Automatic diagnosis of melanoma using machine learning methods on a spectroscopic system

MISSING IMAGE

Material Information

Title:
Automatic diagnosis of melanoma using machine learning methods on a spectroscopic system
Physical Description:
Mixed Material
Language:
English
Creator:
Li, Lin
Zhang, Qizhi
Ding, Yihua
Jiang, Huabei
Theirs, Bruce H.
Wang, James Z.
Publisher:
Bio-Med Central (BMC Medical Imaging)
Publication Date:

Notes

Abstract:
Background: Early and accurate diagnosis of melanoma, the deadliest type of skin cancer, has the potential to reduce morbidity and mortality rate. However, early diagnosis of melanoma is not trivial even for experienced dermatologists, as it needs sampling and laboratory tests which can be extremely complex and subjective. The accuracy of clinical diagnosis of melanoma is also an issue especially in distinguishing between melanoma and mole. To solve these problems, this paper presents an approach that makes non-subjective judgements based on quantitative measures for automatic diagnosis of melanoma. Methods: Our approach involves image acquisition, image processing, feature extraction, and classification. 187 images (19 malignant melanoma and 168 benign lesions) were collected in a clinic by a spectroscopic device that combines single-scattered, polarized light spectroscopy with multiple-scattered, un-polarized light spectroscopy. After noise reduction and image normalization, features were extracted based on statistical measurements (i.e. mean, standard deviation, mean absolute deviation, L1 norm, and L2 norm) of image pixel intensities to characterize the pattern of melanoma. Finally, these features were fed into certain classifiers to train learning models for classification. Results: We adopted three classifiers – artificial neural network, naïve bayes, and k-nearest neighbour to evaluate our approach separately. The naive bayes classifier achieved the best performance - 89% accuracy, 89% sensitivity and 89% specificity, which was integrated with our approach in a desktop application running on the spectroscopic system for diagnosis of melanoma. Conclusions: Our work has two strengths. (1) We have used single scattered polarized light spectroscopy and multiple scattered unpolarized light spectroscopy to decipher the multilayered characteristics of human skin. (2) Our approach does not need image segmentation, as we directly probe tiny spots in the lesion skin and the image scans do not involve background skin. The desktop application for automatic diagnosis of melanoma can help dermatologists get a non-subjective second opinion for their diagnosis decision.
General Note:
Li et al. BMC Medical Imaging 2014, 14:36 http://www.biomedcentral.com/1471-2342/14/36; Pages 1-12
General Note:
doi:10.1186/1471-2342-14-36 Cite this article as: Li et al.: Automatic diagnosis of melanoma using machine learning methods on a spectroscopic system. BMC Medical Imaging 2014 14:36.

Record Information

Source Institution:
University of Florida
Holding Location:
University of Florida
Rights Management:
© 2014 Li et al.; licensee BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly credited. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.
System ID:
AA00026988:00001


This item is only available as the following downloads:


Full Text

PAGE 1

RESEARCHARTICLEOpenAccessAutomaticdiagnosisofmelanomausingmachine learningmethodsonaspectroscopicsystemLinLi1*,QizhiZhang2,YihuaDing3,HuabeiJiang2,BruceHThiers4andJamesZWang3AbstractBackground: Earlyandaccuratediagnosisofmelanoma,thedeadliesttypeofskincancer,hasthepotentialto reducemorbidityandmortalityrate.However,earlydiagnosisofmelanomaisnottrivialevenforexperienced dermatologists,asitneedssamplingandlaboratorytestswhichcanbeextremelycomplexandsubjective.The accuracyofclinicaldiagnosisofmelanomaisalsoanissueespeciallyindistinguishingbetweenmelanomaand mole.Tosolvetheseproblems,thispaperpresentsanapproachthatmakesnon-subjectivejudgementsbasedon quantitativemeasuresforautomaticdiagnosisofmelanoma. Methods: Ourapproachinvolvesimageacquisition,imageprocessing,featureextraction,andclassification.187 images(19malignantmelanomaand168benignlesions)werecollectedinaclinicbyaspectroscopicdevicethat combinessingle-scattered,polarizedlightspectroscopywithmultiple-scattered,un-polarizedlightspectroscopy. Afternoisereductionandimagenormalization,featureswereextractedbasedonstatisticalmeasurements(i.e. mean,standarddeviation,meanabsolutedeviation, L1norm,and L2norm)ofimagepixelintensitiestocharacterize thepatternofmelanoma.Finally,thesefeatureswerefedintocertainclassifierstotrainlearningmodelsfor classification. Results: Weadoptedthreeclassifiers – artificialneuralnetwork,navebayes,andk-nearestneighbourtoevaluate ourapproachseparately.Thenaivebayesclassifierachievedthebestperformance-89%accuracy,89%sensitivity and89%specificity,whichwasintegratedwithourapproachinadesktopapplicationrunningonthespectroscopic systemfordiagnosisofmelanoma. Conclusions: Ourworkhastwostrengths.(1)Wehaveusedsinglescatteredpolarizedlightspectroscopyand multiplescatteredunpolarizedlightspectroscopytodecipherthemultilayeredcharacteristicsofhumanskin.(2)Our approachdoesnotneedimagesegmentation,aswedirectlyprobetinyspotsinthelesionskinandtheimage scansdonotinvolvebackgroundskin.Thedesktopapplicationforautomaticdiagnosisofmelanomacanhelp dermatologistsgetanon-subjectivesecondopinionfortheirdiagnosisdecision.BackgroundMelanomaisalethalformofskincancer,withan estimatedmortalityrateof14%worldwide[1].The AmericanCancerSocietyreported76,690newcasesof melanomaintheUnitedStatesin2013,with9,480estimateddeaths,accordingtorecentannualcancerfacts andfigures[2].Theglobalcancerstatistics[3]also showsthattheincidenceandmortalityratesofmelanomaareinrisingtrend.Fortunately,melanomamaybe treatedsuccessfullywitha10-yearsurvivalratebetween 90and97%yetthecurabilitydependsonitsearlydetectionandexcisionwhenthetumorisstillsmallandthin. Therefore,earlyandaccuratediagnosisofmelanomais particularlyimportant. Withtheuseofdermoscopy[4]andseveralclinicalalgorithmssuchastheABCDrule[5],the7-pointchecklist[6],andtheMenziesmethod[7],thediagnosis accuracyofmelanomahasbeenhigherthanthesimple naked-eyeexamination[8].However,clinicaldiagnosisis inherentlysubjectiveandcomplex,thustheaccuracyis highlydependingonexperiencesofdermatologists, whichisestimatedtobeabout75 – 85%[9]. Toreducesubjectivityandcomplexityofclinicaldiagnosis,itisdesirabletoconductresearchonquantitative *Correspondence: lil@seattleu.edu1DepartmentofComputerScience&SoftwareEngineering,Seattle University,Seattle,WA98122,USA Fulllistofauthorinformationisavailableattheendofthearticle 2014Lietal.;licenseeBioMedCentralLtd.ThisisanOpenAccessarticledistributedunderthetermsoftheCreative CommonsAttributionLicense(http://creativecommons.org/licenses/by/4.0),whichpermitsunrestricteduse,distribution,and reproductioninanymedium,providedtheoriginalworkisproperlycredited.TheCreativeCommonsPublicDomain Dedicationwaiver(http://creativecommons.org/publicdomain/zero/1.0/)appliestothedatamadeavailableinthisarticle, unlessotherwisestated.Li etal.BMCMedicalImaging 2014, 14 :36 http://www.biomedcentral.com/1471-2342/14/36

