Abstract — Heart tracking in ultrasound sequences is a difficult task due to speckle noise, low SNR and lack of contrast. Therefore it is usually difficult to obtain robust estimates of the heart cavities since feature detectors produce a large number of
1
AMethodfortheDynamicAnalysisoftheHeartUsinga
LyapounovBasedDenoisingAlgorithm
JacintoC.NascimentoJo aoM.SanchesJorgeS.MarquesInstitutodeSistemaseRob´otica-InstitutoSuperiorT´ecnico
Abstract—Hearttrackinginultrasoundsequencesisadif culttaskdue
tospecklenoise,lowSNRandlackofcontrast.Thereforeitisusuallydif culttoobtainrobustestimatesoftheheartcavitiessincefeaturedetectorsproducealargenumberofoutliers.Thispaperpresentsanalgorithmwhichcombinestwomainoperations:i)anoveldenoisingalgorithmbasedontheLyapounovequationandii)arobusttracker,recentlyproposedbytheauthors,basedonamodeloftheoutlierfeatures.Experimentalresultsareprovided,showingthattheproposedalgorithmiscomputationallyef cientandleadstoaccurateestimatesoftheleftventricleduringthecardiaccycle.
I.INTRODUCTION
Theestimationoftheleftventricle(LV)boundaryisusefulfromaclinicalpointofviewbecauseitallowstoextractrelevantmeasuresoftheheartdynamicbehavior,amongwhichwestresstheejectionfractionorlocalwallmotion.
Ultrasoundimaging,isoneofthemostusedtechniquestoobservethedynamicsoftheheart.However,thelowsignal-to-noiseratio(SNR)andthemultiplicativenatureofthenoise(speckle)corruptingtheultrasoundimages,maketheLVsegmentationadif culttask.
Edgedetectionalgorithmsfailduetothepresenceofmultiplicativenoiseinheartultrasoundimages.Thestrongestedgesareoftennotlocatedontheendocardium(insomecases,therearenovisualcueforthepres-enceofthemyocardiumborder).In[1]itisproposedtheinstantaneouscoef cientofvariation(ICOV)providinggoodsegmentationresults,butthesocalledproblemof“edgedropout”stillremain(particularlyinthediastolephase).
Therefore,noisereductionmustbeappliedbeforeedgedetection.Severaltechniqueshavebeenproposedtoreducethespecklenoisewithoutdistortingtherelevantclinicaldetails,e.g.,Bayesianmethods[2,3],mixturedistributionoftheRicianpdfwiththeinverseGaussianasamixturedistribution(RiIG)[4],softthresholding[5],waveletbasedmethods[6],waveletsoft-shrinking[7],ormultiresolutionbasedtechniques[8],median ltering[9],andanisotropicdiffusion[10].
Eventhoughthedenoisingalgorithmssigni cantlyreducethespecklenoise,advancedtrackingtechniquesareneededtosegmenttheLVboundaries.Priorartinsegmentationofechocardiographicsequencesoftheheartincludesactiveshapemodels[11,12],activecontours[13]andlevelsettechniques[14].Pairsofcoupledcontourshavealsobeenused.Theupdatemethodsusegeometricrestrictions(e.g.,subspaceshapemodel),dynamicmodelsandthemeasurementsmodelse.g.,usingheteroscedasticnoise[15].
InthispaperwejoinanoveldenoisingalgorithmbasedontheLyapounovequationandarecentrobusttracker[16].Thedenoisingalgorithmmusttoprocessalargenumberofultrasoundimagesinanef cientway[17].Thisisobtainedshowingthatthesolutionofthe lteringprocessveri esaLyapunovequation,whichcanbeef cientlybyknownalgorithms.
Therobustnessofthetrackerisobtainedbyusingmiddlelevelfeatures(strokes),whicharelabelledasvalidorinvalid.Sincethelabelsareunknowntheyarereplacedbytheirprobabilitiescomputedusingaprobabilisticmodeloftheobservations.Adataassociation lteristhenusedtoupdatethecontourparametersunderthepresenceofoutliers.Thetrackingalgorithmproposedinthepaperwasassessedusingasetofimagesequences,segmentedbymedicaldoctors.Theseimages,areusedasagroundtruthtocomputeFOM( guresofmerit).The
performanceofthetrackercomputedinthiswayiscomparedwithothertrackingalgorithms.
Thepaperisorganizedasfollows:SectionIIdescribestheoverallsys-tem.SectionsIII,IVandVdescribethepre-processing,featuredetectionandtrackingstepsrespectively.SectionVIdescribesexperimentalandsectionVIIconcludesthepaper.
II.SYSTEMOVERVIEW
Theproposedsystemaimsistotracktheboundaryoftheleftventricleduringthecardiaccycle.Thesysteminputisasequencesofultrasoundimagessampledat25Hz.
Thesystemperformsthreemainoperations(seeFig.1);i)denoising:toreducethespecklenoiseandenhancethecontrast;duetothelargeamountofdatatobeprocessedanovelalgorithmwasdevelopedtoperformthistask,ii)featuredetection:detectsintensitytransitionsalongorthogonallinestothecontour(seeFig.2).Transitionsareobtainedbyapplyingamatched ltertotheintensitypro lesandcomputingthelocalmaxima[18],andiii)trackingbasedonarobusttrackingalgorithmwhich tsadeformablecurvetothepointsdetectedintheimage.Thisalgorithmmustbeabletodealwithalargenumberofoutliersandtointerpolatetheboundarywhennofeaturesaredetectedduetolowcontrastoftheheartboundary.Thisisspeciallyimportantclosetotheapexandinthepresenceofsuddenmotionchanges(e.g.,inthemitralvalve).Arecenttrackingalgorithmisusedinthisstep.

Fig.1.Systemoverview.
III.PRE-PROCESSING
Theperformanceofthetrackerdependsonthesignaltonoiseratio,SNR,oftheinputimages.TheimageshavealowSNRduetothepresenceofmultiplicativenoise(speckle).ThespecklenoiseismodeledwithaRayleighdistributionwithspacevaryingparameter[19].
Thegoalofthepre-processingstepistoreducethenoisewithoutlosingrelevantinformation.InthispaperaMAPcriterionisusedtoestimatetheoriginalimagesfromthenoisyobservations.Thisapproach,however,isslowandcomputationallydemanding.Thisdif cultyisparticularysevereinthepresentproblembecausewearedealingwithalargenumberofultrasoundimages.
InthispaperweuseafastMAPdenoisingalgorithm.Thedenoisingalgorithmestimatestheoriginalimage,F,byminimizingthefollowingenergyfunction
E(Y,F)= log(p(Y|F)p(F))
(1)
whereYisthenoisyimage,p(Y|F)istheobservationmodelandp(F)isthepriordistributionoftheunknownimage.
Assumingconditionalindependenceofthesites,wehavep(Y|F)=QN,Mobservations,atdifferent
p(yij|fij)wherep(y y2/2f
istheRayleighdistribution.
ijy|f)=eConcerningtheprior,aGibbsdensityfunctionwithquadraticpotentialfunctionsisused,i.e.,
p(F)=
1 αPp,q∈S(fp fq)2Z
e(2)
The Lyapounov exponent of the dynamical systems, ...(26) A New Chaos-Based Encryption Method for ... Security analysis of a chaos-based image ...
The Lyapounov exponent of the dynamical systems, ...(26) A New Chaos-Based Encryption Method for ... Security analysis of a chaos-based image ...
热门文档