Longitudinal white matter hyperintensity detection in multiple sclerosis

Multiple sclerosis (MS) is a degenerative neurological disease affecting the central nervous system. It is characterized by an abnormal immune response resulting in inflammation and progressive damage to myelin, a protective tissue covering the nerve fibers, which is essential for signal transmission. Demyelinated areas in the brain appear as hyperintensities in T2-weighted magnetic resonance (MR) images and are radiologically described as white matter hyperintensities (WMH). At present, the most important radiological diagnostic criterion for MS is based on the identification of these lesions in MR images while follow-up of MS patients primarily consists of comparing MR images of the brain with prior MR examinations. Performing a manual assessment of the prior and current image for the identification and quantification of white matter lesions is time-consuming for the radiologist and the manual assessment is known to exhibit a significant inter-rater and intra-rater variability. For these reasons there is an increasing need for automated tools that are fast, while providing reproducible and accurate measures. The primary aim of this thesis was to develop an automated algorithm that can accurately detect changes in WMH (i.e., new and enlarged lesions) with a high level of sensitivity and within a short time. Secondly, the feasibility of using only a single image contrast instead of two contrasts for lesion segmentation was investigated. Lastly, the viability of using automatically generated weak labels as a reference for training a deep neural network was addressed. The proposed solution was developed using a deep learning approach and implemented in three steps: 1) generation of a brain mask, 2) segmentation of white matter lesions cross-sectionally, and 3) detection and segmentation of new and enlarged lesions in follow-up MR scans. The models for each step were trained primarily on weak labels generated by existing algorithms, due to the limited availability of manual segmentations. To further improve results, fine-tuning on manually corrected labels (when available) was subsequently performed. Several data preprocessing steps and data augmentation routines were implemented, and different network architectures and training approaches were investigated. The brain segmentation model was not only used as a precursor for succeeding processing steps, but also provided pretrained weights for the cross-sectional lesion segmentation. The performance of the algorithm was comparable to the best ranked solutions proposed in public challenges. This thesis showed that a single MS contrast contains sufficient information to provide WMH segmentations and detect changes in lesions. In addition, the use of weak labels to create robust algorithms which can provide accurate WMH segmentation was shown. Special attention was paid to the applicability in a real environment. The achieved inference time of the developed algorithms of about 4 s ensures that a clinical application will be possible in the future.