Computational Imaging Biomarkers of Multiple Sclerosis

Multiple Sclerosis (MS) is a chronic disease of the central nervous system characterized by disseminated neuroinflammation and neurodegeneration, and it is one of the primary causes of disability in young adults worldwide. The combination of neuroinflammation and neurodegeneration leads to the formation of multiple focal lesions and marked brain atrophy which also includes deep gray matter structures.

Magnetic resonance imaging (MRI) is more sensitive in detecting disease activity than clinical assessment, and it is the primary tool to detect cerebral lesions and morphological changes in the brain of MS patients. Manually labeling lesions and various brain structures is time consuming and prone to inter- and intra-rater variability.

Therefore there is a strong need for automatic tools that jointly segment white matter lesions and various neuroanatomical structures. Although in recent years a lot of progress has been made in the development of methods for the segmentation of brain structures, and in particular, the segmentation of white matter lesions, current approaches fall well below the threshold of what is required in the clinic.

This thesis describes segmentation methods that try to bridge the gap between research and clinical applications in MS. In the first part of this thesis, we developed a method for simultaneously segmenting various neuroanatomical structures and white matter lesions from MRI scans of MS patients. By using separate models for anatomical shapes and MRI appearance, the method is adaptive to differences in scanners and MRI sequences.

We validated the method on four datasets, showing robust performance in white matter lesion segmentation while simultaneously segmenting 41 brain structures. In the second part of this thesis, we extended this method to handle longitudinal scans. The method inherits all the properties of the cross-sectional method is built upon, and it has no requirements on the number and timing of longitudinal follow-up scans.

Furthermore, the method produces more reliable segmentations and detects disease effects better than its cross-sectional counterpart.