Nonparametric Bayesian Clustering of Structural Whole Brain Connectivity in Full Image Resolution

Diffusion magnetic resonance imaging enables measuring the structural connectivity of the human brain at a high spatial resolution. Local noisy connectivity estimates can be derived using tractography approaches and statistical models are necessary to quantify the brain’s salient structural organization.

However, statistically modeling these massive structural connectivity datasets is a computational challenging task. We develop a high-performance inference procedure for the infinite relational model (a prominent non-parametric Bayesian model for clustering networks into structurally similar groups) that defines structural units at the resolution of statistical support.

We apply the model to a network of structural brain connectivity in full image resolution with more than one hundred thousand regions (voxels in the gray-white matter boundary) and around one hundred million connections. The derived clustering identifies in the order of one thousand salient structural units and we find that the identified units provide better predictive performance than predicting using the full graph or two commonly used atlases.

Extracting structural units of brain connectivity at the full image resolution can aid in understanding the underlying connectivity patterns, and the proposed method for large scale data driven generation of structural units provides a promising framework that can exploit the increasing spatial resolution of neuro-imaging technologies.