剪切力環(huán)境下的血管平滑肌細(xì)胞圖像分割技術(shù)研究-英文-解讀

2007年4月第44卷第2期四川高校學(xué)報(自然科學(xué)版)JournalofSichuanUniversity(NaturalScienceEdition)Apr.2007Vol.44No.2文章編號:049026756(2007)0220284205剪切力環(huán)境下的血管平滑肌細(xì)胞圖像分割技術(shù)探討李曉寧1,樊瑜波2,3(1.四川師范高校計算機(jī)科學(xué)學(xué)院,成都610068;2.四川高校建筑與環(huán)境學(xué)院生物力學(xué)所,成都610064;3.北京航空航天高校生物工程系,北京100083)摘要:定量分析血管平滑肌細(xì)胞VSMCs(VascularSmoothMuscleCells)在剪切力作用下的形態(tài)改變有助于理解VSMCs的生長機(jī)制以及病理學(xué)探討.圖像分割在圖像分析中扮演著重要角色,但同時也是圖像處理流程中最困難的步驟.針對培育在平板流室中的VSMCs相互交疊的細(xì)胞圖像,作者首先分析了這類細(xì)胞圖像的特征,隨后提出了一種具有四個步驟的圖像分割方法,分別是圖像預(yù)處理、照明校正、數(shù)學(xué)形態(tài)學(xué)重建和分水嶺轉(zhuǎn)換.試驗結(jié)果證明,根據(jù)這種方法分割細(xì)胞圖像,取得了較好的效果.關(guān)鍵詞:定量分析;分水嶺轉(zhuǎn)換;數(shù)學(xué)形態(tài)學(xué);圖像分割中圖分類號:TP751文獻(xiàn)標(biāo)識碼:ACellsegmentationofvascularsmoothmusclecellsundershearstressusingimageprocessingLIXiao2ning,FANYu2bo12,3(1.CollegeofComputerScience,SichuanNormalUniversity,Chengdu610068,China;2.BiomechanicalEngineeringLaboratory,SichuanUniversity,Chengdu610064,China;3.BioengineeringDepartment,BeihangUniversity,Beijing100083,China)Abstract:Quantitativelyanalysisofthemorphologicalresponsetoshearstressonvascularsmoothmusclecells(VSMCs)willfacilitateunderstandingthegrowthmechanismandpathologyofVSMCs.Imagesegmentationplaysakeyroleinimageanalysisandisalsothemostdifficultstepinimageprocessingpipeline.Inthispaper,firstly,thecharacteristicsoftheimagesofVSMCsculturedintheparallelplatechamberisanalyzedindetail,thenonesegmentationstrategywhichhasfourstepsbasedonwatershedoperatorsisproposed.Imageprepro2cessing,illuminationcorrecting,mathematicalmorphologicalreconstruction,watershedtransformationareimplementedaccordingtothesetsequence.Theresultshowsabetterseparationcellsfrombackgroundcanbeachievedbyusingthismethod.Attheendtheresultofimagesegmentationisanalyzedanddiscussed.Keywords:quantitativelyanalysis,watershed,mathematicalmorphology,imagesegmentation收稿日期:基金項目:作者簡介:通訊作者:2006210229國家自然科學(xué)基金(10527001,10672105);北京R&D項目(H060920051030);四川師范高校基金李曉寧(1972-),男,博士,主要探討方向為醫(yī)學(xué)圖像處理.樊瑜波.E2mail:ybfan@yahoo第2期李曉寧等:剪切力環(huán)境下的血管平滑肌細(xì)胞圖像分割技術(shù)探討2851IntroductionCellularmechanicsisafocusofnowadaystissueengineeringstudy.Investigationsintotheeffectsofflowonvascularsmoothmusclecells(VSMCs)aremotivatedbythepossibleinfluenceofflowinvascu2larbiologyandpathobiology.Todeterminetheex2tentoffloweffects,studiesinwhichthefluidme2chanicconditionscanbesystematicallyvariedareneeded.Invivostudyisdifficulttoquantitativelydefinethedetailedcharacteristicsofthehemody2namicenvironment.Exertingmechanicalforcethroughvariouskindsofmechanicsdevicesonlivingcellsinvitroisthemainmethodofcellmechanicsnow.Mechanicalstimuli,suchasstrainandfluidshearstress,arebelievedtobekeyfactorsinregu2latingcardiovasculartissuegrowthandremodelingduringdevelopmentandindiseasestates.Shearstressesonvascularsmoothmusclecellsduetobloodfluidflowplayavitalroleinregulatingtheirmor2phology,structure,growthrateandfunctions.Thechangeinmorphologyisaninstinctiveresponsetobiomechanicalenvironmentandisanindicatorforthefunctionalchangesinthecells.Quantitativelyanalyzingcellmotilityandmorphodynamicproper2tiesfacilitatesunderstandingtherelationsbetweenmechanicalstimuliandmechanismsofcellpathologi2calchanges.Itisnecessarytobuildadigitalimageprocessingsystemwiththeaidofelectronictech2nologiesandimageprocessingforthistarget.Howtoexactlyandaccuratelysegmentcellsfromimagesisthemajorproblemwiththeimageprocessingsys2tem.Imagesegmentationisoneofthehotspotsanddifficultstepofimageprocessingtechnology.Uptonowthousandsofmethodsofimagesegmentationhavebeenpresented.Butunfortunatelythereisnomethodappreciatedforeachkindofimagesegmen2tationproblems.Althoughsomealgorithmsarepro2posedfordyeingandfixingcellimagesegmentationtheyareunsuitablefortheimagesofbiologicalactivecells.Intheexperimentsofcellularmechanicsthematerialsoftenmustbekeptaliveandcouldnotbeendyed.Livingcellsaretransparentandtheirimagepropertiesvarywithexperimentalconditions.Theseallinducethedifficultiesfortheimagesegmenta2tion.Sodevelopingthesegmentationstudiesfortheimagesoflivingcellsculturedinthemechanicalen2vironmentisveryimportant.Thisarticleexploresthetechnologiesofimagesegmentationofvascularsmoothmusclecells(VSMCs).2Materialsandmethod2.1CellcultureVascularsmoothmusclecellswereobtainedfromrataortaaccordingtoageneralmethodusedinmedicine.Aninnovativeparallelplateflowchambersystemisusedtoculturethesecollectedcellsandtodeliverquasi2physiologicalpulsatileflowfieldforcul2turedvascularcells[1].2.2ImageacquiringsystemImageacquiringsystemconsistsofaphasecon2trastinvertedmicroscope(OlympusIX70,Japan),aprofessionalmicroscopyCCD(PixeraPro150ES,USA)andapersonalcomputer.AnalogcellimagesobservedbyOlympusIX70microscopeareconvertedtodigitalimagesthroughthevideooutputcompo2nent,andthesingleporttube(IX2SPT,Japan),aCmountandPixera150ESdigitalCCD.Atlastimagesarecapturedintocomputersystems.DiagramofourimageacquiringsystemisshownasFig.1.Fig.1DiagramoftheimageacquiringsystembasedonprofessionalmicroscopyCCD2.3CharacteristicsofvascularsmoothmusclecellimagesBeforesegmentationstrategyisdetermined,studyingthecharacteristicsofcellimagesisneces2sary.Generally,cellsimagespresentintwoclassicalmodes.Oneisoverlappedandtheotherisnon2over2lapped.Fornon2overlappedcellsimage,onemethod286四川高校學(xué)報(自然科學(xué)版)第44卷isintroducedinouranotherpaper[2].Asfortypicaloverlappedcellsimages,thereareseveralnotableandcommoncharacteristics.(1)Fromthewholeimage,thecontrastofintensitiesbetweenthecellsareaandthebackgroundislow.Althoughtherearesomehalationbetweencellsandbackground,thein2tensitiesofhalationisnotconstantandtheshapeofhalationisnotregular.(2)Theboundariesofthecellsmaybeextremelydifficulttodefineduetothevariationsofcellmorphology.(3)Manyholeslieinthebodyofcell.Andsotheintensitiesofsomeareasofcellbodyareidenticaltothoseofbackground.(4)Cellsaregrowinginoverlappingmode.Andthisconducesenormouslyimpedimentforseparatingcellsfromtheimages.2.4SegmentationStrategyObviously,itisimpossibletoaccuratelydistin2guishthecellboundaryfromthebackgroundonlybysimplemethods.Wementionthatwatershedisuse2fultoseparateoverlappingobjectsfromthescene.Sowedevelopthispipeline(Fig.2)tocompletethesegmentationtaskasfollows.Fig.2Illustrationoftheappliedimageprocessingpipeline2.4.1ImagepreprocessingInputimageisfirstlyenhancedbyaself2multiplemethod.TheRGBval2uesofeachpixelaremultipliedbythemselvesindi2viduallyandtheproductsarerespectivelynormalizedwithintherangefrom0to255.Thenormalizedval2ueisthenewintensityofthepixel.Thisprocedureisdonepixelbypixelandthengreytransformationisusedforeachpixeloftheconsequentimage.ThegrayscaleimageisshowninFig.4(a).2.4.2CorrectinginhomogeneousilluminationAlargenumberofexperimentfactorscancauselight2ingtobeeven,soinhomogeneousilluminationisacommonprobleminquantitativemicroscopy.Addi2tiveilluminationinfluencescanbenormalizedbyre2movinglocalmeansorutilizingderivativesforobjectdetecting[3].However,inthecaseofdiascopicmi2crographsthereisamultiplicationrelationshipbe2tweenthemainlyhigh2frequencycellobjectinforma2tionandratherlow2frequencyvariationsofillumina2tion.Logarithmicimagemodelissuitableforremov2ingmultiplicativeilluminationcomponents.Thisre2lationcanbeexpressedasF(x,y)=(g-A(x,y)3I(x,y)WheretheimagefunctionF(x,y)isregardedastheintensityoflightpassingthroughalightab2sorbingsample.A(x,y)andI(x,y)denotetheabsorptionandmicroscopeillumination,respective2ly.Theconstantgrepresentsthemaximumdigitalintensityvalue,whichis255for82bitimages.Inthelogarithmicdomain,itcanbedescribedasA(x,y)=g-exp(logF(x,y))-log(I(x,y))AccordingtotheworkofVolkerMetzleretc.