Diffusion-based population statistics using tract probability maps
Résumé
We present a novel technique for the tract-based statistical analysis of diffusion imaging data. In our technique, we represent each white matter (WM) tract as a tract probability map (TPM): a function mapping a point to its probability of belonging to the tract. We start by automatically clustering the tracts identified in the brain via tractography into TPMs using a novel Gaussian process framework. Then, each tract is modeled by the skeleton of its TPM, a medial representation with a tubular or sheet-like geometry. The appropriate geometry for each tract is implicitly inferred from the data instead of being selected a priori, as is done by current tract-specific approaches. The TPM representation makes it possible to average diffusion imaging based features along directions locally perpendicular to the skeleton of each WM tract, increasing the sensitivity and specificity of statistical analyses on the WM. Our framework therefore facilitates the automated analysis of WM tract bundles, and enables the quantification and visualization of tract-based statistical differences between groups. We have demonstrated the applicability of our framework by studying WM differences between 34 schizophrenia patients and 24 healthy controls.