PAGE 2

approachesforautomateddetectionofmelanoma.In thelasttwodecades,alargeamountofcomputer-aidedapproacheshavebeendevelopedfordiagnosisofmelanoma. Forexample,A.G.Manousakietal.[10]haveproposedan approachthatincorporatesparametersofgeometry,color andcolortextureasindependentcovariatesfordiscriminatingmelanomafrommelanocyticnevi.IntheworkofH. Gansteretal.[11],amelanomarecognitionsystemthat involvesimageprocessing,segmentation,featurecalculationandselection,aswellask-NNclassificationhasbeen presented.Thework[12]byJ.F.Alcnetal.alsopresents anautomaticimagingsystemthatcombinestheoutcome oftheimageclassificationwithcontextknowledgesuchas skintype,age,gendertoaddconfidencetotheclassification.R.GarnaviandM.Aldeenhavepresentedanapproach[13]thatusesborder-andwavelet-basedtexture andfourdifferentclassifiersindiagnosisofmelanoma. Additionally,manycomputerapplicationshavebeendevelopedformelanomadiagnosis.SomeoftheseapplicationsincludeSolarScan[14],theDANAOSexpertsystem [15],DermoGenius-Ultra[16],andMelaFind[17],etc. Thesemethodshaveachievedgoodclassificationaccuracy, buteithermostofthemobtainedimagesbyhand-held cameras,whichthusneededimagesegmentationtoseparatethelesionfromthebackground,ortheywerelargely basedonthelightintensityspectra,derivedabsorption andscatteringspectra,whichhavebeenshowntobe highlysensitivetotheabnormalchangesintissues.Moreover,mostofthesemethodstreattheskintissueasauniformorhomogeneousmedium. Nevertheless,theskintissueisinhomogeneouswith multilayeredstructure.Itconsistsoftwoprimarylayers thatincludeabottomlayer – thedermis,andatop layer – theepidermis.Underthedermisissubcutaneous tissue,orhypodermis,whichconsistsofconnectivetissue,fibroblasts,andfatcells.Theepidermisconsistsof cellscalledkeratinocytes,whichdevelopinthebasal layeroftheskinatthebottomoftheepidermis.Asthese keratinocytesmigratetothesurface,theyflatten,cornify, andharden,thuslysealingtheskin[18].Theepidermis alsoincludesmelanocyteslocatednearthebaseofthe epidermis.Thesecellssecretethepigmentmelanin whichprotectstheskinfromultravioletradiation.The dermiscanvaryinthicknessfrom1to4mm,andisprimarilycomprisedoffibrousconnectivetissuesuchas collagen.Thedermisalsocontainsthehairfollicles,sweat glands,bloodvessels,andnerves.Theskinisanextremely complexopticalstru cturebecauseitconsistsofmultiple scatterswithmanyshapesandsizes.Aslightpropagates throughtheskin,itisscatteredandabsorbeddifferentlyin eachlayer.Theabsorptionisalsocomplexconsisting ofcontributionsprimarilyfrombloodandmelanin.The multilayeredstructurefurther increasesthecomplexity.Becauseskincanceroftenbeginsintheepidermallayerand invadesdeepertissueovertime,theinformationobtained withthecurrentmultiple-scat teredlight-basedmethodsis averagedoutanddoesnotreflecttheaccuratemorphology ofthespecificdiseasedlayer,a lthoughencouragingresults forskincancerdetectionhavebeenrecentlyobtainedusing multiple-scatteredlight-onl ymethodsinconjunctionwith classificationalgorithms[19,20]. Twotechniqueshavebeendevelopedforuseindecipheringthemultilayeredcharacteristicsofhumanskin: multilayermodel-basedMultiple-ScatteredLightSpectroscopy(MSLS)[21-25]andsingle-scattered,Polarized LightSpectroscopy(PLS)[26-31].ThemultilayermodelbasedMSLSconsiderstherealisticmultilayeredstructure oftheskininthemathematicalmodeloflightpropagation,thusisagoodchoiceforskinstudies.However,the accurateuseofsuchimprovedspectroscopyrequires knowingtheexactthicknessofeachlayerofskinwhichis difficulttoobtain.Ontheotherhand,PLSisabletophysicallydiscriminatebetweenmultiplelayersoftissue,which tendstosimplifythelightpropagationmodelingbecause single-scatteredlightmaintainsitsoriginalpolarization. Nevertheless,comparedtoMSLS,PLShasarelativelylow signal-to-noiseratio. TotaketheadvantagesofbothMSLSandPLS,inthis study,wecapturedskinimageswithacombinationof MSLSandPLStechnologies,anddevelopedanautomated methodfordiagnosisofmelanomaviapatternclassificationoftheskinimages.ThecombinedMSLSandPLS technologiesprovideunprecedentedtissuefunctionalinformationandcellularstructuresandaccuratelyreflect morphologiesinspecificdiseasedlayersofskin.Usinga numberofskinscanscollectedfromaclinicbyourspectroscopicsystemcombiningsingleandmultiple-scattered lightmeasurements,wefirstidentifiedthepixel-by-pixel intensitydifferencesbetweenthegroupofmelanomaand thegroupofbenignskins.Wethenselectedstatistical measurementsofpixelintensitythatpresentedcharacteristicsofmelanomaasfeaturesforclassification.Next,classificationwascarriedoutbyusingtheselectedfeatures. Weevaluatedourapproachusingartificialneuralnetwork (ANN),navebayes(NB),andk-nearestneighbour(k-NN) separately.Theapproachcouldachieve89%sensitivity, 89%specificity,and89%accuracybyusingNB.Basedon theproposedmethod,adesktopapplicationthatrunson thespectroscopicsystemhasbeendevelopedtointegrate imageacquisition,imageprocessing,featureextraction, andclassificationintoone-stopservice.Dermatologists canusethisapplicationtoconductinstantaneousdiagnosisofmelanomatogetasecondopiniontotheirclinical decisions.MethodsDatausedinthisworkwerecollectedinaclinicbyaspectroscopicsystemwithcombinedsingleandmultiple-Li etal.BMCMedicalImaging 2014, 14 :36 Page2of12 http://www.biomedcentral.com/1471-2342/14/36

