Вы находитесь на странице: 1из 9

3/8/2017 Featureselectionusinggeneticalgorithmforbreastcancerdiagnosis:experimentonthreedifferentdatasets

IranJBasicMedSci.2016May19(5):476482. PMCID:PMC4923467

Featureselectionusinggeneticalgorithmforbreastcancerdiagnosis:
experimentonthreedifferentdatasets
ShokoufehAalaei, 1HadiShahraki, 2AlirezaRowhanimanesh, 3andSaeidEslami4,1,5,*
1
DepartmentofMedicalInformatics,SchoolofMedicine,MashhadUniversityofMedicalSciences,Mashhad,Iran
2
DepartmentofElectricalEngineering,FacultyofEngineering,UniversityofBirjand,Birjand,Iran
3
RoboticsLaboratory,DepartmentofElectricalEngineering,UniversityofNeyshabur,Neyshabur,Iran
4
PharmaceuticalResearchCenter,SchoolofPharmacy,MashhadUniversityofMedicalSciences,Mashhad,Iran
5
DepartmentofMedicalInformatics,AcademicMedicalCenter,Amsterdam,TheNetherlands
*
Correspondingauthor:SaeedEslami.PharmaceuticalResearchCenter,SchoolofPharmacy,MashhadUniversityofMedicalSciences,Mashhad,Iran
DepartmentofMedicalInformatics,SchoolofMedicine,MashhadUniversityofMedicalSciences,Mashhad,IranDepartmentofMedicalInformatics,
AcademicMedicalCenter,Amsterdam,TheNetherlands.Tel:+9851380022429email:EslamiS@mums.ac.ir

Received2014May15Accepted2016Mar3.

Copyright:IranianJournalofBasicMedicalSciences

ThisisanopenaccessarticledistributedunderthetermsoftheCreativeCommonsAttributionNoncommercialShareAlike3.0Unported,whichpermits
unrestricteduse,distribution,andreproductioninanymedium,providedtheoriginalworkisproperlycited.

Abstract Goto:

Objective(s):
Thisstudyaddressesfeatureselectionforbreastcancerdiagnosis.Thepresentprocessusesawrapperapproach
usingGAbasedonfeatureselectionandPSclassifier.Theresultsofexperimentshowthattheproposedmodelis
comparabletotheothermodelsonWisconsinbreastcancerdatasets.

MaterialsandMethods:
Toevaluateeffectivenessofproposedfeatureselectionmethod,weemployedthreedifferentclassifiersartificial
neuralnetwork(ANN)andPSclassifierandgeneticalgorithmbasedclassifier(GAclassifier)onWisconsinbreast
cancerdatasetsincludeWisconsinbreastcancerdataset(WBC),Wisconsindiagnosisbreastcancer(WDBC),and
Wisconsinprognosisbreastcancer(WPBC).

Results:
ForWBCdataset,itisobservedthatfeatureselectionimprovedtheaccuracyofallclassifiersexpectofANNand
thebestaccuracywithfeatureselectionachievedbyPSclassifier.ForWDBCandWPBC,resultsshowfeature
selectionimprovedaccuracyofallthreeclassifiersandthebestaccuracywithfeatureselectionachievedbyANN.
Alsospecificityandsensitivityimprovedafterfeatureselection.

Conclusion:
Theresultsshowthatfeatureselectioncanimproveaccuracy,specificityandsensitivityofclassifiers.Resultofthis
studyiscomparablewiththeotherstudiesonWisconsinbreastcancerdatasets.

Keywords:Breastcancer,Classificationfeature,Selectiondatamining

Introduction Goto:

Amajorclassofproblemsinmedicalscienceinvolvesthediagnosisofdisease,basedonanumberoftestsdoneon
thepatients.Becauseofwelterofdata,theultimatediagnosismaybedifficulttoobtain,evenforamedicalexpert.

