Exact Patient-Level Unlearning Without Retraining
One-U RAI Seed Grant project developing MRI models for rare clival tumors that can remove a patient's contribution exactly—without full retraining—when consent is withdrawn.
This project is a One-U RAI Seed Grant Award. Full title: Responsible AI for Rare Clival Tumors: Exact Patient-Level Unlearning Without Retraining. Co-PI: Tyler Richards, Neuroradiology.
As medical AI moves toward clinical use, patients may later withdraw consent for their data. Removing cases from a stored dataset is not enough: once a model is trained, patient-specific information remains encoded in its parameters, and full retraining to honor withdrawal is often impractical. This gap between ethical and legal expectations and technical reality is especially acute for rare diseases, where multi-institutional data sharing is essential and re-identification risk is high.
We are developing an unlearning-capable modeling framework for MRI of rare clival (central skull base) tumors. Distinguishing primary neoplasms from secondary lesions related to systemic malignancy is clinically critical: primary tumors are typically managed with surgery, while secondary tumors are treated with radiation and chemotherapy. An accurate AI classifier could reduce unnecessary biopsies, avoid costly systemic imaging, and speed appropriate care—but building such models requires collaborative datasets that patients and institutions will trust.
Rather than a single monolithic parameter set, the approach represents the deployed model as a fixed core trained on non-revocable data, plus compact per-patient parameter contributions that can be aggregated into the final predictor. Withdrawing consent then corresponds to removing that patient's contribution from the aggregate—exact, auditable, and without full retraining. The work will evaluate tumor segmentation and classification performance alongside measurable privacy protection, establishing a framework for adaptable and accountable medical AI.
