Segmentation of magnetic resonance images of the brain
Simon Keith Warfield · UNSWorks (University of New South Wales, Sydney, Australia) · 2022
The goal of this thesis was to develop computer methods for the automatic segmentation of medical images. Focus was placed upon the segmentation of magnetic resonance images (MRI) of the human brain. Interest in this area has been stimulated by the potential of clinical applications and the possibility for insights into the structure and functioning of the brain. Global segmentation strategies have been widely used for the identification and delineation of brain structures from MRI. Pattern recognition techniques have been applied to the segmentation of tissue types, and methods utilising explicit anatomical models to the segmentation of gross subcortical anatomical structures. A fast and accurate k-Nearest Neighbour algorithm for the classification of multichannel image data was developed. Fast classification is achieved by computing a classification lookup table and taking advantage of the similarity of patterns in the table with a new distance transform algorithm. The k Distance Transform algorithm extends existing distance transform algorithms, which compute a map of the nearest neighbour distance, to allow the computation of a map of the k nearest neighbour distances in an efficient manner. A general conceptual framework of applying local optimisation strategies in an hierarchical iterative segmentation scheme was developed. A segmentation methodology capable of identifying and delineating tissue classes and gross anatomical structures together with abnormalities was developed. This led to the development of a successful algorithm for the segmentation of MRI scans of patients with multiple sclerosis. Validation experiments with real and simulated MRI data are presented and show that the algorithm performs well. The concept of applying local optimisations with global segmentation strategies has broad application in the medical imaging field. The contributions of this thesis are a new distance transform algorithm, a new algorithm for fast k-Nearest Neighbour classification of multichannel image data and a conceptual framework emphasising local optimisation strategies with which to approach new segmentation problems. This framework has been used to construct novel segmentation algorithms which achieve both tissue classification (after resolving the difficulty of overlapping tissue class intensity ranges) and the segmentation of anatomical structures, despite the presence of abnormalities.