Automated brain lesion segmentation in magnetic resonance images

In this thesis, we investigate the potential of automation in brain lesion segmentation in magnetic resonance images. We first develop a novel supervised method, which segments regions in magnetic resonance images using gated recurrent units, provided training data with pixel-wise annotations on what to segment is available. We improve on this method using the latest technical advances in the field of machine learning and insights on possible weaknesses of our method, and adapt it specifically for the task of lesion segmentation in the brain. We show the feasibility of our approach on multiple public benchmarks, consistently reaching positions at the top of the list of competing methods. Adapting our problem successfully to the problem of landmark localization, we show the generalizability of the approach. Moving away from large training cohorts with manual segmentations to data where it is only known that a certain pathology is present, we propose a weakly-supervised segmentation approach. Given a set of images with known pathology of a certain kind and a healthy reference set, our formulation can segment the difference of the two data distributions. Lastly, we show how information from already existing lesion maps can be extracted in a meaningful way by connecting lesions across time in longitudinal studies. We hence present a full tool set for the automated processing of lesions in magnetic resonance images.