In search of functional specificity in the brain : generative models for group fMRI data

Abstract

In this thesis, we develop an exploratory framework for design and analysis of fMRI studies. In our framework, the experimenter presents subjects with a broad set of stimuli/tasks relevant to the domain under study. The analysis method then automatically searches for likely patterns of functional specificity in the resulting data. This is in contrast to the traditional confirmatory approaches that require the experimenter to specify a narrow hypothesis a priori and aims to localize areas of the brain whose activation pattern agrees with the hypothesized response. To validate the hypothesis, it is usually assumed that detected areas should appear in consistent anatomical locations across subjects. Our approach relaxes the conventional anatomical consistency constraint to discover networks of functionally homogeneous but anatomically variable areas. Our analysis method relies on generative models that explain fMRI data across the group as collections of brain locations with similar profiles of functional specificity. We refer to each such collection as a functional system and model it as a component of a mixture model for the data. The search for patterns of specificity corresponds to inference on the hidden variables of the model based on the observed fMRI data. We also develop a nonparametric hierarchical Bayesian model for group fMRI data that integrates the mixture model prior over activations with a model for fMRI signals. We apply the algorithms in a study of high level vision where we consider a large space of patterns of category selectivity over 69 distinct images. The analysis successfully discovers previously characterized face, scene, and body selective areas, among a few others, as the most dominant patterns in the data. This finding suggests that our approach can be employed to search for novel patterns of functional specificity in high level perception and cognition.