Machine Learning Methods for Brain Lesion Delineation

Abstract

Brain lesions are regions of abnormal or damaged tissue in the brain, commonly due to stroke, cancer or other disease. They are diagnosed primarily using neuroimaging, the most common modalities being Magnetic Resonance Imaging (MRI) or Computed Tomography (CT). Brain lesions have a high degree of variability in terms of location, size, intensity and form, which makes diagnosis challenging. Traditionally, radiologists diagnose lesions by inspecting neuroimages directly by eye; however, this is time-consuming and subjective. For these reasons, many automated methods have been developed for lesion delineation (segmentation), lesion identification and diagnosis. The goal of this thesis is to improve and develop automated methods for delineating brain lesions from multimodal MRI scans. First, we propose an improvement to existing segmentation methods by exploiting the bilateral quasi-symmetry of healthy brains, which breaks down when lesions are present. We augment our data using nonlinear registration of a neuroimage to a reflected version of itself, leading to an improvement in Dice coefficient of 13 percent. Second, we model lesion volume in brain image patches with a modified Poisson regression method. The model accurately identified the lesion image with the larger lesion volume for 86 percent of paired sample patches. Both of these projects were published in the proceedings of the BIOSTEC 2020 conference. In the last two chapters, we propose a confidence-based approach to measure segmentation uncertainty, and apply an unsupervised segmentation method based on mutual information.