Table Of ContentUniversité Paris-Sud
Doctoral school 427 : Computer Science
Parietal Team - Inria Saclay
Large-scale functional MRI analysis
to accumulate knowledge on brain functions.
Yannick Schwartz
AdissertationsubmittedforthedegreeofDoctorofScience,
inthesubjectofComputerScience.
Defendedpubliclythe21thofApril2015.
Advisors: DrJean-BaptistePoline UniversityCaliforniaBerkeley,USA
DrBertrandThirion ParietalTeam-InriaSaclay,France
Reviewers: DrFinnÅrupNielsen TechnicalUniversityofDenmark,Denmark
PrTorWager UniversityofColorado,USA
Examiners: DrYvesBurnod FacultédemédecinePierreetMarieCurie,France
PrAlainDenise UniversitéParis-Sud,France
PrSimonEickhoff Heinrich-HeineUniversity,Germany
DrJean-BaptistePoline UniversityCaliforniaBerkeley,USA
DrGaëlVaroquaux ParietalTeam-InriaSaclay,France
Université Paris-Sud
Ecole Doctorale 427 : informatiqe Paris-Sud
Éqipe Parietal - Inria Saclay
Discipline : informatiqe
Thèse de doctorat
Analyse à grande échelle d’IRM
fonctionnelle pour accumuler la connaissance
sur les fonctions cérébrales.
Yannick Schwartz
Présentée et soutenue publiquement le 21 avril 2015.
Directeurdethèse: DrJean-BaptistePoline UniversityCaliforniaBerkeley,USA
Co-directeurdethèse: DrBertrandThirion ParietalTeam-InriaSaclay,France
Rapporteurs: DrFinnÅrupNielsen TechnicalUniversityofDenmark,Danemark
PrTorWager UniversityofColorado,USA
Examinateurs: DrYvesBurnod FacultédemédecinePierreetMarieCurie,France
PrAlainDenise UniversitéParis-Sud,France
PrSimonEickhoff Heinrich-HeineUniversity,Allemagne
DrJean-BaptistePoline UniversityCaliforniaBerkeley,USA
DrGaëlVaroquaux ParietalTeam-InriaSaclay,France
Abstract
Howcanweaccumulateknowledgeonbrainfunctions?Howcanwelever-
ageyearsofresearchinfunctionalMRItoanalysefiner-grainedpsycholog-
icalconstructs,andbuildacomprehensivemodelofthebrain?Researchers
usuallyrelyonsinglestudiestodelineatebrainregionsrecruitedbymen-
tal processes. They relate their findings to previous works in an informal
way by definingregions of interest from the literature. Meta-analysis ap-
proachesprovideamoreprincipledwaytobuildupontheliterature.
Thisthesisinvestigatesthreewaystoassembleknowledgeusingactiva-
tion maps from a large amount of studies. First, we present an approach
thatusesjointlytwosimilarfMRIexperiments,tobetterconditionananal-
ysisfromastatisticalstandpoint.Weshowthatitisavaluabledata-driven
alternative to traditional regions of interest analyses, but fails to provide
a systematic way to relate studies, and thus does not permit to integrate
knowledgeonalargescale. Becauseofthedifficultytoassociatemultiple
studies,weresorttousingasingledatasetsamplingalargenumberofstim-
uliforoursecondcontribution.Thismethodestimatesfunctionalnetworks
associatedwithfunctionalprofiles, wherethefunctionalnetworksarein-
teractingbrainregionsandthefunctionalprofilesareaweightedsetofcog-
nitivedescriptors.Thisworksuccessfullyyieldsknownbrainnetworksand
automaticallyassociatesmeaningfuldescriptions. Itslimitationslieinthe
unsupervisednatureofthismethod,whichismoredifficulttovalidate,and
theuseofasingledataset.Ithoweverbringsthenotionofcognitivelabels,
which is central to our last contribution. Our last contribution presents a
methodthatlearnsfunctionalatlasesbycombiningseveraldatasets.[Hen-
son2006]showsthatforwardinference,i.e.theprobabilityofanactivation
givenacognitiveprocess,isoftennotsufficienttoconcludeontheengage-
mentofbrainregionsforacognitiveprocess. Conversely,[Poldrack2006]
describesreverseinferenceastheprobabilityofacognitiveprocessgiven
anactivation,butwarnsofalogicalfallacyinconcludingonsuchinference
fromevokedactivity.Avoidingthisissuerequirestoperformreverseinfer-
encewithalargecoverageofthecognitivespace.Wepresentaframework
thatusesa"meta-design"todescribemanydifferenttaskswithacommon
vocabulary, and use forward and reverse inference in conjunction to out-
linefunctionalnetworksthatareconsistentlyrepresentedacrossthestud-
ies.Weuseapredictivemodelforreverseinference,andperformprediction
onunseenstudiestoguaranteethatwedonotlearnstudies’idiosyncrasies.
