A Bayesian General Linear Modeling Approach to fMRI Data Analysis

• Activity Number: 547
• Type: Topic Contributed
• Date/Time: Wednesday, August 6, 2014 : 2:00 PM to 3:50 PM
• Sponsor: SSC
• Author(s): Ryan Yue*+
• Companies: City University of New York

Abstract

The general linear model (GLM) (Worsley and Friston, 1995) has arguably become the dominant statistical method for analyzing functional Magnetic Resonance Imaging (fMRI) data. It models the fMRI time series as a linear combination of different signal components and tests whether activity in a brain region is systematically related to a number of pre-specified input functions. In this work we use a Bayesian GLM approach, which has a number of advantages compared to existing methods. First, we account for the spatial-temporal structure in fMRI data using a class of sophisticated spatial processes to model baseline and amplitude fields, while assuming autoregressive errors in the temporal domain. Second, we propose a novel joint posterior probability map (PPM) method that allows us to identify regions of activation, rather than just voxels. Third, we extend the methodology from the single-subject to the multi-subject case, thus facilitating population inference. Finally, we employ an approximate Bayesian inference tool to relieve the computational burden imposed by the massive size of fMRI data. The method is validated through simulation studies and applied to data from an fMRI study