FrankH.Guenther,AyoubDaliri,AlfonsoNieto-Castanon,MeganThompson,andJasonA.Tourville

Advancesinneuroimagingtechniquessuchasfunctionalmagneticresonanceimaging (fMRI)andelectrocorticography(ECoG)overthepasttwodecadeshaveresultedinagreatlyimprovedunderstandingoftheneuralmechanismsunderlyinghumansensory, motor,andcognitivecapabilities,leadingtoincreasinglysophisticatedneuralmodelsof thesefunctions.Functionalneuroimaginghasbeenapowerfulmeansforevaluatingand refiningsuchmodels.Todate,however,theseevaluationshavebeenalmostexclusively qualitative.Quantitativeevaluationshavebeenhamperedbytheabsenceofageneralcomputationalframeworkfor(i)generatingpredictedfunctionalactivationfromamodelthatcanbedirectlyandquantitativelycomparedtoempiricalfunctionalneuroimaging data,and(ii)testingbetweenmodelstoidentifythemodelthatbestfitsexperimentaldata.Herewepresentageneralcomputationalframeworktoovercometheseissues.In thisframework,thebrainnetworkresponsibleforataskisbrokenintoasetof computationalnodes,eachofwhichislocalizedtoanMNIstereotacticcoordinateinthebrain.Associatedwitheachnodeisacomputationalloadfunctionthatlinksthenode’s activitytoacomputationinvolvingquantifiablemeasuresfromthetask.Theinstantaneousneuralactivityateachlocationinthebrain(e.g.,eachvoxelofanfMRI image,oreachelectrodeofanECoGarray)isthencalculatedbysummingthecontributionsofallmodelnodesatthatlocation,witheachnodetreatedasaGaussian activitysourcecenteredatthenode’slocation.TheparametersoftheGaussians(i.e., spreadandamplitude)areoptimizedtoproducethebestfittothefunctionaldata.

Quantitativecomparisonsbetweendifferentmodelsarebasedontheoverallfitleveland numberoffreeparametersusingtheAkaikeInformationCriterion(AIC).Thisframework hasbeenappliedtotwolargefMRIdatabases,oneinvolvingstudiesofworking    memoryandanotherinvolvingstudiesofspeechproduction.Theresultsoftheseanalyses highlighttheframework’sabilitytoprovidequantitativefitstothefMRIdata              fromavarietyofcognitivemodels,aswellasitsutilityforselectingthebestmodelamongstcompetingmodelsbasedonamountofinformationlost.[SupportedbyNIH grantsR01DC002852,R01DC007683.]