing. Springer; 2018.

p. 41–62.

[42]Said S, Mostafa A, Houssein EH, Hassanien AE, Hefny H. Moth-flame optimization-basedsegmentationforMRIliver images.In:Internationalcon- ferenceonadvancedintelligentsystemsandinformatics;2017.

[43]Neggaz N,Houssein EH,Hussain K.AnefficientHenrygassolubilityoptimiza- tionforfeatureselection.ExpertSystAppl2020:113364.

[44]Houssein EH,Hosney ME,Oliva D,Mohamed WM,Hassaballah M.Anovelhy- bridHarrishawks’optimizationandsupportvectormachinesfordrugdesign anddiscovery.ComputChemEng2020;133:106656.

[45]Hussien AG,Houssein EH,Hassanien AE.A binarywhaleoptimizationalgo- rithmwithhyperbolictangentfitnessfunctionforfeatureselection.In:2017 eighthinternationalconferenceonintelligentcomputingandinformationsys- tems(ICICIS);2017.