Multimodal Deep Learning Facilitates Automatic Identification of Chronic Active Lesions in Multiple Sclerosis

This article introduces a novel framework based on 3D multimodal deep learning that enables automatic classification of paramagnetic rim lesions (Rim⁺) in multiple sclerosis (MS) via asymmetric QSM-FLAIR modeling. It aims to address the challenges of manual Rim⁺ lesion detection and provide an automated solution for the detection of biomarkers for chronic active inflammation.

• Original Authors/Institution: This article is from an arXiv preprint, and the author team focuses on medical image analysis and the application of deep learning in neurological disease diagnosis
• Source Platform: arXiv
• Original Title: 3D Classification of Paramagnetic Rim Lesions in Multiple Sclerosis via Asymmetric QSM-FLAIR Modeling
• Original Link: http://arxiv.org/abs/2606.16756v1
• Publication Date: June 15, 2026

Multiple sclerosis (MS) is a chronic inflammatory demyelinating disease of the central nervous system. Paramagnetic rim lesions (Rim⁺), as biomarkers for chronic active inflammation, are closely associated with long-term disability progression in patients. However, Rim⁺ lesion detection faces challenges: it requires professional equipment and interpretation by experienced physicians, manual assessment is time-consuming with large inter-observer variability, and the low proportion of Rim⁺ lesions leads to class imbalance issues, making automated analysis difficult.

The research team proposed an innovative 3D multimodal deep learning framework with core features including:

1. Asymmetric Multimodal Modeling: Uses two modalities—Quantitative Susceptibility Mapping (QSM) and FLAIR MRI. QSM serves as a susceptibility-driven signal, while FLAIR provides structural context constraints, fully leveraging their different roles in lesion characterization;
2. Self-Supervised Pre-training and Contrastive Learning: Learns modality correspondence through pre-training on large-scale unlabeled data to extract generalized features, then fine-tunes with limited labeled data and introduces contrastive regularization to enhance class discrimination ability.

The study was validated on a clinical cohort of 88 MS patients, with expert annotations as the gold standard. The results show:

• The asymmetric multimodal modeling method significantly improved performance in Rim⁺/Rim⁻ classification tasks compared to existing architectures;
• Effectively alleviated class imbalance issues, maintaining high sensitivity while controlling the false positive rate;
• The automated results showed good consistency with expert assessments, demonstrating potential for clinical auxiliary diagnosis.

This study provides a new tool for precise diagnosis and prognosis evaluation of MS. By automatically identifying chronic active lesions, it helps clinicians more accurately assess disease activity and adjust treatment plans in a timely manner. Additionally, the method is expected to be extended to image analysis of other neurodegenerative diseases, promoting the development of medical imaging AI.

Future research directions include: expanding the scale of validation cohorts, exploring more MRI sequence fusion strategies, and developing real-time analysis systems to support clinical workflows.