Towards Automatic Spinal Cord MRI Analysis for Improved Estimation of Imaging Biomarkers

Automatic image analysis of the spinal cord offers key insights into its morphology, enabling consistent characterization of anatomical variations across populations and facilitating the diagnosis and monitoring of neurological diseases. Morphometric measures derived from spinal cord magnetic resonance imaging (MRI) scans serve as key biomarkers in improving our understanding of the pathophysiology of various diseases. Segmentation of the spinal cord and lesions is a crucial prerequisite for extracting clinical biomarkers. However, the inherent challenges associated with spinal cord imaging, namely, the small physical structure of the cord, its susceptibility to noise and motion artifacts combined with partial volume effects make automatic segmentation extremely challenging. Consequently, existing tools lack robustness and are unable to generalize, suggesting the need for automatic tools that can segment the spinal cord across several MRI contrasts and pathologies. The overarching goal of this thesis is to develop generalizable, automatic tools for the robust segmentation of the spinal cord and lesions across various MRI contrasts and pathologies for better estimation of imaging biomarkers. This goal is spread across three contributions: (i) a contrast-specific method for segmenting intramedullary lesions in spinal cord injury, (ii) a more generalist, contrast-agnostic approach for spinal cord segmentation, and (iii) a framework designed to facilitate continuous training of spinal cord segmentation models.