Adaptive feature unlearning for trustworthy medical imaging privacy

Deep learning has become integral to medical imaging, but its tendency to memorize training data poses serious risks for patient privacy. Machine unlearning offers a potential remedy by revoking sensitive information, yet existing approaches face three key limitations: (1) they often achieve only output-level changes while residual feature representations remain; (2) they rely on batch retraining, making real-time removal of individual patient images infeasible; and (3) they lack rigorous metrics to verify forgetting in feature space. We propose AdaptForget, a domain-adaptive feature-level unlearning framework for privacy-preserving medical image analysis. AdaptForget introduces out-of-distribution (OOD) guidance to disentangle forgotten data from retained data in the feature manifold, supported by a theoretical feature-level unlearning bound. To prevent feature collapse, we design an OOD-driven feature-output disentanglement loss that enforces structured removal of forgotten data. To enable timely revocation, we formalize the task of single-entry forgetting, allowing immediate erasure of individual patient records. For objective auditing, we propose the isolation verification distance, a novel metric that quantifies feature separation and provides interpretable evidence of forgetting. Extensive experiments on four medical imaging benchmarks (histopathology, retinal fundus, dermatology, and OCT) as well as complementary healthcare record datasets demonstrate that AdaptForget achieves state-of-the-art privacy protection while preserving model utility. Code is publicly available at https://github.com/wangbrav/AdaptForget.