Population-wise consistent segmentation of diffusion weighted magnetic resonance images

Abstract

In this thesis, we investigate unsupervised and semi-supervised methods to construct anatomical atlases and segment medical images. We propose an integrated registration and clustering algorithm to compute an anatomical atlas of fiber-bundles as well as deep gray matter structures from a population of diffusion tensor MR images (DT-MRI). We refer to this algorithm as "Consistency Clustering" since the outputs of the algorithm include population-wise consistent segmentations and correspondence between the subjects. The consistency is ensured through using a single anatomical model for the whole population, which is similar to the atlases used by experts for manual labeling. We experiment with both parametric and non-parametric models for the gray matter and white matter segmentation problems, each model resulting in a different kind of atlas. Consistent population-wise segmentations require development of several integrated algorithms for clustering, registration, atlas-building and outlier rejection. In this thesis we develop, implement and evaluate these tools individually and together as a population-wise segmentation tool. Together, Consistency Clustering enables automatic atlas construction in DT-MRI for a population, either normal or affected by a neural disorder. Consistency Clustering also provides the user the choice to include prior knowledge through a few labeled subjects (semi-supervised) or compute an anatomical atlas in a completely data driven manner (unsupervised). Furthermore, resulting anatomical models are compact representations of populations and can be used for population-wise morphometry. We implement and evaluate these methods using in vivo DT-MRI datasets. We investigate the benefits of population-wise segmentation as opposed to individually segmenting subjects, as well as effects of noise and initialization on the segmentations.