This final contribution permits to learn functional atlases, i.e. functional
networksassociatedwithacognitiveconcept.
4
We explored different possibilities to jointly analyse multiple fMRI ex-
periments. Wehavefoundthatoneofthemainchallengesistobeableto
relatetheexperimentswithoneanother.Asasolution,weproposeacom-
monvocabularytodescribethetasks. [Henson2006]advocatestheuseof
forward and reverse inference in conjunction to associate cognitive func-
tionstobrainregions,whichisonlypossibleinthecontextofalargescale
analysistoovercomethelimitationsofreverseinference. Thisframingof
theproblemthereforemakesitpossibletoestablishalargestatisticalmodel
of the brain, and accumulate knowledge across functional neuroimaging
studies.
Acknowledgements
IfirstwouldliketoexpressmygratitudetomytwoadvisorsJean-Baptiste
PolineandBertrandThirion,whogavemethisgreatopportunityandhelped
mealong myPhD withtheir supervision. Ialso wouldlike tothank Gaël
Varoquaux,whowasformenotonlyashadowadvisor,butalsotheperfect
spoonforourvarioustravels. Thankyouforthisamazingexperience,and
forintroducingmetotherealmofscience. IamalsogratefultoFinnÅrup
Nielsen and Tor Wager, for taking time to review my thesis and all their
insightful comments, as well as Yves Burnod, Simon Eickhoff, and Alain
Deniseforbeingpartofmydefensecommittee.
IwouldliketogivespecialthankstoalltheParietalteam,bothpresent
and past members: you are the reason why this team is so uniquely and
trollinglydisruptive.SobigthankstoVincentMichel(a.k.a.NickelMitchell)
for showing me that technical skills were optional to do a PhD, and Alan
Tucholka (a.k.a. the Graou) for teaching me the relational skills I needed
inaprofessionalenvironment,CéciliaDamonwhowaslikeabigsisterto
me, Alexandre Gramfort (malade), Pierre "Commissaire" Fillard, Philippe
"tchoutchou" Ciuciu, Fabian "Chicken" Pedregosa, Viviana "No no no" Si-
less,VirgileFritsch,petaflopicBenoîtDaMota,MichaelEickenberg,Olivier
Grisel,LoïcEstève,AndrésHoyosIdrobo,AlexandreAbraham,Konstantin
Shmelkov, Danilo "triple that awsome" Bzdok, Aina Frau, Solveig Badillo,
Salma Bougacha, Mehdi Rahim, Nicolas Chauffert, Ana Luísa Pinho, and
KamalakarReddy. SpecialthankstoRégineBricquet,BarbaraMoulin,and
MaikeGilliotwhosavedmefrommanyadministrativehurdles.
I have spent a bit more than seven and a half years at Neurospin, and
hadthechancetomeetmanypeoplethatmadethisexperiencerichbothon
theprofessionalandpersonal levels. Itmayjustlooklikealong enumer-
ation of names but I really enjoyed working and/or spending (too much)
time around a coffee with each and every single one of you. Thanks to
all the IMAGEN team, in particular Benjamin Thyreau, who taught me
right from wrong, Alexis Barbot, Fanny Gollier-Briant, and Eric Artiges.
