Deep-learning-based image segmentation for uncommon ischemic stroke From infants to adults

Developing deep learning-based algorithms that accurately segment structures in scans that are relevant to treatment or evaluation of the outcome of uncommon stroke is a difficult task. The difficulty is due to the presence of image artefacts, few data being available to train the networks, and the small volume of some of the target structures. Hence, the aim of this thesis was to investigate, develop, and evaluate deep learning-based algorithms for automatic segmentation of images of uncommon sub-types of stroke.
In chapter two, transfer learning strategies for automated medical image segmentation were evaluated. Our results showed that pre-training on a segmentation task on the same domain as the target segmentation task yielded the greatest improvement in spatial agreement. However, our results have also shown that the choice of source task and domain have an inconsistent effect on the detection rate.
In chapters three and four, segmentation algorithms for scans of patients suffering from posterior circulation stroke were developed. In chapter three, deep transfer learning was used to improve segmentation of lesions caused by posterior circulation stroke. In chapter four an algorithm, which restricted inference to the area surrounding the brain stem, was developed to segment thrombi in the posterior circulation.
In chapter five, two instances of an algorithm were developed to segment brain tissue types and the ischemic lesion per hemisphere in patients suffering from perinatal arterial ischemic stroke. One instance segmented scans acquired at term, the other instance segmented scans acquired at follow-up.