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4923467/ 1/9
3/8/2017 Featureselectionusinggeneticalgorithmforbreastcancerdiagnosis:experimentonthreedifferentdatasets

Improvementsinfacilitiescausedverylargedatabasescanbecollectedinmedicinewhichneedstodiscover
relationshipsburiedindata.Dataminingapproachesinmedicaldomainareusingintensivelyforthesepurposes(1,
2).Oneoftheapplicationareasofanalysingdatabaseisautomateddiagnosticsystems.Thesesystemscanhelp
doctorsintheirdecisionmaking.Anotherapplicationisfindingwaystoimprovepatientoutcome,reducecostand
enhanceclinicalstudies.Inaddition,needforautomateddiagnosishasbeenmostacuteincaseofdeadlydisease
likecancerwhereearlydetectioncangreatlyenhancethechancesoflongtermsurvivalandreducethecosts.
Breastcancerconsideredthemostcommoninvasivecancerinwomen.InUSA,itisconsideredtobesecond
leadingcauseofmortalityamongwomenandthemostcommoncauseofmortalityintheagegroup40to55years
women(3).Theeffectivenessofearlydetectionhasbeenproventoreducealotofmortalityamongpatientswith
breastcancer(4).

Therearethreeclassicalmethodsavailablefordetectingbreastcancer:physicalexam,mammographyandbiopsy
includingFineneedleaspirationbiopsy(FNABorFNAC),Coreneedlebiopsy,Surgicalbiopsy,Lymphnode
biopsy(5).

Mammographyisoneofthemostusedmethodstodetectthebreastcancer.Inliterature,radiologistsshow
considerablevariationininterpretingamammography(6).Accuracyofmammographyvariesfrom68%to79%
(7).Whenmammographydetectsatumour,biopsyisrequiredtodetermineitsmalignancy.Theaccuracyof
surgicalbiopsyisnearly100%butitiscostly,invasive,timeconsumingandpainful.FNACisalsowidelyadopted
inthediagnosisofbreastcancer.TheaccuracyofFNACwithvisualinterpretationvariesfrom35%to95%
dependingontheexperienceofadoctor(8).So,itisnecessarytodevelopbetteridentificationmethodsto
recognizethebreastcancer.Theseidentificationmethodscanhelptoassignpatientstoeitherabenigngroupthat
doesnothavebreastcanceroramalignantgroupwhohasstrongevidenceofhavingbreastcancer.

Malignanttumoursgenerallyaremoreseriousthanbenigntumours.Asmentioned,earlydetectionofbreastcancer
leadstomuchhigherchancesofsuccessfultreatment.Inordertoreachthisgoal,itisnecessarytohavediagnostic
systemswithhighlevelsofaccuracyandreliabilitythathelpdoctorstodistinguishbetweenbenignbreasttumours
andmalignantones.

Oneoftheproblemsindiagnosticsystemsisthemultiplicityoffeatures.Irrelevancyandredundancyinthese
featuresincreasetheconfusionofclassificationalgorithmanddecreaselearningprecision(9,10).Featureselection
isoneofthemethodsthatcancopewiththisproblemandplaysanimportantroleinclassification.Featureselection
isoneofthepreprocessingtechniquesindataminingandextensivelyusedinthefieldsofstatistics,pattern
recognitionandmedicaldomain.

TherearethreeapproachesforfeatureselectionincludingWrapper,FilterandEmbedded(11).Inwrapperapproach
thegoodnessofselectedsubsetoffeaturesdeterminedbylearningandevaluatingaclassifierusingonlythe
variablesincludedintheproposedsubset.Filterapproachusessometechniquestoscoretheselectedsubset,
ignoringclassifieralgorithm.Inotherwordgoodnessofselectedsubsetoffeaturesdeterminedbyusingonly
intrinsicpropertiesofthedata(12).Inembeddedapproach,selectingthebestsubsetoffeaturesisperformedduring
themodelconstructionprocess.

