OpenWell: Energy-Based Interactive Discovery for Open-World Medical Segmentation

Volumetric medical segmentation typically relies on closed-set assumptions, failing to identify or adapt to anatomical anomalies outside a fixed training vocabulary. We propose OpenWell, an energy-based framework that bridges the gap between Open-World discovery and interactive adaptation. Unlike deterministic approaches, OpenWell models anatomical classes as probabilistic energy wells on the hypersphere, using von Mises-Fisher (vMF) distributions. This formulation captures both the semantic identity and intra-class uncertainty of tissues, enabling out-of-distribution detection via a Free Energy score. Furthermore, we introduce an Instant Enrollment mechanism: triggered by a single user interaction, the system initializes a new energy well, allowing the model to adapt to previously unseen structures in real-time without gradient-based retraining. Extensive evaluations on AMOS, BTCV, MSD-Pancreas, and TotalSegmentator-MRI demonstrate that OpenWell improves Out-of-Distribution (OOD) detection AUROC over closed-set and foundation models, while achieving rapid, one-shot interactive adaptation to novel classes.