[4],morphologicalilluminationcorrectionhasmanyadvantagesoverlinearlow2passfiltering.TheirmethodillustratedinFig.3isusedtocorrecttheinhomogeneousilluminationwedealwithandgoodresultisobtained.Onedifferencefromtheirmethodsisthesizeofstructureelementthatweadoptinmorphologicalreconstructionstep.Wehavementionedthatinourcapturingconditionsthesizeofcellbodyisalwayssmallerthan50pixels,soweFig.3Flowchartoftheilluminationbymorphologicalmethod第2期李曉寧等:剪切力環(huán)境下的血管平滑肌細(xì)胞圖像分割技術(shù)探討287Fig.4Resultimages(a)Graytransformation,(b)Illuminationcorrectionbymorpho2logicalmethods,(c)Morphologicalreconstruction,(d)Distancetransformationfor(c),(e)Markerimagebydomeextraction(h=4),(f)Overlayingmarkerimageondistanceimageerodetheimageusingstructureelementwhichsizeis25pixelsandconsequentlythecellularpixelinfor2mationiserased.Thenwereconstructtheimagebygeodesicdilationandcangetanimageconsistingofmostbackgroundinformation.Ifwereducetheo2riginalimagewiththisconsequentlyone,wecanob2tainanimagethatonlyincludescellbodies.Theob2tainedimageisshowninFig.4(b).AsshowninFig.4(b),althoughcellbodiesareseparatefrombackground,therearesomeback2groundareasarealsoretainedwhichisnotwanted.Soweutilizeanothermorphologicalreconstructionoperatortoeraseallthebackgroundareas.Differentfromlaststep,thisreconstructionoperatorisappliedtorestorecellbodyinformationandtogetridofbackgroundcomponentornoise.Becausecellsareconjointtogetherandbackgroundareaornoiseisin2solate,wecanerodetheimagewithlargestructureelementandtheobtainedimageislookedasmakerimagewhichsignallthecellareas.ReconstructiontheimageshowninFig.4(b)usingthismakerim2age,canclearupallthebackgroundandnoise.TheobtainedimageisshowninFig.4(c).Inthisimageonlycellareasarepresented.2.4.3WatershedtransformationAlthoughthebackgroundandnoisearealleliminated,cellsareoverlappedanditisnecessarytoseparatethemfromeachother.Watershedtransformationiseffectiveforsolvingthiskindofproblems,butinfact,thebrutalcomputationofwatershedsdoesnotconstituteagoodsegmentationmethod.Indeedthesimplecom2putationofanimage’swatershedsmostlyresultsinanover2segmentation.,i.e.,thecorrectcontourslostinamassofirrelevantones.Inordertogetridofthisover2segmentation[5],onemustmaketwopreparations.Oneistransformingoriginalimagetogradientoneordistanceone,whichwillbeinputimageofwatershed.Weusedistanceimagethoughdistancetransformationasinputimageofwater2shed.ThecorrespondingdistanceimageisshowninFig.4(d).Theotherisextractingmakerpointswhichsignobjectstobedissected.Universallyonemakerpointindexesoneobject.Onehastousetheknowledgeavailableontheproblemtodesignaro2bustalgorithmforextractingmarkersofdifferentre2gionstobesegmented.Inthispaper[6],h2dometransformationispresentedanditisverifiedthath2dometransformationcanextractlightstructureswithoutinvolvinganysizeorshapecriterion.Theonlyparameter(h)isrelatedtotheheightofthesestructures.Wedesignadometransformationwhichparameter(h)isequalto4andutilizeittogetmarkerpointsfromdistanceimage.Aftersuperpos2ingthemakerimage(Fig.4(e)withFig.4(d)andthenreversetheconsequentimagewecangetimageasshowninFig.4(f).Watershedoperatorisappliedtothisimage(asFig.4(f)).TheresultisshownFig.5(a).3ResultsanddiscussionAsshowninFig.5(a),thesegmentationlinesarepresentedinmanyconjunctionofcellbound2aries.Wecanlabelthesegmentedimageandgetthecontoursofcellbodyusingotherimageprocessingtechniques.Hence,quantitativeanalysiscanbesuc2ceeded.288四川高校學(xué)報(自然科學(xué)版)第44卷Fig.5Resultsofwatershedsegmentationwithdifferentmakerimages(a)h=4;(b)h=3;(c)h=5Asawhole,thelocationsofsegmentationlinesarecoincidentwiththeintuitionofthehumaneye2whendomeheightequalsto4.Butsomewhereseg2mentationlineslocateincongruously.Becausetheaccurateofwatershedseparationmethodseriouslydependsonthemakerimage,wetriedthreeheightsofdomewhichareparameterh=3,4,5,respec2tively.TheresultsareshowninFig.5(b),Fig.5(a),Fig.5(c).Itisevidentthatover2segmentationispresentedath=3andsomecellarenottotallysepara