Unlearning for Clinical Deployment: Scalable Data Erasure in Medical Image Classification Models
Urbana Jaman Orthee, Abu Saleh Muhammad Naeem, Tamanna Khatun, Meharun Ohona, Nafees Mansoor
Conference Paper. 2025 IEEE International Conference on Biomedical Engineering, Computer and Information Technology for Health, BECITHCON 2025, pp. 279–284 (2025).
Abstract
Machine learning has become integral to modern healthcare. However, models often inherit dataset biases or outdated information that compromise generalizability and clinical reliability. Machine unlearning (MU) offers a means to selectively remove the influence of unwanted data cohorts without necessitating full retraining. This research introduces a sparsity-enhanced unlearning framework for medical image classification, featuring two proposed methods: Fine-Tuning Confusion (FT Confusion) and Gradient Ascent Repair (GA Repair), which are extensions of the standard Fine-Tuning and Gradient Ascent approach. Each method is evaluated across both class-level and instance-level forgetting tasks. Experimental results demonstrate that FT Confusion achieves retraining-level performance in class-wise forgetting, whereas GA Repair most closely approximates retraining in random instance-level scenarios. Furthermore, model sparsity consistently narrowed the gap between approximate unlearning and complete retraining, thereby improving efficiency and stability. These findings indicate that no single strategy is universally optimal; instead, a complementary portfolio is needed, with FT Confusion most effective for class-level bias mitigation and GA Repair for heterogeneous patient-level removal. Our contribution lies in enhancing two widely used MU baselines (fine-tuning and gradient ascent) with optimization-driven mechanisms, providing clearer theoretical grounding and improved reliability over existing approximate unlearning methods. Overall, this work highlights the feasibility of scalable unlearning in clinical workflows, supporting the development of adaptive, bias-aware, and regulator-ready medical AI systems. © 2025 IEEE.
Keywords
Deep Learning, Healthcare, Machine Unlearning, Model Sparsification, Optimization