Agoodamountofresearchonbreastcancerdatasetsusingfeatureselectionmethodsisfoundinliteraturesuchas
antcolonyalgorithm(13),adiscreteparticleswarmoptimizationmethod(14),wrapperapproachwithgenetic
algorithm(15),supportvectorbasedfeatureselectionusingfisherslineardiscriminateandsupportvectormachine
(16),fastcorrelationbasedfeatureselection(FCBF),multithreadbasedFCBFfeatureselectionanddecision
dependentdecisionindependentcorrelation(DDCDIC)(17),RoughsetKMeansClustering(18),modification
correlationroughsetfeatureselection(MCRSFS)(19).

Inthisstudyawrapperfeatureselectionmethodisproposedbasedongeneticalgorithmbasedfeatureselection.
Thismodelemployedparticleswarmoptimizationalgorithmbasedclassifier(PSclassifier)asfitnessfunction.The
modelevaluatedonWisconsinbreastcancerdatabases.

MaterialsandMethods Goto:

DatasetDescription(Wisconsinbreastcancerdatabases)
Inthisstudy,theWisconsinbreastcancerdatasetsfromUCIMachineLearningRepositoryisused(20).Theyhave
beencollectedbyDr.WilliamH.Wolberg(19891991)attheUniversityofWisconsinMadisonHospitals.The

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4923467/ 2/9
3/8/2017 Featureselectionusinggeneticalgorithmforbreastcancerdiagnosis:experimentonthreedifferentdatasets

detailofthesedatasetsisshownintable1.

Table1
Wisconsinbreastcancerdatasets(18)

InWBCdatasetthereare699recordsthateachrecordhasnineattributesexpectofidnumberandclass.Thesenine
attributesaregradedonanintervalscalefromanormalstateof110,with10beingthemostabnormalstate(
Table2).Inthisdatabase,241(65.5%)recordsaremalignantand458(34.5%)recordsarebenign.

Table2
Wisconsinbreastcancer(WBC)Attribute(20)

InWDBCthereare569recordsthateachrecordhasthirtyattributesexpectofidnumberandclass.Featuresare
computedfromadigitizedimageofafineneedleaspirate(FNA)ofabreastmass.Theydescribecharacteristicsof
thecellnucleipresentintheimage.

Tenrealvaluedfeaturesarecomputedforeachcellnucleus:

a.radius(meanofdistancesfromcentertopointsontheperimeter)
b.texture(standarddeviationofgrayscalevalues)
c.perimeter
d.area
e.smoothness(localvariationinradiuslengths)
f.compactness(perimeter^2/area1.0)
g.concavity(severityofconcaveportionsofthecontour)
h.concavepoints(numberofconcaveportionsofthecontour)
i.symmetry
j.fractaldimension(coastlineapproximation1)(20).

Themean,standarderror,andworstorlargest(meanofthethreelargestvalues)ofthesefeatureswerecomputed
foreachimage,resultingin30features.Forinstance,field3isMeanRadius,field13isRadiusSEandfield23is
WorstRadius.

TheWPBCandWDBChavethesamefeaturesyettheWPBChastwoadditionalfeaturesasfollows:

Tumoursizethatisthediameteroftheexcisedtumourincentimetersandlymphnodestatusthatisnumberof
positiveaxillarylymphnodesobservedattimeofsurgery.

Featureselection
Featureselectionisaprocessthatreducesthenumberofattributesandselectsasubsetoforiginalfeatures.Feature
selectionisoftenusedindatapreprocessingtoidentifyrelevantfeaturesthatareoftenunknownpreviousand
removesirrelevantorredundantfeatureswhichdonothavesignificanceinclassificationtask.Featureselection
aimstoimprovetheclassificationaccuracy(9).

Geneticalgorithm
Geneticalgorithm(GA),originallydevelopedbyHolland,isacomputationaloptimizationparadigmmodelledon
theconceptofbiologicalevolution(21).TheGAisanoptimizationprocedurethatoperatesinbinarysearchspaces
andmanipulatesapopulationofpotentialsolutions.Apointinthesearchspaceisrepresentedbyafinitesequence
of0sand1s,calledachromosome.Thequalityofpossiblesolutionsisevaluatedbyafitnessfunction.The
probabilityofsurvivalisproportionaltothechromosomesfitnessvalue.InGA,theinitialpopulationisrandomly
generatedbythreeoperators:selection,crossover,andmutation.Theselectionoperatorselectselitestotransfer
directlytonextgeneration.Thecrossoveroperatorrandomlyswapsaportionofchromosomesbetweentwochosen
parentstoproduceoffspringchromosomes.Themutationoperatorrandomlyalertsabitinchromosomes.

