Graph-Theoretical Methods for Statistical Inference on MR Connectome Data DARPA

Abstract

We are developing analytical tools to perform statistical inference on the connectome. Previous work has shown that simple measures of brain connectivity (e.g. total volume of white matter) are correlated with general cognitive functions such as intelligence. Because the connectome can be represented as a large interconnected graph (in which nodes are neuroanatomical regions and synapses are bundles of white matter tracts), we hypothesize that the development of algorithms based on principles of graph theory will allow for greater prediction of performance on measures of specific cognitive functions. To test this hypothesis, we have: (i) collected multimodal MR data from a large cohort of subjects from the Baltimore Longitudinal Study of Aging (BLSA), (ii) developed and applied a high-throughput fully-automated pipeline for extracting braingraphs from multimodal MR images, (iii) derived asymptotically optimal algorithms for graph classification, and (iv) applied these algorithms to simulations based on the BLSA data set. We show that our data processing pipeline is both efficient and robust. Furthermore, given a relatively small number of subjects, simulated classification accuracy approaches optimality. These results suggest that the developed methods may be useful for unraveling the detailed connectivity underlying many cognitive functions.