Bayesian clustering methods for morphological analysis of MR images

Determining the relationship between structure (i.e. morphology) and function is a fundamental problem in brain research. In this paper we present a new framework based on Bayesian clustering methods for the voxel-wise statistical morphology-function analysis of registered MR images. We construct a Bayesian network to automatically identify the significant associations between voxel-wise morphological variables and functional variables, such as cognitive performance. A Bayesian latent variable induction method is applied to locate the homogeneous association regions on registered maps of morphological variables. Experimental results on images with simulated atrophy confirm that the new method outperforms conventional statistical method, based on linear statistics.