To forget is to preserve: Machine Unlearning for 3D medical image segmentation
Researchers have explored machine unlearning techniques to comply with data privacy regulations like GDPR, which allow individuals to request data removal from trained models. A study applied four unlearning strategies to the MRBrainS18 dataset using a 3D ResNet-50 backbone pre-trained with Med3D. The "Noisy Label" strategy demonstrated the best balance, reducing data in the forget set by 93% while retaining 84% accuracy on the retained set after 50 epochs, outperforming other methods that caused significant performance degradation. AI
IMPACT This research provides a baseline for subject-specific unlearning, offering clear criteria for practitioners to select appropriate strategies for data privacy compliance in AI models.