Novel mathematical methods for analysis of brain white matter fibers using diffusion MRI

Official URL: http://risc01.sabanciuniv.edu/record=b1817050 (Table of Contents)

White matter fibers connect and transfer information among various gray matter regions of the brain. Diffusion Magnetic Resonance Imaging (DMRI) allows in-vivo estimation of fiber orientations. From the estimated orientations, a 3D curve representation of the trajectory of fibers can be reconstructed in a process known as tractography. Automatic classification of these \tracts" into classes of anatomically known fiber bundles is a very important problem in neuroimage computing. In this thesis, three automatic fiber classification methods are proposed. The first two are based on combining neuroanatomical priors with density-based clustering. The first method includes brainstem heuristics but the second is more general and can be applied to any fiber pathway in the brain. Further, the second method introduces a novel fiber representation, Neighborhood Resolved Fiber Orientation Distribution(NRFOD), that represents a tract as a set of histograms that encode the distribution of fiber orientations in its neighborhood. The third method utilizes the NRFOD representation to directly map a tract to a probability estimate for each bundle class in a supervised classification framework. A practical training and validation set creation methodology is proposed. Additionally, the thesis includes statistical significance tests to investigate whether the structural change between pre-operative and post-operative fiber bundles after a tumor resection operation are related to the change in patient's cognitive performance scores. To this end, a fiber bundle to fiber bundle registration method and various quantitative measures of the structural change are proposed. We present results over DMRI data with clinical evaluations of 30 patients with brainstem tumors.