Stochastic Systems Group
Home Research Group Members Programs  
Demos Calendar Publications Mission Statement Alumni

SSG Seminar Abstract

Functional Hierarchy: Representation and Modeling of Spatial Patterns of Activation in fMRI

Polina Golland
Computer Science and Artificial Intelligence Laboratory, Massachusetts Institute of Technology

In this talk, I will present a novel approach to computational modeling of spatial activation patterns observed through fMRI. Functional connectivity analysis is widely used in fMRI studies for detection and analysis of large networks that co-activate with a user-selected `seed' region of interest. In contrast, our method is based on clustering; it simultaneously identifies interesting seed time courses and associates voxels with the respective networks. This generalization eliminates the sensitivity to the threshold used to classify voxels as members of a network and enables discovery of co-activated networks without user selection of seed regions.

Based on the empirical observation that the detected patterns of co-activation are inherently hierarchical, we propose a new representation for spatial patterns of functional organization. Just like the anatomical hierarchies represent the structure of the brain as a tree of increasingly simple systems, we believe that the functional description of the brain should also be of a hierarchical nature. We introduce Functional Hierarchy, a top-down representation that encapsulates the notion that functionally defined regions should be viewed at different resolutions, as dictated by the observed activation pattern. We construct the functional hierarchy through an iterative decomposition that utilizes clustering for network subdivision at each step.

The experimental results demonstrate that the functional region hierarchy provides a robust and anatomically meaningful model for spatial patterns of co-activation in fMRI. The hierarchical representation leads to insights into the structure of the functional networks that are not immediately apparent from flat representations that segment the brain into a large number of small regions. In addition, subject-specific region hierarchies tend to share common tree structure, further confirming the validity of this representation for modeling group-wise patterns of co-activation.

Problems with this site should be emailed to