Representation learning for multi-source MRI-based prediction of disability progression in multiple sclerosis

Despite the availability of detailed MRI acquisitions and clinical variables, predicting short-term disability in patients with multiple sclerosis (MS) remains difficult. This limitation prevents clinicians from offering a reliable prognosis and may delay access to the most appropriate therapeutic options. Current prognostic approaches rely mainly on predefined imaging biomarkers, such as lesion volume, lesion count, or regional atrophy measures. While these markers are clinically relevant, their correlation with disability progression is often only partial (Barkhof, 2002). Reducing complex MRI data to summary features inevitably discards information regarding spatial organization, lesion morphology, and subtle tissue alterations that may carry additional prognostic value (Storelli et al., 2022).

This PhD project proposes to reformulate early disability worsening prediction as a multimodal representation learning problem. The central hypothesis is that learning structured latent representations directly from raw brain and spinal cord MRI data, jointly integrated with relevant non-imaging variables (e.g., age, sex, genetic characteristics, physical condition), may improve prediction of disability progression compared with traditional biomarker-based approaches.

Several methodological challenges will be addressed. First, learning a direct mapping from raw MRI (potentially including repeated timepoints when available) to short-term disability outcomes is challenging because the imaging input is high-dimensional and highly variable, whereas clinical endpoints (e.g., EDSS-based worsening) provide a compact and coarse summary of progression. To address this, the project will investigate alternative training strategies, comparing direct end-to-end supervised learning to self-supervised pretraining followed by outcome-driven fine-tuning. Self-supervised learning provides a principled framework for pretraining feature extractors without manual annotations (Chen et al., 2020; He et al., 2022; Balestriero et al., 2023), and is well suited to leverage large MRI cohorts. Second, combining brain and spinal cord MRI introduces a representation alignment challenge due to differences in anatomy, field of view, resolution, and acquisition artifacts. The project will develop dedicated encoders for each anatomical domain and explore latent-space fusion mechanisms to integrate information coherently. Third, multimodal fusion strategies will be studied to integrate imaging with clinical variables while ensuring balanced use of each modality. Finally, interpretability analyses will be performed to relate learned representations and predicted outcomes to anatomically meaningful patterns.

The project will rely on large clinical cohorts, including OFSEP data with longitudinal MRI and disability assessment. It aims to contribute to methodological development in multimodal representation learning for medical imaging and to support more accurate prognostic modeling in MS.