Machine Unlearning Risks in Medical Image Classification

A recent study published on arXiv (2604.23854) investigates the implications of machine unlearning on clinical safety within binary medical image classification. The researchers discovered that conventional unlearning techniques—Fine-Tuning, Random Labeling, and SalUn—tend to diminish test utility and elevate false-negative rates, thereby increasing clinical risk. To mitigate this issue, they introduce SalUn-CRA (Clinical Risk-Aware), which substitutes random relabeling with an entropy-based forgetting mechanism for malignant samples in the forget set. Testing on the DermaMNIST and PathMNIST datasets validates their method. This research underscores the importance of reconciling data protection regulations with patient safety in deep learning applications for medical diagnosis.