End-to-End Patch-Based Semantic Segmentation for Paramagnetic Rim Lesion Detection

Zineb El yamani, Antoine Theberge, Manon Edde, Olivier Grimard, Emmanuelle Lapointe, Stefano Magon, Muhamed Barakovic, Maxime Descoteaux, Francois Rheault

Keywords: Multiple Sclerosis, Chronic Active Lesions, Paramagnetic Rim Lesions, Segmentation, Magnetic Resonance Imaging, Quantitative Susceptibility Mapping

TL;DR: We introduce a patch-based multiclass segmentation framework for whole-brain detection of paramagnetic rim lesions in multiple sclerosis, removing the need for lesion-centered preselection and achieving fair performance in a low-data setting.

Abstract: Paramagnetic Rim Lesions (PRLs) are an emerging biomarker of chronic active inflammation in Multiple Sclerosis (MS) but their visual identification on susceptibility-sensitive MRI remains challenging and time-intensive. Due to the scarcity of PRLs, existing automated methods rely on patch-based classification, where a lesion-centered 3D patch is classified as PRL or non-PRL. However, MS lesions often occur in clusters, so a single patch may contain multiple types of lesions. Moreover, this approach requires prior extraction of lesion-centered patches which complicates the reconstruction of whole-brain predictions. To overcome this, we propose an end-to-end, whole-brain pipeline that generates patches on the fly and directly delineates PRLs within them, eliminating the need for lesion-centered extraction and enabling more precise and user-friendly automated PRL detection.

Submission Number: 90