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

2026-06-15 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision & Pattern Recognition, Health & Medical Research · Depth: Expert, quick

Summary

A study on machine unlearning for 3D medical image segmentation addresses compliance with data privacy laws like GDPR, which mandate personal data erasure from trained models. Researchers investigated several approximate unlearning strategies, based on four mechanics, applied to the MRBrainS18 dataset. They utilized a 3D ResNet-50 backbone, pre-trained with the Med3D framework, for segmentation. Evaluation involved assessing retention accuracy on "retain" and "forget" subjects using Dice similarity coefficient and mean absolute error (MAE) values over 20 and 50 training epochs. The Noisy Label strategy demonstrated the optimal trade-off, achieving a 93% decrease in the forget set while maintaining 84% accuracy for the retained set after 50 epochs. Other strategies showed significant forgetting but also catastrophic degradation in retained set performance at higher epoch numbers. This research establishes a strict baseline for subject-specific unlearning performance.

Key takeaway

For Machine Learning Engineers developing 3D medical image segmentation models requiring GDPR compliance, you should prioritize evaluating the Noisy Label unlearning strategy. This approach demonstrated a 93% decrease in forgotten data while retaining 84% accuracy on preserved data after 50 epochs, offering a robust trade-off. Establish strict performance baselines for subject-specific unlearning to ensure both privacy and model utility are met effectively in your deployments.

Key insights

Effective machine unlearning balances data erasure with preserving model utility for retained information.

Principles

• Unlearning strategies often degrade retained data performance.
• Noisy Label strategy offers a strong unlearning-retention trade-off.

Method

Approximate unlearning strategies were applied to a Med3D pre-trained 3D ResNet-50 on MRBrainS18, evaluated by Dice and MAE across 20/50 epochs.

In practice

• Consider Noisy Label strategy for medical image unlearning.
• Establish performance baselines for subject-specific unlearning.

Topics

• Machine Unlearning
• 3D Medical Imaging
• GDPR Compliance
• ResNet-50
• Noisy Label Strategy
• Data Privacy

Best for: CTO, VP of Engineering/Data, Director of AI/ML, AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• PURGE: Projected Unlearning via Retain-Guided Erasure — PURGE unlearns data by projecting gradients to preserve retain-set performance and erasing hidden representations towards retain distribution.
• Fairness Beyond Demographics: Optimizing Performance Across Appearance-Based Hidden Cohorts in Medical Imaging — Optimizing fairness across appearance-based hidden cohorts improves medical image analysis models without relying on demographic labels.
• Learning What to Forget: Improving LLM Unlearning via Learned Token-Level Importance — Effective LLM unlearning requires implicitly identifying and selectively targeting forget-specific tokens based on their conflict with retaining general capabilities.
• Class Unlearning via Depth-Aware Removal of Forget-Specific Directions — DAMP offers a one-shot, closed-form weight-surgery for class unlearning by removing forget-specific directions.
• Forgetting That Sticks: Quantization-Permanent Unlearning via Circuit Attribution — Quantization can reverse machine unlearning, necessitating methods that ensure forgetting persists post-compression.
• SHRED: Retain-Set-Free Unlearning via Self-Distillation with Logit Demotion — LLM unlearning can be retain-set-free by selectively demoting high-information tokens via self-distillation, preserving general utility.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.