To forget is to preserve: Machine Unlearning for 3D medical image segmentation

Researchers have benchmarked machine unlearning techniques for 3D medical image segmentation, a step toward complying with data privacy regulations that grant individuals the right to have their information erased from trained models. The study, submitted to arXiv on June 15, 2026, evaluates several approximate unlearning strategies applied to the MRBrainS18 dataset [1]. The work is driven by laws such as the General Data Protection Regulation (GDPR), which allows individuals to request the deletion of their personal data from machine learning models [1][2]. The authors use a 3D ResNet-50 as the backbone architecture for segmentation, pre-trained with the Med3D framework [1][2]. Researchers assessed performance using the Dice similarity coefficient and mean absolute error across two training horizons: 20 and 50 epochs [1]. The Noisy Label strategy delivered the strongest overall trade-off. After 50 epochs, it achieved a 93% decrease in the forget set while maintaining 84% accuracy for the retained set [1][2]. Other strategies showed extreme forgetting at higher epoch counts but also suffered catastrophic degradation of retain-set performance [1]. The paper provides a strict baseline of performance metrics for unlearning on a subject-specific level [1][2]. Nitesh Kumar Singh is listed as the corresponding author on the submission [1]. The research appears on arXiv, a preprint server that has integrated interactive demos through a collaboration with Hugging Face Spaces to make machine learning papers more accessible [3][4]. This integration allows users to find open-source demos linked to papers in computer science, statistics, and electrical engineering categories [4][5]. Machine unlearning addresses a growing tension between the capabilities of large-scale models and individual privacy rights. Large language models, for instance, are trained with self-supervised learning on vast amounts of text [7]. The field has seen rapid commercial development, with companies such as DeepSeek releasing open-weight models under permissive licenses [6]. Contextual AI, co-founded by former Hugging Face Head of Research Douwe Kiela, focuses on building grounded AI agents for enterprise knowledge bases [8]. The unlearning benchmarks established in the new study offer practitioners clear criteria for selecting strategies when deletion requests arrive [1].