Precision Monitoring for Disease Progression in Patients with Multiple Sclerosis: A Deep Learning Approach

Abstract

Artificial intelligence has tremendous potential in a range of clinical applications. Leveraging recent advances in deep learning, the works in this thesis has generated a range of technologies for patients with Multiple Sclerosis (MS) that facilitate precision monitoring using routine MRI and clinical assessments; and contribute to realising the goal of personalised disease management.

MS is a chronic inflammatory demyelinating disease of the central nervous system (CNS), characterised by focal demyelinating plaques in the brain and spinal cord; and progressive neurodegeneration. Despite success in cohort studies and clinical trials, the measurement of disease activity using conventional imaging biomarkers in real-world clinical practice is limited to qualitative assessment of lesion activity, which is time consuming and prone to human error. Quantitative measures, such as T2 lesion load, volumetric assessment of lesion activity and brain atrophy, are constrained by challenges associated with handling real-world data variances. In this thesis, DeepBVC was developed for robust brain atrophy assessment through imaging synthesis, while a lesion segmentation model was developed using a novel federated learning framework, Fed-CoT, to leverage large data collaborations. With existing quantitative brain structural analyses, this work has developed an effective deep learning analysis pipeline, which delivers a fully automated suite of MS-specific clinical imaging biomarkers to facilitate the precision monitoring of patients with MS and response to disease modifying therapy. The framework for individualised MRI-guided management in this thesis was complemented by a disease prognosis model, based on a Large Language Model, providing insights into the risks of clinical worsening over the subsequent 3 years. The value and performance of the MS biomarkers in this thesis are underpinned by extensive validation in real-world, multi-centre data from more than 1030 patients.