Hierarchical Dirichlet Processes for Modeling fMRI Brain Activation Patterns

Seyoung Kim

U. of California (Irvine)

Abstract

In this talk I will describe the use of Dirichlet process mixtures for modeling of spatial patterns across multiple images. A motivating application is the analysis of functional magnetic resonance imaging (fMRI) data for brain image. I present a mixture-based response-surface technique for extracting and characterizing spatial clusters of image intensity. This approach provides a richer and more powerful representation of the image data compared to more traditional voxel-based hypothesis testing methods. Dirichlet process priors are used to automatically select the number of activation clusters in a single image. This model can be further extended with hierarchical Dirichlet processes to extract common activation clusters from multiple images. In addition, I combine hierarchical Dirichlet processes with random effects model to capture the image-level variabilities in the shape of local clusters.

I describe an MCMC sampling method to simultaneously estimate both the shape parameters and the number of local activations, and demonstrate the application of the algorithm to fMRI brain images.

Bio:

Seyoung Kim is a Ph.D. candidate in the Department of Computer Science at UC, Irvine. Her research interests include statistical machine learning and Bayesian methods.
Maintained by: Fei Sha