In multiple sclerosis, lesions with a peripheral paramagnetic rim is a negative prognostic imaging biomarker and represents a potential outcome measure in MRI-based clinical trials. Nowadays, the presence or absence of paramagnetic rims is determined through visual inspection by medical experts, which is tedious, time consuming and prone to observer variability. So far, few solutions to the automated classification of rims problem have been proposed. These studies present limitations that represent an obstacle to full automation of the rim analysis process and its large-scale application. Our goal is to implement and assess a fully automated algorithm capable of identifying rim lesions in MRI. In this work, we explore a Region-proposal CNN deep learning approach to solve the fully automated rim lesions classification problem that perform instance segmentation by object detection and have shown promising results in recent challenges, particularly in medical imaging. After different approaches focus on implifying the task, Mask RCNN with MobileNet v2 as backbone using attention gaussian filtering to the input images showed better performance than the rest with rates of 0.42 TPR and 0.61 FPR for the test set. However, the achieved results reveal the weaknesses of our approach and the difficulty of our classification problem.