WhyPreprocess?
• Dataisnoisy!• Subjectsmove!
• ThingschangeovertheJmeofyourexperiment(moreifit’slongandboring)!
• Thereagiantbulldozersandexcavators10feetawayfromyourscanner!Wait,that’snotnormal..?
• Etcetcetc…
WhyPreprocess?• PreprocessingaUemptstoincreaseSignal(e.g.,BOLDcontrast)toNoise(variancefrommovement,subject,scannerarJfacts,otheruncontrollables)raJo
• HelpsyoumeetassumpJonssoyoucandostatsonyourdata
PreprocessingResources
• Fullcreditforalltheinfoandmostoftheslides(andnicepictures)goesto:
• FSLtutorials‐– hUp://www.fmrib.ox.ac.uk/fslcourse/
• UCLANITPSummerCourse(SeeMonJ7/12Preprocessing)– hUp://www.brainmapping.org/NITP/Summer2011.php
PreprocessingSteps
• Pre‐Preprocessing– DICOMtransformaJon,ImagereconstrucJon,BET
• MoJoncorrecJon• Slice‐JmingcorrecJon
• SpaJalfiltering• Temporalfiltering• GlobalintensitynormalizaJon
• RegistraJon/Normalizing(technicallypost‐preprocessing)
PreprocessingSteps
• Pre‐Preprocessing– DICOMtransformaJon,ImagereconstrucJon,BET
• MoJoncorrecJon• Slice‐JmingcorrecJon
• SpaJalfiltering• Temporalfiltering• GlobalintensitynormalizaJon
• RegistraJon/Normalizing(technicallypost‐preprocessing)
Pre‐Preprocessing
• Scanneroutput:DICOMS• Analysisprograminput:Analyze(SPM),NiFTi(FSL,SPM),BrainVoyager
• SPM,BrainVoyagerhavebuilt‐infuncJons
• FSL–useMRICRON’sdcm2niiconvert– Cansetuppreferencesondcm2nii.inifile– Putalldcmsinonefolder
– dcm2nii–b/ApplicaJons/mricronmac/dcm2nii.inifirscile.dcm
Pre‐Preprocessing
• ImageregistraJon– Lookatyourdata!UseFSLvideomode.
– OsirixViewer(oranyotherDICOMviewer):• hUp://www.osirix‐viewer.com/
• BrainextracJon– MostlyinFSL,useBETtoolbox
PreprocessingSteps
• Pre‐Preprocessing– DICOMtransformaJon,ImagereconstrucJon,BET
• MoJoncorrecJon• Slice‐JmingcorrecJon
• SpaJalfiltering• Temporalfiltering• GlobalintensitynormalizaJon
• RegistraJon/Normalizing(technicallypost‐preprocessing)
MoJonCorrecJon
• “Pleasebevery,very,verysJll….”• Movementthat’s1%ofvoxelsizecaninducea1%signalchange,whichissomeJmesgreaterthanyouractualBOLDsignal!(saytheFSLgods)
• Also,yourvoxelisnolongerthesamevoxel…
MoJonCorrecJon
• Chooseavolumetoregisteralltheothervolumesto(e.g.,first,middle,last,mean,standardspace;inFSLitsmiddle)
• Withinsubject:use6DOFrigid
• X,y,z,roll,pitch,yaw• Realignstoreferencetominimizevariance
MoJonCorrecJon
• Othernotes:– Canpreventwithtraining,makesubjectsfeelrelaxed(saying“don’tmove!”usuallymakesthemmorenevous!)
– GivethemaminutetogetseUledandcomfortable– Makesuretaskisn’tcorrelatedw/moJon
– ToomuchmoJon?Throwoutrun/subject..
