Bayesian Spatial Modeling of fMRI Data: A Multiple Subject Analysis.

Bayesian Hierarchical Model;Mixture Model;Reversible Jump MCMC;Human Brain Mapping;Multiple Subject FMRI Data Analysis;Dirichlet Process Prior;Public Health;Health Sciences;Biostatistics

Xu, LeiQin, Zhaohui ;

University of Michigan

关键词: Bayesian Hierarchical Model;    Mixture Model;    Reversible Jump MCMC;    Human Brain Mapping;    Multiple Subject FMRI Data Analysis;    Dirichlet Process Prior;    Public Health;    Health Sciences;    Biostatistics;   

Others  :  https://deepblue.lib.umich.edu/bitstream/handle/2027.42/57656/leix_1.pdf?sequence=2&isAllowed=y

瑞士|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

The aim of this dissertation is to develop spatial models formulti-subject fMRI data. While there has been much work on univariate modeling of each voxelfor single- and multi-subject data, and some work on spatial modelingfor single-subject data, there has been no work on spatial models thatexplicitly account for intersubject variability in activationlocation. In this dissertation, we present explicit spatial models in an hierarchical Bayesian framework. At the first levelwe model ``population centers;;;; that mark the centers of regions of activation.At the second level,each subject may have zero or moreassociated ``individual components;;;; fora given population center.At the third level, we useGaussian mixtures for the probability that a voxel belongs to anactivated region. We start with a finite mixture model for the prior distribution of the individual component centers. We then extend it to an infinite mixture model via a mixture of Dirichlet process priors. Our approach incorporates the unknown number of mixture components into themodel as a parameter whose posterior distribution is estimated byreversible jump Markov chain Monte Carlo. Although we are motivated by fMRI data, this model can easily bemodified to handle other types of imaging data. We demonstrate our method with simulated data sets and a fMRI study of visual working memory and show better precision of localization with our method relative to the standard mass-univariate method.