Longitudinal automated assessment of paramagnetic rim lesions in multiple sclerosis using RimNet

Wynen, Maxence [UCL] La Rosa Francesco [EPFL] Sellimi, Amina [UCL] Barquero Germán [EPFL] Perrotta Gaetano [ULB] Lolli Valentina [ULB] Van Pesch, Vincent [UCL] Granziera Cristina [University of Basel] Kober Tobias [Siemens Healthcare AG] Sati Pascal [National Institute of Health] Macq, Benoît [UCL] Reich, Daniel S. [National Institute of Health] Absinta Martina [National Institute of Health] Bach Cuadra Meritxell [Université de Lausanne] Maggi, Pietro [UCL] et al.[show all]

The automatic assessment of paramagnetic rim lesions in multiple sclerosis is important, and a deep learning-based algorithm called RimNet has recently been proposed. This work evaluates the generalizability of RimNet and its longitudinal performance on MRI data acquired at different clini-cal centers. We found that RimNet’s performance was nearly as good on totally unseen data as in the original paper (receiver-operating-characteristic area-under-the-curve (AUC) 0.88 vs. 0.94, pre-cision-recall AUC 0.69 vs. 0.70), and it made consistent predictions on longitudinal data (binary con-sistency 82%, probability consistency 93%).