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4923467/ 3/9
3/8/2017 Featureselectionusinggeneticalgorithmforbreastcancerdiagnosis:experimentonthreedifferentdatasets

InthisworkGAisusedtoeliminateinsignificantfeatures.Inordertoreachthispurpose,wedefinedchromosomes
asamaskforfeatures.Inotherword,eachchromosomeisasubsetoffeatures.Thesizeofchromosome(numberof
genes)isequaltothenumberoffeaturesthatrepresentthespecificationofacancerpatient.Asmentioned,a
chromosomeisrepresentedinformofbinarystringthatis0or1.1meansthecorrespondingfeatureisselectedand
0meansitisnotselected(Figure1).

Figure1
Generatinginitialpopulation

Evaluationfunction
Thegoaloftheproposedmodelisselectingthebestsubsetoffeaturesthatcanproducethehighestclassification
accuracyfordiagnosisandprognosisthebreastcancer.Therefore,thebestsubsetoffeaturesshouldbeselected.
Forselectingthebestsubset,afunctionisneededtoevaluatetheresultofselectingeachsubsetoffeatures
(chromosome).

Inthisworkweusedaclassifierbasedontheparticleswarmoptimizationalgorithm(PSclassifier)whichisanovel
classifierthatproposedbyZahiriandSeyedin(22).

TheparticleswarmoptimizationdevelopedbyKennedyandEberhart(23).Thisoptimizationmethodisbasedon
thebehaviourofswarmofbeesorflockofbirdswhilesearchingforfood.InPSO,theparticlesflythroughthe
problemspacebyfollowingtheoptimalparticles.Eachparticleremembersthebestpositionthatithasvisited
(Pbest)andalsobestpositionamongalltheparticlesinthepopulation(Gbest).Thepositionofeachparticle
changesaccordingtothePbestandGbestintheproblemspace.

InPSclassifier,PSOalgorithmisusedtofindthedecisionhyperplanesbetweenthedifferentclasses.Decision
hyperplanesareemployedtodividefeaturespaceintoindividualregions.Eachregionisassignedtoaspecific
class.

Ageneralhyperplaneisintheformof

whereX=(x1,x2,,xn)andW=(w1,w2,,wn+1)arecalledtheaugmentedfeatureandweightvector,
respectively.nisthefeaturespacedimension.

Inageneralcase,thereareanumberofhyperplanesthatseparatethefeaturespacetodifferentregions,thateach
regiondistinguishesanindividualclass(Figure2).

Figure2
Separatingtwoclasseswithonehyperplane

ThePSclassifiermustfindWj(j=1,2,,H)insolutionspace,whereHisthenecessarynumberofdecisionhyper
planes.

FitnessfunctionofPSclassifierisdefinedasfollow:

whereMissisthenumberofmisclassifieddatapointsbyW.

Featureselectionprocess
ThefeatureselectionprocessisrepresentedinFigure3.ItisobservedthatGAselectssubsetoffeaturesas
chromosomesandeachchromosomeissenttothePSclassifierforcalculatingfitnessvalue.PSclassifieruseseach
chromosomeasmaskforfeatures.Sothateachgeneonchromosomedeterminesthecorrespondingfeatureshould

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4923467/ 4/9
3/8/2017 Featureselectionusinggeneticalgorithmforbreastcancerdiagnosis:experimentonthreedifferentdatasets

beusedinPSclassifierornot.PSclassifierdeterminesafitnessvalueforeachchromosomesandGAusesthese
fitnessvaluestotheprocessofchromosomeevolution.FinallyGAfindsanoptimalsubsetoffeatures.

