MedCRP-CL beats catastrophic forgetting with 73.3% Dice on 16 medical tasks

Medical image segmentation faces a fundamental continual learning challenge: data arrives sequentially from heterogeneous sources, and models often forget earlier tasks when learning new ones. Existing approaches either apply uniform constraints across all tasks (causing forgetting when tasks conflict) or require predefined task groupings that cannot anticipate future diversity. MedCRP-CL, introduced by Ziyuan Gao and accepted at ICML 2026, solves this by dynamically discovering task structure on the fly. It leverages the Chinese Restaurant Process (a Bayesian nonparametric method) to infer semantic modalities from clinical text prompts as tasks arrive, without needing to know the number of groups in advance. These modalities capture finer-grained structure than imaging modalities alone by combining anatomical regions and pathological context. The framework then maintains a set of semantic modality-specific LoRA adapters, each regularized by intra-modality Elastic Weight Consolidation (EWC), ensuring parameters are isolated across dissimilar groups while allowing knowledge transfer within similar ones. Notably, MedCRP-CL is replay-free, storing only aggregate statistics rather than raw patient data.

Experiments on 16 diverse medical segmentation tasks across four imaging modalities demonstrate the framework’s effectiveness. MedCRP-CL achieves a 73.3% Dice score with only 4.1% forgetting, outperforming the best baseline by 8.0% in absolute Dice while requiring 6× fewer parameters. This efficiency comes from its dual strategy of grouping tasks intelligently and applying lightweight adapter modules that are trained only on data from their assigned cluster. The framework also respects privacy constraints by eliminating the need to store or replay previous examples. The code is publicly available, and the paper provides a strong foundation for deploying continual learning in clinical settings where new imaging protocols and pathologies appear regularly. For professionals building AI systems that must adapt over time without degrading performance, MedCRP-CL represents a practical, scalable approach to lifelong medical image analysis.