PAGE 3

scatteredlightmeasurements.Thestudywasapprovedby theInstitutionalReviewBoard(IRB)forHumanResearch inMedicalUniversityofSouthCarolina.Patientshave consentedtoparticipateinthestudyandthestudyhasnot usedpatientidentifiableinformation.Thespectroscopic systemwascomposedofaPCwithamonitor,aCCDcamera,aCCDcontroller,aspectrometer,alightsource,opticalfibers,apolarizer,gradient-indexlens,andaprobe. Figure1showsthepicturesofthesystem,includingthe frontpanelremovedtoshowtheinside(1a),thelayoutof ac bLight Source CCD Controller CCD Spectrometer Figure1 Picturesofthespectroscopicsystem.(a) frontviewofthesystem; (b) instrumentsonthe2ndlayerofthecart; (c) systemschema. Li etal.BMCMedicalImaging 2014, 14 :36 Page3of12 http://www.biomedcentral.com/1471-2342/14/36

PAGE 4

instrumentsinsidethesystem(1b),andthesystemschema (1c).Inthedesignoftheopticalprobe,opticalfiberswith gradient-indexlensbuiltatthedistalendwereemployed insteadofattachingaseparatedlenssystemwiththe sourceanddetectorfibers.Figure2isaschemaofthe probe.ParticipantsThedatasetinthisstudyincluded187samplesfrom79 participants,assomeparticipantsprovidedacoupleof sampleslocatedindifferentareasofskin.Theparticipantswereagedfrom22to79.The187samplesconsistedof19melanoma[16%female,84%male,mean age=61.32(12.88)years]and168benignscans(i.e.benignnevus)[51%female,49%male,meanage=34.63 (11.40)years].Thedatasamplesweresentforpathologyandthelabels(i.e.benignandmelanomas)were confirmedcorrect.ImageacquisitionandpreprocessingForeachskinsample,ourspectroscopicsystemgathered imagesfromthreespots.Weperformedtwopairsof scansoneachlesionskinarea.Eachpairincludedone ‘ P ’ scanning(usingparallellight)andone ‘ V ’ scanning (usingverticallight).Ifthelesionorabnormalareawas largerthantheprobetip,wemovedtheprobetip slightlytotakethetwopairsontwodifferentspotsof thelesionarea.Ifthelesionareawastoosmall,onlyone spotwaschosenandscannedtwicetoobtaintwopairs ofscans.Wealsotookonepairofscansonthenormal skinareanearbyforcomparison.Therefore,sixspectral imageswerecollectedforeachsample.Figure3depicts theswitchontheprobeforselectingPscanorVscan (left)andtheskinlesionimagingbyatechnician(right). Thespectralimagesareinabinaryformat.Eachspectral imagecontains32512pixels.Thevalueofeachpixelis theintensityobtainedbytheCCDcamera.Considering 10c m ab Figure2 Aschemaoftheprobe. Athree-dimensionalview (a) andatwo-dimensionalschema (b) oftheopticalprobe. Li etal.BMCMedicalImaging 2014, 14 :36 Page4of12 http://www.biomedcentral.com/1471-2342/14/36

PAGE 5

thattheimagesmaycontainnoisesduetoenvironmentaleffectsduringcollection,weperformednoisereductionusing themedianfilter.Afternoiser eduction,pixelintensities werenormalizedwithmin-maxnormalizationmethod,as showninformula(1), I0 i ; j ; k Ii ; j ; k min max min max0 min0 min0 1 where I( i j k )representstheoriginalpixelintensityatposition(i,j)ofimagek(1 i 32,1 j 512),and I '( i j k )is thecorrespondingnormalizedintensity. max and min arethelargestandsmallestintensitiesintheoriginal imagerespectively; max '= 1 and min '= 0 TakePscanasanexample,Figure4presentsaPscanof amelanomalesion,wherewepickedline15tovisualize thespectraaspixelsinthislinehavethemostsignificant intensities.Vscansalsodemonstratesimilarspectra.FeatureextractionAsthelightgeneratedbyourspectroscopicsystem propagatesthroughthebenignandmalignantskin,itis The Switch Figure3 Demonstrationofimageacquisition. Left:theswitchontheprobefromwhichtoperformeither ‘ P ’ scanor ‘ V ’ scan.Right:weplace thetipoftheprobeontheskininaperpendicularwayandletthelightspotlocateinthelesionareaoftheskin. Figure4 Pscanofamelanomaskinlesion. Li etal.BMCMedicalImaging 2014, 14 :36 Page5of12 http://www.biomedcentral.com/1471-2342/14/36