Figure3
Proposedfeatureselectionflowchart

Inproposedmodel,thenumberofchromosomesineachpopulation(sizeofpopulation)is150andmaximum
iterationis300.Themutationrateis0.4andcrossoveris0.5andeliterateis0.1.AlsoforPSclassifier,swarmsize
of150wasselectedandinitialinertiaweightwaschosen0.7.

Predictionmodels
Inthisstudyweuseddifferentclassifieralgorithmsnamelyartificialneuralnetwork(ANN),PSclassifierandGA
classifierassubsetevaluatingmechanismonWisconsinbreastcancerdatasets(WBCD).

Inthisworkwebuildthree3layerneuralnetworksbyusingnprtoolinMatlabsoftware.Artificialneuralnetworks
areacomputationaltool,basedonthepropertiesofbiologicalneuralsystems.GAclassifierisanotherclassifierthat
isusedtoevaluateproposedmethodanditispresentedbyBandyopadhyayetal(24).Thenumberofchromosomes
ineachpopulation(sizeofpopulation)is150andmaximumiterationis300.Themutationrateis0.4andcrossover
is0.5andeliterateis0.1.ThethirdselectedclassifierisPSclassifierthatwasdescribedbefore.

Inordertoevaluatetheclassificationefficiency,threemainmetricsincludingaccuracy,sensitivityandspecificity
havebeencomputedfortheclassifiers.Thesemetricsarecalculatedfrom:

WhereTNisnumberofTrueNegatives,TPisnumberofTruePositives,FNisnumberofFalseNegativesandFP
isnumberofFalsePositives.

Ourtrainingandtestingwasiterated30timesforeachclassifierandaverageofresultswasexpressedasthefinal
result.80%ofdataisallocatedtotrainingsetandtheremaining20%isallocatedtotestset(incaseofANN,20%
ofdataallocatedtovalidatingset).

Itshouldbenotedthatparameterstuningoftheclassifiersareequalbeforeandafterfeatureselection.

Results Goto:

ProposedfeatureselectionmethodwasappliedonWisconsinbreastcancerdatabasesandTable3showsselected
relevantfeatures.

Table3
Selectedfeaturesafterapplyingfeatureselectionmethod

Inneuralnetwork,thelayersincludeaninputlayerof9,30and33discretevariableswithWBC,WDBC,WPBC
datasets,respectivelywithoutfeatureselection.Afterfeatureselectionwebuildlayersincludeaninputlayerof4,
14and16discretevariables.Inallnetworksweconsideredahiddenlayerwith5nodesandanoutputlayerwith2
nodes.

Wisconsinbreastcancerdataset(WBC)

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4923467/ 5/9
3/8/2017 Featureselectionusinggeneticalgorithmforbreastcancerdiagnosis:experimentonthreedifferentdatasets

WeusedclassifierswithandwithoutfeatureselectionwithWBCdataset.ResultsaresummarizedintheTable4.

Table4
TheSensitivity,specificityandaccuracyof3classifierswithandwithout
featureselection(FS)usingWBCdataset

Wisconsindiagnosisbreastcancer(WDBC)
WeemployeddescribedclassifiersonWDBC.Thecomparisonofaverageaccuraciesforthethreeclassifiers
(ANN,PSclassifier,GAclassifier)withandwithoutfeatureselectionisshowninTable5.

Table5
TheSensitivity,specificityandaccuracyof3classifierswithandwithout
featureselection(FS)usingWDBCdataset

Wisconsinprognosisbreastcancer(WPBC)
ResultsofemployingthreedescribedclassifiersonWPBCaresummarizedintheTable6.

Table6
TheSensitivity,specificityandaccuracyof3classifierswithandwithout
featureselection(FS)usingWPBCdataset

Discussion Goto:

InthisstudyafeatureselectionmodelwithGAbasedonfeatureselectionisdesignedtoidentifyrelevantfeatures.
GAhasmorerecentlydevelopedincomparetodifferentfeatureselectionalgorithms.GAcanbeusefultofeature
selectionwhentheproblemhasexponentialsearchspace.TherearemanyadvantagesoftheGAsforfeature
selectionthathavepublishedinvariousliteratures(25,26).

