Long-MS-Diff: Towards Generating Anatomically Plausible Lesion Progression in MS Imaging using Diffusion Models

Prateek Mathur, Brendan S Kelly, Ronan P. Killeen, Aonghus Lawlor

Keywords: Diffusion Models, Synthetic Medical Imaging, Multiple Sclerosis

TL;DR: Long-MS-Diff generates realistic MS follow-up MRIs using diffusion models conditioned on lesion features. It improves lesion fidelity with segmentation guidance and boosts downstream performance with synthetic data augmentation.

Abstract: Monitoring lesion progression in Multiple Sclerosis (MS) is vital for assessing disease activity and guiding treatment decisions. However, the limited availability of annotated longitudinal MRI data presents a challenge for developing robust deep learning models. We introduce Long-MS-Diff, a conditional diffusion-based framework for generating anatomically plausible follow-up brain scans in MS. The model is conditioned on baseline images, lesion change masks, and scalar lesion-level features. To address the extreme sparsity of new lesions, we incorporate an auxiliary segmentation task and propose an adaptive weighted loss to balance anatomical reconstruction with lesion-specific fidelity. A radiologist assessment of synthetic scans confirms high image quality, anatomical plausibility, and lesion adherence. In downstream segmentation experiments, moderate augmentation with Long-MS-Diff improves performance, outperforming models trained on real data alone. These results highlight the value of controlled synthetic data in modelling disease progression and demonstrate the utility of diffusion-based generation in data-scarce clinical settings.

Submission Number: 81