I would also like to acknowledge my adoptive parents Edouard Duches-
nayandDimitriPapadopoulos,aswellasmycollaboratorsfromtheother
teams:JulienLefèvre,DavidGermanaud,ClaraFischer,PaulineRoca,Edith
LeFloch,SoizicLaguitton,MatthieuPerrot,PamelaGuevara,UrielleTho-
prakarn,GrégoryOperto,MathieuDubois,YannCointepas,DenisRivière,
AlexisRoche,LaurentRisser,NicolasSouedet,Jean-FrançoisMangin,Vin-
centFrouin,MélanieStrauss,DenisEngemann,MurielleFabre,LauraDu-
pas,ElodieCauvet,CatherineWacongne,MarieAmalric,ValentinaBorghe-
6
sani,ChristophePallier,LucieCharles,AnneKozem,andAurélienMassire.
AndIalsowouldliketocongratulateBaptisteGauthierforhisPhD.
Thanksalsotoallmyfriendswhoremindedmethatthereismoretolife
than academic research: my forever roommates Colin and Rémi, Ahmed,
Pieb, Piouf, Aurélien, Ienien, Chris, and Soubouzou San. Finally special
thankstoallmyfamily,inparticulartomyparents,andCédricandChris-
tel for their continued support. Last but not least, thanks to my beloved
Rebecca,whowhippedmewhenIneeded,andbelievedinmewhenIcould
not.
Contents
17
I State of the art: a brief introduction to neuroimaging
19
1 From brain images to the study of the mind
1.1 Frombrainlesionstofunctionalimaging 20
1.2 BOLDFunctionalMRI 20
1.3 Mappingmentalprocessestothebrain 22
1.4 Conclusion 25
27
2 Tools for neuroimaging data modeling
2.1 Statisticalinference 28
2.2 StatisticallearningforfMRI 32
2.3 Quantitativemeta-analyses 39
2.4 Conclusion 42
47
II Contributions: from an image database to learning brain functions
49
3 Scaling up from individual studies
3.1 Findingthedata 50
3.2 Fromdiversedatasourcestocuratedbrainmaps 51
3.3 Conlusion 57
65
4 Functional localization by meta-analysis
4.1 Improvingaccuracyandpowerwithtransferlearningusingameta-analyticdatabase 69
4.2 OnspatialselectivityandpredictionacrossconditionswithfMRI 75
4.3 Conlusion 80
8
83
5 Learning functional networks
5.1 Introduction 84
5.2 Amulti-subjectsparse-codingmodelofbrainresponse 85
5.3 EfficientlearningofRFX-structureddictionaries 87
5.4 Resultsonsimulateddata 89
5.5 LearningacognitivebrainatlasfromfMRI 90
5.6 Conlusion 93
97
6 Learning functional atlases
6.1 Annotatingbrainmaps 98
6.2 Inferringconcept-specificnetworks 99
6.3 Functionalatlases 108
6.4 Conclusion 111
117
A Appendix: datasets
A.1 BalloonAnalogRisk-takingTask(ds000001) 117
A.2 Classificationlearning(ds000002) 118
A.3 Rhymejudgment(ds000003) 119
A.4 Mixed-gamblestask(ds000005) 120
A.5 Living-nonlivingdecisionwithplainormirror-reversedtext(ds000006) 121
A.6 Stop-signaltaskwithspoken&manualresponses(ds000007) 122
A.7 Stop-signaltaskwithunconditionalandconditionalstopping(ds000008) 123
A.8 Thegeneralityofself-control(ds000009) 124
A.9 Classificationlearningandtone-counting(ds000011) 125
A.10Classificationlearningandstop-signal(1yeartest-retest)(ds000017) 126
A.11Cross-languagerepetitionpriming(ds000051) 127
A.12Classificationlearningandreversal(ds000052) 128
A.13Simontask(ds000101) 129
A.14Flankertask(event-related)(ds000102) 130
A.15Visualobjectrecognition(ds000105) 131
A.16Wordandobjectprocessing(ds000107) 132
A.17Prefrontal-SubcorticalPathwaysMediatingSuccessfulEmotionRegulation(ds000108) 133
A.18Falsebelieftask(ds000109) 134
A.19Incidentalencodingtask(PosnerCueingParadigm)(ds000110) 135
Description:A.31 A Parametric Empirical Bayesian Framework for the EEG/MEG Inverse Problem: widespread experimental procedure to localize brain functions with fMRI, Montreal Neurological Institute (MNI) [7] brain spaces. typical case is that of early-stage clinical trials, for which the group size is.