Thecomparisonofaverageaccuraciesforthethreeclassifiers(ANN,PSclassifier,GAclassifier)withandwithout
featureselectiononWBCdatasetshowedthatwithoutfeatureselectiontheaccuracyofANN(96.8%)isthebest
andtheaccuracyobtainedbyPSclassifierisbetterthanthatproducedbyGAclassifier(96.2vs.96.08).Itis
observedthatfeatureselectionimprovedtheaccuracyofallclassifiersexpectofANNandthebestaccuracywith
featureselectionachievedbyPSclassifier(96.9%).Alsoitisapparentfromresultsobtainedthatspecificityand
sensitivityhasbeenapproximatelyimprovedbyfeatureselection.

Table7showsacomparisonbetweenclassificationaccuraciesofotherpublishedstudieswhichuseddifferent
featureselectionmethodsandtheaccuraciesobtainedbyANN,PSclassifierandGAclassifierinthisworkon
WBCdataset.

Table7
Comparisonofexperimentalresultsofproposedmethodandotherpapersin
WBC

ForWDBCdataset,ANNclassifiershowsthebestaccuracy(96.5%).FromTable5itisobviousthattheANN
accuracywithWDBCiswellthanPSclassifierandGAclassifieraccuraciesrespectively(96.4vs.96.1).Results
showfeatureselectionimprovedaccuracyofallthreeclassifiersandthebestaccuracywithfeatureselection
achievedbyANN(97.3%).AlsoTable5showsthatspecificityandsensitivitycanimproveafterfeatureselection.

Table8showsacomparisonbetweenclassificationaccuraciesofotherpublishedstudieswhichuseddifferent
featureselectionmethodsandtheaccuraciesobtainedinthisworkonWDBCdataset.

Table8
Comparisonofexperimentalresultsofproposedmethodandotherpapersin
WDBC

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4923467/ 6/9
3/8/2017 Featureselectionusinggeneticalgorithmforbreastcancerdiagnosis:experimentonthreedifferentdatasets

ThecomparisonofaverageaccuraciesforthedescribedclassifierswithandwithoutfeatureselectiononWPBC
showedthatwithoutfeatureselectiontheaccuracyofPSclassifier(77.8%)isthebestandtheaccuracyobtainedby
ANNisbetterthanthatproducedbyGAclassifier(77.4vs.76.3).Itisclearthatfeatureselectionimprovedthe
accuracyofallthreeclassifiersandthebestaccuracywithfeatureselectionachievedbyANN(79.2%).Alsoascan
beseenfromthetable8,thespecificityandsensitivityimprovedafterfeatureselection.Theresultofthisdatasetis
comparablewithotherstudies(35).

Table9showsacomparisonbetweenclassificationaccuraciesofotherpublishedstudieswhichuseddifferent
featureselectionmethodsandtheaccuraciesobtainedbythreedifferentclassifiersinthisworkonWPBCdataset.

Table9
Comparisonofexperimentalresultsofproposedmethodandotherpapersin
WPBC

Itshouldbenotedwhiledataminingcanfacilitateanalysingoflargedatabasesandhelpmedicalstaffindecision
makingweshouldconsiderthelimitationsofwhatitcando.dataminingtechniquescandiscoverpatternburiedin
databutitcantreplacephysiciansinsights(36).Alsosometimestheincreaseinthenumberoffeaturesleadstothe
decreaseinthespeedofthealgorithm.Thereforeidentifyingpatternsmaybetimeconsuming.

Conclusion Goto:

Inthispaper,weproposedafeatureselectionmethodusingGAforselectingthebestsubsetoffeaturesforbreast
cancerdiagnosissystem.

ANN,PSclassifierandGAclassifierwereusedtoevaluateproposedfeatureselectionmethodonWisconsin
BreastCancerDatasets.InWBC,theclassificationusingPSclassifierissuperiortootherclassification.InWDBC
andWPBC,ANNachievedthebestaccuracy.Theresultsshowthatfeatureselectioncanimproveaccuracyof
classifiers.ResultofthisstudyiscomparablewiththeotherstudiesonWisconsinbreastcancerdatasets.