PreprocessingSteps
• Pre‐Preprocessing– DICOMtransformaJon,ImagereconstrucJon,BET
• MoJoncorrecJon• Slice‐JmingcorrecJon
• SpaJalfiltering• Temporalfiltering• GlobalintensitynormalizaJon
• RegistraJon/Normalizing(technicallypost‐preprocessing)
Slice‐TimingCorrecJon• Scanners(likeourSiemens3T)mayacquireslicesinaninterleavedfashion(e.g.,0,2,4..1,3,5)toavoidcontaminaJngneighboringslice
• Cancorrectandputslicesbackinorder• Currentconsensusisnoneed(FSL,MonJ)– It’sanotherinterpolaJon(theless,thebeUer)– Itdoesn’thelpthatmuch,couldmakethingsworse
– Instead,addTEMPORALDERIVATIVEtomodel
PreprocessingSteps
• Pre‐Preprocessing– DICOMtransformaJon,ImagereconstrucJon,BET
• MoJoncorrecJon• Slice‐JmingcorrecJon
• SpaJalfiltering• Temporalfiltering• GlobalintensitynormalizaJon
• RegistraJon/Normalizing(technicallypost‐preprocessing)
SpaJalFiltering(Smoothing)
• Averagesonevoxel’svalueswithitsneighbors• Minimum“smoothness”neededifusingGaussianrandomfieldtheory(FSL)
• PROS:canincreaseSNRbydecreasingvariance
• CONS:mayreducesignalifsmallacJvaJons– Ifyouexpectthis,useasmallerkernel
– AlsoreducesspaJalresoluJon
SpaJalFiltering(Smoothing)
• Onaverage,akernelof5mmisadequate– Picksomethingabout2.5xFWHM
– Cangoupto10‐15mmifexpecJnglargeacJvaJons
– Ornotuse(otheropJonsforthresholding)
PreprocessingSteps
• Pre‐Preprocessing– DICOMtransformaJon,ImagereconstrucJon,BET
• MoJoncorrecJon• Slice‐JmingcorrecJon
• SpaJalfiltering• Temporalfiltering• GlobalintensitynormalizaJon
• RegistraJon/Normalizing(technicallypost‐preprocessing)
TemporalFiltering• Temporalnoise:– Temporaldriofromscanner;physiologicalcycles(cardiac,repiratory)
• Thesecanmaskyouractualsignal!
• Highpassfilterletshighfrequenciesthrough• Lowpassfilterletslowfrequenciesthrough(removeshigh‐frequencyfluctuaJons)butcanremovesignal,espifevent‐related.
• Usuallyjustusehighpass–FSLusesprewhiteningtoavoidlow‐pass
TemporalFiltering
• GenerallyaHPFilterof100sisadquate– Changeifyourmodelisdifferent(egverylongevent)
• SPM:modelslowdriostocatchvariance;usescosinebasisset
• FSL:removeslowdriosandthenconvolvessignalwithgaussian‐weightedrunningline
PreprocessingSteps
• Pre‐Preprocessing– DICOMtransformaJon,ImagereconstrucJon,BET
• MoJoncorrecJon• Slice‐JmingcorrecJon
• SpaJalfiltering• Temporalfiltering• GlobalintensitynormalizaJon
• RegistraJon/Normalizing(technicallypost‐preprocessing)
GlobalIntensityNormalizaJon• FromMonJ2011slides:• GOOD:• Between‐sessions(grandmeanscaling)soyouarecomparingallrunsbyasinglefactor(canbearbitrary)centeredonthesamemean
• BAD:• Within‐session(globalscaling),forceseachVOLUMEtohavesamemeanintensity,that’ssilly
PreprocessingSteps
• Pre‐Preprocessing– DICOMtransformaJon,ImagereconstrucJon,BET
• MoJoncorrecJon• Slice‐JmingcorrecJon
• SpaJalfiltering• Temporalfiltering• GlobalintensitynormalizaJon
• **ICADenoising• RegistraJon/Normalizing(technicallypost‐preprocessing)
ICADenoising
• Manuallyremovenoisecomponents
• FeedrawdataintoMELODIC(FSL;MulJvariateExploratoryLinearOpJmisedDecomposiJonintoIndependentComponents)
• Modelfree–pullsoutallthe“components”• CanmanuallygothroughcomponentsandidenJfynoise![demo]
• SeeKelly2010JNeurosciMethodsfordetailsonwhattoremove
• EspeciallygoodifresJngstatedata;oroddmovements!Badifnoiserelatestosignal…
RegistraJon
• Allowscomparingbetweenindividualsbymappingalltoatemplatebrain(fromMonJ2011slides)
• Rigidbody:– 6DOF–3rotaJons,3translaJons,typicallywithinsubj– 7DOF–addglobalscaling,goodforfunctoanatomical
• Affine:– 12DOF–3rotaJons,3translaJons,3scalings,3skewings– SubjecttotemplaterotaJon
• Non‐linear:– RepresentedbyadeformaJonfield(notamatrix)
RegistraJon
• Re:Non‐lineartransformaJons
• SeeFSLspecificallyonregistraJonandMonJformoredetailedinformaJon(FSLFLIRT,FNIRT)
• SPMDARTELisalsogood(soIhear)• LotsofopJonsoverall!
RegistraJon
• Whenregistering,youhavetointerpolateagain–whattodowithnewemptyspaces?– Nearestneighbor?Trilinear(4neighbors)?Sinc(10)?
• FSL:analyzesinnaJvespace,thennormalizes• SPM:normalizesfirst,thenGLM,decreasesvariance&SD
• Canaccountforsomedifferencesinresults!