George Chen Evelina Fedorenko Nancy Kanwisher Polina Golland 12162011 NIPS MLINI Workshop 2011 1 Talk Outline Finding correspondences between functional regions in the brain ID: 1048178
Download Presentation The PPT/PDF document "Deformation-Invariant Sparse Coding for ..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
1. Deformation-Invariant Sparse Coding for Modeling Spatial Variability of Functional Patterns in the BrainGeorge Chen, Evelina Fedorenko, Nancy Kanwisher, Polina Golland12/16/2011NIPS MLINI Workshop 20111
2. Talk OutlineFinding correspondences between functional regions in the brainA new generative modelResults for language fMRI study12/16/2011NIPS MLINI Workshop 20112
3. Functional Region Correspondences12/16/2011NIPS MLINI Workshop 20113Given stimulus, get functional activation regionsSubject 1Subject 2Align to common anatomical spaceFunctional variability!Goal: Find correspondences between “parcels” contiguous region in braingroup-level parcelsParcel: contiguous region in brainBiology:brain compartmentalized into functional modules parcels represent these modules
4. Functional VariabilityStandard approach: just average in common anatomical space12/16/2011NIPS MLINI Workshop 20114Functional variability less pronounced activation in group averagespaceSubject 1Subject 2spaceAveragespaceAlignedspace
5. Previous WorkThirion et al. 2007: treat parcels as discrete objects and find parcel correspondences across subjects by matchingXu et al. 2009: generative, hierarchical model representing activation regions as Gaussian mixturesSabuncu et al. 2010: groupwise functional registration12/16/2011NIPS MLINI Workshop 20115
6. Previous WorkThirion et al. 2007: treat parcels as discrete objects and find parcel correspondences across subjects by matchingXu et al. 2009: generative, hierarchical model representing activation regions as Gaussian mixturesSabuncu et al. 2010: groupwise functional registration12/16/2011NIPS MLINI Workshop 20116
7. Our Generative Model12/16/2011NIPS MLINI Workshop 20117To generate image for a subject:Choose weights for eachgroup-level parcelForm weighted sum ofgroup-level parcelsDeform pre-image and add noisePre-imagee.g. (0.2, 1) Deformation: Group-level parcels1:2: sparse, no deformations sparse coding i.i.d. entries i.i.d. prior Goal: Estimate group-level parcels and deformations
8. Estimating Group-level Parcels and DeformationsPriors on group-level parcels and deformations from image registrationWant to be parcel, have sparse support, and smoothWant MAP estimate:Use generalized EM algorithm for MAP estimation 12/16/2011NIPS MLINI Workshop 20118 Don’t get to observe ’s! sparsitysmoothnessparcelidentifiability
9. Language fMRI StudyDataSubstantial functional variability!33 subjectsContrast: reading sentences vs. pronounceable nonwords are t-statistic images from standard fMRI preprocessingAll images initially brought into common anatomical spaceWhat we’ll showEstimated group-level parcels correspond to language processing regionsEstimated deformations improve fMRI group analysis 12/16/2011NIPS MLINI Workshop 20119
10. Left frontal lobeLeft temporal lobeEstimated Group-level ParcelsCorrespond to known language processing regions12/16/2011NIPS MLINI Workshop 201110Spatial support of group-level parcelsRight temporal lobeRight cerebellumExample group-level parcels
11. Apply estimated deformation to fMRI data for each subject and redo standard fMRI group analysis on separate data12/16/2011NIPS MLINI Workshop 201111Modeling functional variability increases statistical significance in each group-level parcelGroup-level Parcel IndexNegative log p-valueImproving fMRI Group Analysis with Estimated Deformations
12. Improving fMRI Group Analysis with Estimated Deformations12/16/2011NIPS MLINI Workshop 201112spaceSubject 1Subject 2AverageAlignedspacespacespaceWhy is the variance so high for statistical significance values for our model?
13. Improving fMRI Group Analysis with Estimated Deformations12/16/2011NIPS MLINI Workshop 201113AveragespaceWhy is the variance so high for statistical significance values for our model?Group-level parcel supportVariation using anatomical alignment onlyVariation using our model
14. Apply estimated deformation to fMRI data for each subject and redo standard fMRI group analysis12/16/2011NIPS MLINI Workshop 201114Modeling functional variability increases statistical significance in each group-level parcelGroup-level Parcel IndexNegative log p-valueImproving fMRI Group Analysis with Estimated Deformations
15. ContributionsGenerative model for finding group-level parcelsRepresent discrete set of parcels as imagesModel implicitly represents correspondences Just look at where -th group-level parcel shows up in each subject!Get deformations out of model, not just parcel correspondences! Improves fMRI group analysisFuture directionsUse estimated parcels in other fMRI studies as markers for language processing (and other stimuli!) 12/16/2011NIPS MLINI Workshop 201115