Parcellation-Free prediction of task fMRI activations from dMRI tractography

Abstract

The relationship between brain structure and function plays a crucial role in cognitive and clinical neuroscience. We present a supervised machine learning based approach that captures this relationship by predicting the spatial extent of activations that are observed with task based functional Magnetic Resonance Imaging (fMRI) from the local white matter connectivity, as reflected in diffusion MRI (dMRI) tractography. In particular, we explore three different feature representations of local connectivity patterns that do not require a pre-defined parcellation of cortical and subcortical structures. Instead, they employ cluster-based Bag of Features, Gaussian Mixture Models, and Fisher vectors. We demonstrate that our framework can be used to test the statistical significance of structure-function relationships, compare it to parcellation-based and group-average benchmarks, and propose an algorithm for visualizing our chosen feature representations that permits a neuroanatomical interpretation of our results.