Acknowledgements Goto:

WethankDrWilliamHWolbergattheUniversityofWisconsinforsupportinguswiththebreastcancerdataset
whichwehaveusedinourexperiments.

References Goto:

1.SarbazM,PournikO,GhalichiL,KimiafarK,RazaviAR.DesigningaHumanTLymphotropicVirusType1
(HTLVI)DiagnosticModelUsingtheCompleteBloodCount.IranJBasicMedSci.201316:247.
[PMCfreearticle][PubMed]

2.TayaraniA,BaratianA,SistaniMB,SaberiMR,TehranizadehZ.Artificialneuralnetworksanalysisusedto
evaluatethemolecularinteractionsbetweenselecteddrugsandhumancyclooxygenase2receptor.IranJBasicMed
Sci.201316:1196.[PMCfreearticle][PubMed]

3.Breastcancer.org:Knowingyourriskcansaveyourlife[Internet]Breastcancer.org.2016.[cited12May2016].
Availablefrom:http://www.breastcancer.org.

4.BashaSS,PrasadKS.Automaticdetectionofbreastcancermassinmammogramsusingmorphological
operatorsandfuzzycmeansclustering.JTheorApplInfTechnol.2009:5.

5.Howisbreastcancerdiagnosed?[Internet]Cancer.org.2016.[cited12May2016].Availablefrom:
http://www.cancer.org/cancer/breastcancer/detailedguide/breastcancerdiagnosis.

6.ElmoreJG,WellsCK,LeeCH,HowardDH,FeinsteinAR.Variabilityinradiologistsinterpretationsof
mammograms.NEnglJMed.1994331:14931499.[PubMed]

7.FletcherSW,BlackW,HarrisR,RimerBK,ShapiroS.Reportoftheinternationalworkshoponscreeningfor
breastcancer.JNatCancerInst.199385:16441656.[PubMed]

8.WillemsSM,VanDeurzenCH,VanDiestPJ.Diagnosisofbreastlesions:fineneedleaspirationcytologyorcore
needlebiopsy?Areview.Jclinpathol.201265:287292.[PubMed]
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4923467/ 7/9
3/8/2017 Featureselectionusinggeneticalgorithmforbreastcancerdiagnosis:experimentonthreedifferentdatasets

9.KohaviR,JohnGH.Wrappersforfeaturesubsetselection.ArtifIntell.199797:273324.

10.AbeN,KudoM,ToyamaJ,ShimboM.Adivergencecriterionforclassifierindependentfeatureselection.
AdvancesinPatternRecognition:Springer2000:668676.

11.GuyonI,ElisseeffA.Anintroductiontovariableandfeatureselection.JMachLearnRes.20033:11571182.

12.BermejoP,GmezJA,PuertaJM.AGRASPalgorithmforfasthybrid(filterwrapper)featuresubsetselection
inhighdimensionaldatasets.PatternRecognitLett.201132:701711.

13.AghdamMH,GhasemAghaeeN,EhsanBasiriM.Applicationofantcolonyoptimizationforfeatureselection
intextcategorization.EvolutionaryComputation,2008CEC2008(IEEEWorldCongressonComputational
Intelligence)IEEECongresson2008:IEEE

14.UnlerA,MuratA.Adiscreteparticleswarmoptimizationmethodforfeatureselectioninbinaryclassification
problems.EurJOperRes.2010206:528539.

15.KaregowdaAG,JayaramM,ManjunathA.Featuresubsetselectionproblemusingwrapperapproachin
supervisedlearning.IntJComputAppl.20101:1317.

16.YounE,KoenigL,JeongMK,BaekSH.SupportvectorbasedfeatureselectionusingFisherslinear
discriminantandSupportVectorMachine.ExpSystAppl.201037:61486156.

17.DeisyC,SubbulakshmiB,BaskarS,RamarajN.Efficientdimensionalityreductionapproachesforfeature
selection.ConferenceonComputationalIntelligenceandMultimediaApplications,2007InternationalConference
on2007:IEEE

18.SrideviT,MuruganA.AnintelligentclassifierforbreastcancerdiagnosisbasedonKMeansclusteringand
roughset.IntJComputAppl.201485:3842.

