Deep-PRL: a deep learning network for the identification of paramagnetic rim lesions in multiple sclerosis

Motivation: PRLs are an important diagnostic biomarker in people with multiple sclerosis (pwMS). Their identification on MRI is time-consuming and subject to high inter-rater variability. However, the use of AI could support this identification process. Goal(s): We leverage multi-contrast MRI to improve the identification of PRLs. Approach: Deep-PRL is an attention-based CNN, fusing features of T1-w and unwrapped phase images from 185 pwMS. The approach consists in a nested cross-validation with patient stratification. Results: The test performance outperformed state-of-the-art methods, achieving a mean F1 score of 0.860 ± 0.048 and an AUC of 0.982 ± 0.007. Impact: These results represent a significant step towards the integration of an AI tool to assist clinicians in the identification of PRLs, thereby improving the clinical management of pwMS.