PAGE 6

scatteredandabsorbeddifferently,whichresultstothe intensitydifferencesinthecollectedspectralimages. TakePscanasanexample,Figure5demonstrates theintensitydistributionsofallbenign(left)andall malignantmelanomaskinlesions(right).Itiseasyto findthatthepixelintensitiesofmalignantmelanoma skinimagesspreadoutintheintervalfrom0to1, withmostpixelshavinghighintensity,whereasthe majorityofintensityvaluesofbenignskinimagesfall intheintervalfrom0.2to0.4.Wealsofoundthat theintensitydistributionofVscanhasdemonstrated similardifferencesbetweenallbenignandallmelanomaskinlesions. Sincetheintensitydistributionsofthebenignand melanomaskinimagesaregreatlydifferentasshownin Figure5,wethereforeconsiderthepixelintensitiesas thekeyfeaturetodistinguishbenignskinfrommelanoma.Tofullydescribethepixelintensitydistribution ineachimage,weadoptedfivestatisticalmeasuresto quantifyeachscan.Thefivestatisticalvariablesaremean ,standarddeviation ,meanabsolutedeviationMAD, L1norm I i ; j 1,and L2norm I i ; j 2,where I i ; j isthecorrectedintensityofpixel(i,j).WecalculatedthestatisticalvaluesseparatelyforPscansandVscans.That istosay,eachsamplehas10statisticalmeasures(i.e. 5forPscansand5forVscans).Additionally,asfor thepixelintensityofPscansorVscans,ineachsample,wetookboth2scansoflesionareaand1scan ofnormalareanearbyintoconsideration.Thecorrected intensityofpixel(i,j)inPscansofsamplek, I i ; j ; k ; p ,was computedbyformula(2), I i ; j ; k ; p I0 i ; j ; k ; p 1 I0 i ; j ; k ; p 2 = 2 I0 i ; j ; k ; p 3 2 where I '( i j k p 1)isthenormalizedintensityofpixel(i,j) inp1scan-Pscanofthe1stselectedspotinthelesion area, I '( i j k p 2)isthenormalizedintensityofpixel(i,j)in p2scan – Pscanofthe2ndselectedspotinthelesionarea, and I '( i j k p 3)isthenormalizedintensityofpixel(i,j)inp3 scan – Pscanofaspotselectedinthenormalareanearby. ThesamecomputationwasconductedonVscans.Measuringthepixelintensitywithformula(2)cancapturethe differenceoflesionskinandnormalskinregardingthe samesubject,whichmakesmoresensethanusingintensitiesoflesionskinalone,asintensityvaluesmaybeinfluencedbyfactorslikeskincolor,age,gender,etc. Weusethefollowingformulas(3) – (7)tocalculate thestatisticalvaluesinPscansofsamplek.Intheformulas,m=32andn=512asthereare32512pixels perscan.Theindexiisintherange[1,32],andthe indexjisintherange[1,512].Similarly,theformulas areappliedtocalculatethestatisticalmeasurementsof Vscansbyreplacing I i ; j ; k ; p with I i ; j ; k ; v 1 mn Xm i 1Xn j 1I i ; j ; k ; p 3 Figure5 IntensitydistributionofPscansofallbenignandallmelanomaskinlesions. Left:Intensitydistributionofbenignskinlesions; Right:intensitydistributionofmalignantmelanomaskinlesions. Li etal.BMCMedicalImaging 2014, 14 :36 Page6of12 http://www.biomedcentral.com/1471-2342/14/36