19.SrideviT,MuruganA.Anovelfeatureselectionmethodforeffectivebreastcancerdiagnosisandprognosis.Int
JComputAppl.201488:2833.

20.UCIMachineLearningRepository:BreastCancerWisconsin(Diagnostic)DataSet[Internet]
Archive.ics.uci.edu.2016.[cited12May2016].Availablefrom:http://archive.ics.uci.edu/ml/
datasets/Breast+Cancer+Wisconsin+%28Diagnostic%29.

21.HollandJH.Adaptationinnaturalandartificialsystems:Anintroductoryanalysiswithapplicationstobiology,
control,andartificialintelligence.UMichiganPress1975.

22.ZahiriSH,SeyedinSA.Swarmintelligencebasedclassifiers.JFranklinInst.2007344:3623676.

23.KennedyJ,EberhartR.Particleswarmoptimization.ProceedingsoftheIEEEInternationalConferenceon
NeuralNetworks.1995

24.BandyopadhyayS,MurthyCA,PalSK.Theoreticalperformanceofgeneticpatternclassifier.JFranklinInst.
1999336:387422.

25.OhIS,LeeJS,MoonBR.Hybridgeneticalgorithmsforfeatureselection.IEEETransPatternAnalMach
Intell.200426:14241437.[PubMed]

26.HadizadehF,VahdaniS,JafarpourM.QuantitativeStructureActivityRelationshipStudiesof4Imidazolyl1,
4dihydropyridinesasCalciumChannelBlockers.IranJBasicMedSci.201316:910916.[PMCfreearticle]
[PubMed]

27.LavanyaD,RaniDK.Analysisoffeatureselectionwithclassification:Breastcancerdatasets.IndianJournalof
ComputerScienceandEngineering(IJCSE)20112:756763.

28.KarabatakM,InceMC.Anexpertsystemfordetectionofbreastcancerbasedonassociationrulesandneural
network.ExpSystAppl.200936:34653469.

29.ChenHL,YangB,LiuJ,LiuDY.Asupportvectormachineclassifierwithroughsetbasedfeatureselection
forbreastcancerdiagnosis.ExpSystAppl.201138:90149022.

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4923467/ 8/9
3/8/2017 Featureselectionusinggeneticalgorithmforbreastcancerdiagnosis:experimentonthreedifferentdatasets

30.SenturkZK,KaraR.BreastCancerDiagnosisviaDataMining:PerformanceAnalysisofSevendifferent
algorithms.ComputerScience&Engineering.20144:35.

31.NoruziA,SahebiH.Agraphbasedfeatureselectionmethodforimprovingmedicaldiagnosis.AdvComput
Sci.20154:3640.

32.ZhaoJY,ZhangZL.Fuzzyroughneuralnetworkanditsapplicationtofeatureselection.Advanced
ComputationalIntelligence(IWACI),2011FourthInternationalWorkshopon2011:IEEE

33.LiuY,ZhengYF.FS_SFS:Anovelfeatureselectionmethodforsupportvectormachines.PatternRecognit.
200639:13331345.

34.DumitruD.PredictionofrecurrenteventsinbreastcancerusingtheNaiveBayesianclassification.Annalsof
theUniversityofCraiovaMathematicsandComputerScienceSeries.200936:9296.

35.JacobSG,RamaniRG.Efficientclassifierforclassificationofprognosticbreastcancerdatathroughdata
miningtechniques.ProceedingsoftheWorldCongressonEngineeringandComputerScience.2012

36.RichardsG,RaywardSmithVJ,SonksenPH,CareyS,WengC.Dataminingforindicatorsofearlymortality
inadatabaseofclinicalrecords.ArtifIntellMed.200122:215231.[PubMed]

ArticlesfromIranianJournalofBasicMedicalSciencesareprovidedherecourtesyofMashhadUniversityof
MedicalSciences

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4923467/ 9/9