Improving Automated Quantification of MS Injuries with Deep Learning

Abstract

Multiple Sclerosis (MS) is a demyelinating disease of the central nervous system which affects nearly 1 million people in the US. Magnetic resonance imaging (MRI), as the most sensitive non-invasive technique, is used to visualize the disease course. In this dissertation, I present techniques to improve automated quantification of MS injuries based on MRI.

Currently, focal white matter (WM) lesions and cortical gray matter (cGM) injuries are considered important biomarkers for diagnosing MS. Among these, WM lesion quantification has attracted more attention and there are many proposed methods focusing on WM lesion segmentation. However, due to the difficulty of this task, there is no fully-satisfactory method available to the community. To improve the performance of MS lesion segmentation, I develop a deep neural network and propose a data representation named 2.5D stacked slices. This approach achieved state-of-the-art performance for MS lesion segmentation and is currently listed at the top of one MS lesion segmentation challenge leaderboard.

On the other hand, evaluation of cortical thinning, as one aspect of cGM injury quantification, can be improved by lesion inpainting. Towards this goal, I introduced the ``edge prior'' to guide the lesion inpainting process. With the edge prior, the deep networks are able to inpaint the images and better preserve the boundaries between tissues.

Further, to make the proposed methods practical for clinical use, I explored the robustness of these deep learning models. One aspect is the models' robustness to adversarial attacks. Another aspect is the models' performance change on unseen domains (e.g., different clinic sites). To generalize the models to unseen datasets, I use many categories of data augmentations during the training process, assuming these augmentations can fill the gap between domains. Also, to address MRI sequence variations across sites, I apply modality dropout and let the models learn to make reasonable predictions even when some modalities are missing at the test time. With all these efforts, we are able to provide a robust system for MS injuries quantification.