Deep learning automates paramagnetic rim lesion detection on QSM in multiple sclerosis

QSM-only deep learning detected paramagnetic rim lesions in MS with 72.4% sensitivity (vs 62.1% for QSM+FLAIR) on internal test; automated PRL burden linked to cognitive impairment (p=0.002). Retrospective study.

• Retrospective, single-center study: 106 patients for model development (84 exploration [54 train, 30 internal test] + 22 temporal test), plus 117 for clinical correlation. Manual segmentation as ground truth.
• In temporal test set, lesion-level sensitivity was 45.3% (QSM-only) vs 54.7% (QSM+FLAIR) (p=0.181); both models achieved 100% patient-level sensitivity.
• Key limitation: small test sets (n=30 internal, n=22 temporal), no external multi-center validation, retrospective design.

RadPigeon summaries are original and for information only. They are not clinical advice.