PAGE 7

1 mn Xm i 1Xn j 1I i ; j ; k ; p 2v u u t 4 MAD 1 mn Xm i 1Xn j 1I i ; j ; k ; p 5 I i ; j ; k ; p 1 Xm i 1Xn j 1I i ; j ; k ; p 6 I i ; j ; k ; p 2 Xm i 1Xn j 1I i ; j ; k ; p 2s 7 ClassificationAfterextractingthediagnosticfeatures,wetraineddifferentclassifiers(ANN,NB,andk-NN)todistinguish melanomafrombenignskin.Givenatrainingsampleset withnsamples D xi; yi fgn i 1where xiisasample and yiistheassociatedclasslabel,wefocusedonthe binaryclassificationproblem,i.e., yiisfromalabelspace {1}where+1denotesthecancerclassand 1denotes thenon-cancerclass.Thecancerclasswascomposedof thesamplesofmelanoma,whereasthenon-cancerclass consistedofthesamplesfrombenignskin.Eachsample inthetrainingsethas10featurestobefedintotheclassifier.ThetoolWEKA[32]wasusedfortrainingand testing.WithANN,amulti-layernetworkthatusedback propagationwasbuilt.Theinputlayerhad10input units,whichwerethe10selectedfeatures.Theoutput layerhad2units,representingtwoclasses – benignand melanoma.Thehiddenlayerwasinitiallysettohave6 unitsintrainingasnormallythenumberofhiddenunits issettothehalfofthesumofinputunits(10inour study)andoutputunits(2inourstudy).Wekeptthe defaultparameters(i.e.learningRate=0.3,momentum= 0.2,seed=0,trainingTime=500,validationThreshold= 20)inWEKAforANN.AsforNB,theclassifierusedestimatorclasses.Numericestimatorprecisionvalueswere chosenbasedonanalysisofthetrainingdata.Theclassifierusedanormaldistributionfornumericattributes. Wekeptthedefaultparameters(i.e.useKernelEstimator= false,useSupervisedDiscretization=false)inWEKAfor NBclassifier.Thechoiceofkink-NNaffectstheperformanceofthisclassifier.Inourstudy,weused3-NN, whichcombinesrobustnesstonoiseandcostslesstime forclassificationthanusingalargerk[33].Forotherparametersofk-NNinWEKA,wesetcrossValidatetofalse, didnotusedistanceweighting,andusedthebruteforce searchalgorithmfornearestneighboursearch.ResultsPatternofmelanomaMelanomaandbenigngroupcomparisonacrossthe10 featuresdemonstratedthecharacteristicofmelanoma. Theeffectofmelanomacanbeseeninthedistribution ofthepixelintensitydifferencesofabnormalskinand normalskinofindividualsubjects.Figure6demonstratesthattheinfluenceonmelanomaintermsofpixel intensitycanbeobservedincolorscalebycomparinga melanomacasetoabenigncase.ClassificationaccuracyInourexperiment,the187caseswererandomlydivided intoatrainingsetof60casesandatestsetof127cases byrandomsamplingwithoutreplacement.Weconductedtheexperimentusingthreeclassifiers(i.e.ANN, NB,and3-NN)separately.Theexperimentwasrepeated 25timesforeachclassifier.Inaddition,weevaluatedthe performanceofclassificationusingPscanalone(i.e.5 statisticalmeasuresofPscanforclassification),Vscan alone(5statisticalmeasuresofVscanforclassification), aswellasthecombinedPscanandVscan(10features forclassification).Table1demonstratestheperformance ofusingNBwiththecombinedPscanandVscan,of which,onaverage,theaccuracy,specificity,andsensitivitywere89%,89%,89%respectively.Theprobabilityof errorinNBwas0.16.Theaverageaccuracy,specificity andsensitivityofusingANNwiththecombinedPscan andVscanwere88%,93%,and49%respectively.3-NN withthecombinedPscanandVscandemonstrated88% accuracy,92%specificity,and47%sensitivityonaverage. TheprobabilityoferrorinANNand3-NNwas0.24and 0.26respectively.Table2showstheaverageperformance oftheclassifiersusingPscanandVscan,usingPscan aloneorusingVscanalone.Duetospacelimitation,the performancedetailsofeachrunofexperimentwithdifferentclassifiers(i.e.ANN,3-NN)anddifferentsetsof features(i.e.PscanaloneorVscanalone)wereskipped. Inaddition,weperformedtheexperimentwithadifferentcombinationofdatabyrandomlydividingthedata intoatrainingsetof30casesandatestsetof157cases. TheaverageperformanceofeachclassifierusingPscan andVscan,usingPscanaloneorusingVscanaloneis showninTable3.DesktopapplicationForautomaticdetectionofmelanoma,webuiltadesktopapplicationwhichrunsonthemachine(Figure1a) thatconnectstothespectroscopicdevice.Clinicianscan launchtheapplicationtocollectskinscansusingthespectroscopicdevice,andfollowthewizardtomaketheapplicationautomaticallyconductimageprocessing,feature extraction,andclassificationsoastoachieveinstantaneous diagnosisofmelanoma.TheNBclassifierwasintegratedinLi etal.BMCMedicalImaging 2014, 14 :36 Page7of12 http://www.biomedcentral.com/1471-2342/14/36

PAGE 8

thisapplication.Figure7presentsascreenshotoftheautomatedtool.Wekeeptheinterfacesimplefordermatologiststouse.Dermatologistsselectimages1P,2P,3Pand 1V,2V,3Vthathavebeencollectedbythespectroscopic deviceforasubject,byclickingonthe6buttonsonthe veryleftpaneloftheinterface.Thestatusboxshows whetheranimageisloadedsuccessfully.Userscaninput andsavecommentsforthediagnosisinthe “ Comment ” area.Onceimagesareallloadedsuccessfully,usersjust needtoclickonthe “ Diagnose ” buttontostartthediagnosis,whichlaunchestheexecutionofback-endprograms forfeatureextractionandclassification.Oncethediagnosis isdone,amessageboxpopsupshowingthediagnosisresult – eithermelanomaorbenign.DiscussionComputer-aideddiagnosisofmelanomagenerallyinvolvessixsteps:imageacquisition,imageprocessing, a b c e d Figure6 Patternofmelanoma.(a) Thecolorbarrepresentingthecolorschemeusedforpixelintensity; (b) Thecolormapofpixelintensity calculatedbyformula(2)inPscanofabenigncase; (c) Thecolormapofpixelintensitycalculatedbyformula(2)inVscanofthesamebenign caseasthatwasusedin (b) ; (d) Thecolormapofpixelintensitycomputedbyformula(2)inPscanofamelanomacase; (e) Thecolormapof pixelintensitycomputedbyformula(2)inVscanofthesamemelanomacaseasthatwasusedin (d) Li etal.BMCMedicalImaging 2014, 14 :36 Page8of12 http://www.biomedcentral.com/1471-2342/14/36

