Multi-scanner/multi-center evaluation of the progression of multiple sclerosis

Multiple Sclerosis (MS) is a chronic autoimmune disorder that affects the central nervous system. Over time, this condition can lead to severe disability, but a correct course of treatment can slow down the progression of the disease. In order to optimize treatment, MS patients undergo regular follow-ups, which include the acquisition of brain images through Magnetic Resonance Imaging (MRI). Additionally to typical MS lesions, these patients exhibit an increased rate of brain atrophy when compared to healthy subjects. Although brain atrophy has been linked to impairment, it can not be reliably estimated over short periods of time, which is why this biomarker is not commonly considered in clinical practice. One of the main causes for this is the high variability in the images caused by scanner- and center-specific factors, which influences the performance of the tools used to estimate brain volumes from structural MRI. For such measurements to be reliable, the current guidelines are to use the same MR scanner when following-up a patient. Nevertheless, in real clinical conditions patients are often scanned in different machines over the course of their lives, which hinders the reliability of a quantitative analysis of disease progression.
In the present thesis we describe three different approaches to deal with this scannerbias problem. In a first stage, we explore the relationships between image-extracted information and volume estimations and use the findings to develop a statistical harmonization method. Since this is not sensitive enough for single patient analysis, we move our focus to the improvement of the underlying method for brain tissue and structure segmentation. To this end, we propose two alternative approaches: (i) a tissue-specific augmentation of MRI images for segmentation of brain structures using Deep Learning (DL) and (ii) adversarial training of a DL network to promote scanner invariance. After careful analysis, we consider that methods which promote the generalization capabilities of the volume estimation models are the most promising for future research, given their simplicity and good performance on the task of brain segmentation.