Multi-subject Connectivity-Based Parcellation

Abstract

Connectivity has been proposed as a criterion for functional-anatomic segregation of cortical areas. In this thesis, I present a new method of characterizing the DTI-based connectivity profile of cortical voxels using Gaussian mixture models (GMMs). A variety clustering techniques are applied to perform the connectivity-based parcellation (CBP). I first parcellated the human inferior parietal lobule (IPL) on connectivity profiles using spectral clustering and a hidden Markov random field (HMRF) model. I applied our approach to multi-subject parcellation. I then segmented other cortical areas such as precentral and postcentral cortex, using spectral non-parametric Bayes models. A new approach resolving crossing fibers with compressed sensing (CS) was also examined. Using the multi-subject GMM-HMRF approach, results in a smoother segmentation of IPL that is independent of the set of subjects and visually consistent with the Juelich Atlas. The spectral non-parametric Bayes models enable data learn the number of segments automatically. The compressed sensing method is shown to significantly reduce the amount of data required and the computing time while preserving the accuracy.