PAGE 9

imagesegmentation,featureextraction,featureselection, andclassification.Inourwork,wegatheredimageswith combinedsingleandmultiple-scatteredlightmeasurements,fordecipheringthemultilayeredcharacteristics ofhumanskin.Thisisonescientificcontributionofour work.Theotherstrengthofourworkisthatscansoflesionareadonotinvolvebackgroundskinaswedirectly probetinyspotsinthelesionskin.Imagesegmentation wasskippedinourwork.Theobjectiveofsegmentation indiagnosisofmelanomaistodetecttheborderofthe lesionsoastoseparatethelesion – regionofinterest fromthebackgroundskin.Sinceweusetheprobeonly withintheregionofinterest(ROI),ourapproachdoes notneedimagesegmentation.Anapproachwithoutsegmentationreducestheriskofdiagnosiserrorthatmay beintroducedbyimpropersegmentationandreduces thecostandtimecausedbysegmentation.Inaddition, wedidnotperformfeatureselectionbecausethe10statisticalparametersextractedforeachsamplewereallneeded forclassification. Classificationmethodsthathavebeenappliedto computer-aideddiagnosisofmelanomaincludediscriminantanalysis[34,35],ANN[36,37],k-NN[11],support vectormachine(SVM)[38],anddecisiontrees,etc.To evaluateourapproach,weperformedclassificationusing threedifferentclassifiers:ANN,NBandk-NN,withthe combinationofPscanandVscan,Pscanalone,orV scanalone.ComparedwithANNandk-NN,NBcould achievethebestaccuracyandsensitivity.AlthoughANN andk-NNachievedbetterspecificitythanNB,thesensitivitywasnotgood.WithNB,theexperimentusingthe combinationofPscanandVscanachievedbetteraccuracyandspecificitythanusingPscanorVscanalone, whiletheexperimentwithVscanalonepresentedthe bestsensitivity.WepickedtheclassifierNBusingthe combinedPscanandVscanwhichachievedthehighest accuracyandintegrateditinourapplicationtoprovide automateddiagnosisofmelanoma. Table1PerformanceofNBusingPscanandVscan together(60trainingsamples,127testingsamples)SensitivitySpecificityAccuracy Run#11008687.4 Run#276.992.190.6 Run#310091.292.1 Run#484.691.290.6 Run#576.988.687.4 Run#692.387.788.2 Run#710087.789 Run#810088.689.8 Run#910087.789 Run#1010089.590.6 Run#1169.289.587.4 Run#1276.990.489 Run#1392.389.589.8 Run#1492.390.490.6 Run#1576.988.687.4 Run#1676.994.792.9 Run#1784.689.589 Run#1810087.789 Run#1992.387.788.2 Run#2084.687.787.4 Run#2192.390.490.6 Run#2210091.292.1 Run#2310085.186.6 Run#2410086.888.2 Run#2546.29388.2 Average88.689.389.2Allvaluesarein%. Table2Theaverageperformanceoftheclassifiersusing PscanandVscan,usingPscanaloneorusingVscan alone(60trainingsamples,127testingsamples)SensitivitySpecificityAccuracy NB(usingP&Vscan)88.689.389.2 NB(usingPscan)83.777.878.4 NB(usingVscan)95.785.186.2 ANN(usingP&Vscan)49.292.888.4 ANN(usingPscan)13.296.688.2 ANN(usingVscan)19.796.989.0 3NN(usingP&Vscan)46.892.187.5 3NN(usingPscan)24.693.386.3 3NN(usingVscan)32.095.689.0Allvaluesarein%. Table3Theaverageperformanceoftheclassifiersusing PscanandVscan,usingPscanaloneorusingVscan alone(30trainingsamples,157testingsamples)SensitivitySpecificityAccuracy NB(usingP&Vscan)73.090.889.0 NB(usingPscan)74.082.281.4 NB(usingVscan)81.886.085.6 ANN(usingP&Vscan)66.589.887.4 ANN(usingPscan)30.094.287.8 ANN(usingVscan)49.092.888.3 3NN(usingP&Vscan)70.389.787.7 3NN(usingPscan)28.892.686.1 3NN(usingVscan)51.092.388.1Allvaluesarein%.Li etal.BMCMedicalImaging 2014, 14 :36 Page9of12 http://www.biomedcentral.com/1471-2342/14/36

PAGE 10

Therearesomelimitationsinourwork.First,although ourapproachcouldachieve89%sensitivity,89%specificity,and89%accuracywithNB,thesensitivitywith ANNandk-NNwerenotgoodenough.Thesmallnumberofmelanomasandmuchlargersetofbenignsamplesmightmakethebenignsamplesdominatethe classification.Inaddition,becauseofthelimitednumber ofmelanomasinourexperiment,itislikelythattheperformanceofourapproachisspecifictothedatasets usedinthispreliminarystudy.Itisunclearhowthe performancevarieswhenourapproachisappliedtoa largeror/andnewdataset.Follow-upstudiesthatincorporatelargersizeofmelanomamayhavethemost reliabilityandprovidegreaterconfidenceforthediagnosisthantheclassifierdevelopedinthispreliminary study.Second,somefactorslikethechangesinthedevice,thepersonwhousesthedevice,etc.mighthave effectontheoutcome.Particularly,variationinskin color,age,genderinthesamplesetmayinfluencethe diagnosisresult.ConclusionThispaperpresentsacomputer-aidedapproachforautomaticandaccuratediagnosisofmelanoma.Weevaluated ourapproachwiththreeclassifiers-ANN,NB,andk-NN, andourapproachcouldachieve89%sensitivity,89%specificity,and89%accuracywithNB.Inthefuture,wewill dofollow-upstudybycollectingmorerealdata,especially melanomacases,tofurtherevaluateourapproach.Abbreviations ANN: Artificialneuralnetwork;K-NN:K-Nearestneighbour;MAD:Mean absolutedeviation;MSLS:Multiple-scatteredlightspectroscopy;NB:Nave bayes;PLS:Polarizedlightspectroscopy;SVM:Supportvectormachine. Competinginterests Theauthorsdeclarethattheyhavenocompetinginterests. Authors ’ contributions LLimplementedtheimageacquisition,performedimageprocessing, designedtheapproachforfeatureextractionandclassification,carriedout theevaluation,developedthedesktopapplication,anddraftedthe manuscript.QZZdevelopedthespectroscopicdevice.YHDconductedthe extractionoffeatures.HBJdesignedthespectroscopicapproachforimage Figure7 Ascreenshotofthedesktopapplicationforautomateddiagnosisofmelanoma. Li etal.BMCMedicalImaging 2014, 14 :36 Page10of12 http://www.biomedcentral.com/1471-2342/14/36

