Magnetic resonance imaging (MRI) is commonly used for the diagnosis and prognosis of multiple sclerosis (MS). Analysis and extraction of MRI information could be performed manually by radiologists or domain experts, nevertheless, these tasks are tedious, time consuming, require domain expertise, and are subject to inter-rater variability. Thus, automation of MRI analysis tasks has been considered to address these limitations and to handle the large amount of data we encounter in the Big Data era. In this thesis, we propose end-to-end pipelines using deep learning to analyze MRI and extract information relevant to MS. The suite of tools includes MS lesion segmentation, new lesion segmentation/detection, and disability status estimation (EDSS) from MRI and clinico-demographic data. In designing each pipeline, we proposed methodological contributions that addressed various technical challenges, such as domain bias, data sparsity and data imbalance.
Our pipelines are hosted on volBrain, to make them freely and easily usable by the MS community without the need for software or hardware. In doing so, our users benefit from state-of-the-art performance with just a few clicks on their web browser and get a compact, easy-to-read report summarizing the results.