segmentation-unlearning

Application of machine unlearning to image segmentation tasks, particularly in medical imaging. Addresses the problem of models learning spurious correlations (shortcuts) from coarse segmentation masks by formally connecting mask refinement to the “forgetting” of dilation artefacts. Extends certified unlearning from classification to pixel-level predictions, enabling provable reduction of shortcut learning without expensive fine-grained annotations.

Papers Analyzed

Paper	Year	Key Contribution
Towards Certified Shortcut Unlearning in Medical Imaging	2026	First bridge between segmentation refinement and certified unlearning; certified pixel-level unlearning, global spurious mutual information
Certified Unlearning for Neural Networks	2025	(epsilon,delta)-certified gradient clipping for unlearning and model clipping for unlearning algorithms via privacy amplification by iteration
Reducing Reliance on Spurious Features in Medical Image Classification with Spatial Specificity	2022	spatial specificity framework showing finer annotations reduce shortcut reliance
A Case for Reframing Automated Medical Image Classification as Segmentation	2023	segmentation-for-classification with information-theoretic justification
A Survey of Machine Unlearning	2025	Comprehensive taxonomy of exact unlearning and approximate unlearning methods
Remember What You Want to Forget Algorithms for Machine Unlearning	2021	Foundational certified approximate unlearning definition, deletion capacity, Newton-step unlearning
Towards Unbounded Machine Unlearning	2023	SCRUB and NegGrad+ algorithms for practical unlearning
Towards Certified Unlearning for Deep Neural Networks	2024	local convex approximation for certified unlearning + LiSSA for inverse Hessian approximation for DNNs
Certified Machine Unlearning with Randomized Gradient Smoothing and Quantization	2022	randomized gradient smoothing for Hessian-free certified unlearning
Certified Machine Unlearning via Noisy Stochastic Gradient Descent	2024	PNSGD for certified unlearning with W-infinity distance tracking for unlearning
Debiasing Skin Lesion Datasets and Models Not So Fast	2020	Systematic analysis of artefact biases in dermoscopy datasets

Key Concepts and Connections

The central theoretical chain is:

1. spatial specificity (Saab 2022): Finer annotations reduce I(S;Y|Y_tilde) — but requires expensive masks
2. unlearning isomorphism (main paper): Mask refinement = forgetting dilation artefacts — no new annotations needed beyond a small retain set
3. certified pixel-level unlearning (main paper): Projects (epsilon,delta)-indistinguishability to pixel-wise conditional output space
4. global spurious mutual information (main paper): Certified unlearning provably upper-bounds this metric, formally guaranteeing shortcut reduction

The certified unlearning algorithms form a progression:

• Newton-step unlearning (Sekhari 2021): Exact Hessian-based, convex losses only
• randomized gradient smoothing (Zhang Z. 2022): Hessian-free via noise smoothing
• local convex approximation for certified unlearning (Zhang B. 2024): Extends to non-convex DNNs via l2 regularization
• gradient clipping for unlearning / model clipping for unlearning (Koloskova 2025): Handles non-convex losses with DP-style privacy amplification
• PNSGD for certified unlearning (Chien 2024): Unified SGD-based approach with sequential unlearning

Open Questions

• Model collapse: Certified operators can collapse large models (observed on melanoma detection). Can more robust certified algorithms avoid this?
• Assumption relaxation: The disjoint support assumption (spurious features don’t overlap with pathology) fails for some artefacts (e.g., hair overlapping with lesions)
• Multi-class scaling: Performance degrades in the 3-class (background/benign/malignant) setting with aggressive unlearning (90%)

---

topic