PAGE 11

acquisition.BHTcollectedrealdatasetinthisstudy.JZWledthedesignof thestudy.Allauthorsreadandapprovedthefinalmanuscript. Acknowledgements ThisworkispartiallysupportedbyNationalScienceFoundation[DBI-0960586 andDBI-0960443];NationalInstituteofHealth[1R15CA131808-01and1R01 HD069374-01A1]. Authordetails1DepartmentofComputerScience&SoftwareEngineering,Seattle University,Seattle,WA98122,USA.2DepartmentofBiomedicalEngineering, UniversityofFlorida,Gainesville,FL32611,USA.3SchoolofComputing, ClemsonUniversity,Clemson,SC29634,USA.4DepartmentofDermatology, MedicalUniversityofSouthCarolina,Charleston,SC29425,USA. Received:3June2014Accepted:3October2014 Published:13October2014 References1.JemalA,SiegelR,WardE,MurrayT,XuJ,ThunMJ: Cancerstatistics. CACancerJClin 2007, 57: 43 – 66. 2. Cancerfactsandfigures. 2013[http://www.cancer.org/research/ cancerfactsfigures/cancerfactsfigures/cancer-facts-figures-2013] 3.JemalA,BrayF,CenterM,FerlayJ,WardE,FormanD: Globalcancer statistics. CACancerJClin 2011, 61: 69 – 90. 4.BraunR,RabinovitzH,OlivieroM,KopfA,SauratJ: Dermoscopyof pigmentedlesions. JAmAcadDermatol 2005, 52 (1):109 – 121. 5.StolzW,RiemannA,CognettaA: ABCDruleofdermatoscopy:Anew practicalmethodforearlyrecognitionofmalignantmelanoma. EurJDermatol 1994, 4: 521 – 527. 6.ArgenzianoG,FabbrociniG,CarliP,GiorgiV,SammarcoE,DelfinoM: Epiluminescencemicroscopyforthediagnosisofdoubtfulmelanocytic skinlesions:ComparisonoftheABCDruleofdermatoscopyanda new7-pointchecklistbasedonpatternanalysis. ArchDerm 1998, 134: 1563 – 1570. 7.MenziesSW,IngvarC,McCarthyWH: Asensitivityandspecificityanalysis ofthesurfacemicroscopyfeaturesofinvasivemelanoma. MelanomaRes 1996, 6 (1):55 – 62. 8.GellerAC,SwetterSM,BrooksK,DemierreM,YarochAL: Screening,early detection,andtrendsformelanoma:currentstatus(2000 – 2006)and futuredirections. JAmAcadDermatol 2007, 57: 555 – 572. 9.ArgenzianoG,SoyerHP,ChimentiS,TalaminiR,CoronaR,SeraF,BinderM, CerroniL,DeRosaG,FerraraG,Hofmann-WellenhofR,LandthalerM, MenziesSW,PehambergerH,PiccoloD,RabinovitzHS,SchiffnerR, StaibanoS,StolzW,BartenjevI,BlumA,BraunR,CaboH,CarliP,De GiorgiV,FlemingMG,GrichnikJM,GrinCM,HalpernAC,JohrR, etal : Dermoscopyofpigmentedskinlesions:resultsofaconsensusmeetingvia theInternet. JAmAcadDermatol 2003, 48: 679 – 693. 10.ManousakiAG,ManiosAG,TsompanakiEI,PanayiotidesJG,TsiftsisDD, KostakiAK,ToscaAD: Asimpledigitalimageprocessingsystemtoaidin melanomadiagnosisinaneverydaymelanocyticskinlesionunit:a preliminaryreport. IntJDermatol 2006, 45 (4):402 – 410. 11.GansterH,PinzA,RohrerR,WildlingE,BinderM,KittlerH: Automated melanomarecognition. IEEETransMedImaging 2001, 20 (3):233 –239. 12.AlcnJF,CiuhuC,KateW,HeinrichA,UzunbajakavaN,KrekelsG,SiemD, deHaanG: Automaticimagingsystemwithdecisionsupportfor inspectionofpigmentedskinlesionsandmelanomadiagnosis. IEEEJ SelectTopSignProcess 2009, 3 (1):14 – 25. 13.GarnaviR,AldeenM: Computer-aideddiagnosisofmelanomausing border-andwavelet-basedtextureanalysis. IEEETransInfTechnolBiomed 2012, 16 (6):1239 – 1251. 14.MenziesSW,BischofL,TalbotH,GutenevA,AvramidisM,WongL, LoSK,MackellarG,SkladnevV,McCarthyW,KellyJ,CranneyB,Lye P,RabinovitzH,OlivieroM,BlumA,VarolA,De'AmbrosisB,McCleod R,KogaH,GrinC,BraunR,JohrR: TheperformanceofSolarScan: anautomateddermoscopyimageanalysisinstrumentfor thediagnosisofprimarymelanoma. ArchDermatol 2005, 141: 1388 – 1396. 15.HoffmannK,GambichlerT,RickA: Diagnosticandneuralanalysisofskin cancer(danaos).Amulticentrestudyforcollectionandcomputer-aided analysisofdatafrompigmentedskinlesionsusingdigitaldermoscopy. BrJDermatol 2003, 149: 801 – 809. 16.JamoraMJ,WainwrightBD,MeehanSA,BystrynJC: Improvedidentification ofpotentiallydangerouspigmentedskinlesionsbycomputerizedimage analysis. ArchDerm 2003, 139: 195 – 198. 17.MonheitG,CognettaAB,FerrisL,RabinovitzH,GrossK,MartiniM,Grichnik JM,MihmM,PrietoVG,GoogeP,KingR,ToledanoA,KabelevN,WojtonM, Gutkowicz-KrusinD: TheperformanceofMelaFind:aprospectivemulticenter study. ArchDermatol 2011, 147 (2):188 – 194. 18.McCanceKL,HuetherSE,BrashersVL,RoteNS: Pathophysiology:TheBiologic BasisforDiseaseinAdultsandChildren. 6thedition.MarylandHeights, Missouri:MosbyInc.;2009. 19.Garcia-UribeA,Keht arnavazN,MarquezG,PrietoV,DuvicM,Wang LV: Skincancerdetectionbyspectroscopicoblique-incidence reflectometry:classificationandphysiologicalorigins. ApplOpt 2004, 43: 2643 – 2650. 