Attention-based Multiple Instance Learning for Classification of Inflammatory Activity in Crohn’s Disease Patients from MR Enterography

This study proposes a deep learning framework that combines a foundation model, DINOv2, with an attention-based multiple instance learning (MIL) strategy to automatically assess inflammatory activity in Crohn’s disease from magnetic resonance enterography (MRE) images. The proposed method treats lesion-level MRE patches as multiple instances and employs DINOv2 as a feature extractor, followed by a Top-K attention pooling strategy to focus on clinically relevant lesions and derive patient-level sMaRIA scores. Experimental validation on a dataset of 208 patients demonstrated that the model integrating DINOv2-Small with attention-based pooling and Top-K instance selection (K = 3) achieved the highest classification accuracy of 77.78%. Qualitative results further confirmed that the model effectively attends to clinically meaningful inflammatory regions.