Department of Statistical Science
Duke University

presents:

Christopher R. Genovese
genovese@stat.cmu.edu
Carnegie Mellon University

"Spatio-Temporal Inference and Functional Magnetic
Resonance Imaging"

Abstract:

The data obtained from an fMRI experiment are a realization of a very complex spatio-temporal process $Y_{vt}$, where $v$ ranges over a three-dimensional volume and $t$ over time. This process is subject to many sources of variation, both biological and technological. Neural activity manifests itself, via a blood-flow response in the brain, as slight changes in the signal which are embedded in a complex noise process.

Functional MRI serves as a prototype for a class of statistical problems that arises frequently in applications; namely, large data sets realized by a process with substantial spatial {\em and\/} temporal extent. Most work on spatio-temporal inference has focused on prediction, particularly extending common spatial techniques ({\em e.g.}, kriging). However, fMRI poses new challenges, both methodological and inferential, that require a different approach. The temporal component of the data contains critical information and is distinct from and intertwined with the spatial component. The spatial structure is complicated by highly irregular boundaries and non-homogeneous (and often non-local) correlations. There is a great deal of substantive prior information available that should be used. Clear scientific questions motivate the experiments, but the questions relate complicated functions of the entire data realization.

To address these challenges, I construct a family of nonlinear spatio-temporal hierarchical models for fMRI data and develop methods for applying these models to answer questions of scientific interest. I will illustrate the approach and describe computational techniques that should prove useful for more general spatio-temporal inferences.

February 21, 1997

4:00 pm - 5:00 pm

116 Old Chem Building