20.ElbaumM,KopfA,RabinovitzH,LangleyR,KaminoH: Automatic differentiationofmelanomafrommelanocyticneviwithmultispectral digitaldermoscopy:Afeasibilitystudy. JAmAcadDermatol 2001, 44: 207 – 218. 21.PhamT,SpottT,SvaasandL,TrombergB: Quantifyingthepropertiesof two-layerturbidmediawithfrequency-domaindiffusereflectance. ApplOpt 2000, 39: 4733 – 4745. 22.HielscherA,LiuH,ChanceB,TittelF,JacquesS: Time-resolvedphoton emissionfromlayeredturbidmedia. ApplOpt1996, 35: 719 – 728. 23.KienleA,PattersonM,DognitzN,BaysR,WagnieresH: Noninvasive determinationoftheopticalpropertiesoftwolayeredturbidmedia. ApplOpt 1998, 37: 779 – 791. 24.FranceschiniM,FantiniS,PaunescuL,MaierJ,GrattonE: Influenceofa superficiallayerinthequantitativespectroscopicstudyofstrongly scatteringmedia. ApplOpt 1998, 37: 7447 – 7458. 25.MartelliF,BiancoS,ZaccantiG,PifferiA,TorricelliA,BassiA,TaroniF, CubedduR: Phantomvalidationandinvivoapplicationofan inversionprocedureforretrievingt heopticalpropertiesofdiffusive layeredmediafromtime-resolvedreflectancemeasurements. Opt. Lett. 2004, 29: 2037 – 2039. 26.BackmanV,GurjarR,BadizadeganK,ItzkanI,DasariRR,PerelmanLT, FeldMS: Polarizedlightscatteringspectroscopyforquantitative measurementofepithelialcellularstructuresinsitu. IEEEJSelectTop QuantElectron 1999, 5: 1019 – 1026. 27.MourantJR,JohnsonTM,FreyerJP: Characterizingmammaliancellsand cellphantomsbypolarizedbackscatteringfiber-opticmeasurements. ApplOpt 2001, 40: 5114 – 5123. 28.MourantJR,JohnsonTM,CarpenterS,GuerraA,AidaT,FreyerJP: Polarizedangulardependentspectroscopyofepithelialcellsand epithelialcellnucleitodeterminethesizescaleofscattering structures. JBiomedOpt 2002, 7: 378 – 387. 29.JacquesS,RomanJ,LeeK: Imagingsuperficialtissueswithpolarizedlight. LasersSurgMed 2000, 26: 119 – 129. 30.MorganS,RidgwayM: Polarizationpropertiesoflightbackscatteredfrom atwolayerscatteringmedium. OptExpress 2000, 7: 395 – 402. 31.GurjarRS,BackmanV,PerelmanLT,GeorgakoudiI,BadizadeganK,ItzkanI, DasariRR,FeldMS: Imaginghumanepithelialpropertieswithpolarized lightscatteringspectroscopy. NatMed 2001, 7: 1245 – 1248. 32.MarkH,EibeF,GeoffreyH,BernhardP,PeterR,IanW: TheWEKAData miningsoftware:anupdate. SIGKDDExplorNewsl 2009, 11 (1):10 – 18.33.WettschereckD,AhaDW,MohriT: Areviewandempiricalevaluationof featureweightingmethodsforacla ssoflazylearningalgorithms. ArtifIntellRev 1997, 10: 1 – 37. 34.NimunkarA,DhawanA,RelueP,PatwardhanS: Waveletandstatistical analysisformelanoma. In ProceedingsofSPIEMedicalImaging:Image Processing. EditedbyMilanS,MichaelFitzpatrickJ.SanDiego,CA:SPIE; 2002:1346 – 1352.Feb.23,2002. 35.BurroniM,CoronaR,Dell ’ EvaG,SeraF,BonoR,PudduP,PudduP, PerottiR,NobileF,AndreassiL,RubegniP: Melanomacomputer-aided diagnosis:reliabilityandfeasibilitystudy. ClinCancerRes 2004, 10: 1881 – 1886. 36.WalvickR,PatelK,PatwardhanS,DhawanA: Classificationofmelanoma usingwavelet-transform-basedoptimalfeatureset. In ProceedingsofSPIE MedicalImaging:ImageProcessing. EditedbyMilanS,MichaelFitzpatrickJ. SanDiego,CA:SPIE;2004:944 – 951.Feb.14,2004.Li etal.BMCMedicalImaging 2014, 14 :36 Page11of12 http://www.biomedcentral.com/1471-2342/14/36

PAGE 12

37.ZagroubaE,BarhoumiW: Anacceleratedsystemformelanoma diagnosisbasedonsubs etfeatureselection. JComputInfTechnol 2005, 1: 69 – 82. 38.LiL,ZhangQZ,DingYH,JiangHB,ThiersBT,WangJZ: AComputer-aided spectroscopicsystemforearlydiagnosisofmelanoma. In Proceedingsof IEEE25thInternationalConferenceonToolswithArtificialIntelligence. Edited byRandallB,ZoeyV.Washington,DC:IEEE;2013:145 – 150.doi:10.1186/1471-2342-14-36 Citethisarticleas: Li etal. : Automaticdiagnosisofmelanomausing machinelearningmethodsonaspectroscopicsystem. BMCMedical Imaging 2014 14 :36. Submit your next manuscript to BioMed Central and take full advantage of: € Convenient online submission € Thorough peer review € No space constraints or color “gure charges € Immediate publication on acceptance € Inclusion in PubMed, CAS, Scopus and Google Scholar € Research which is freely available for redistribution Submit your manuscript at www.biomedcentral.com/submit Li etal.BMCMedicalImaging 2014, 14 :36 Page12of12 http://www.biomedcentral.com/1471-2342/14/36