acceptodds
Under review as a conference paper at ICLR 2027

Towards Certified Shortcut Unlearning in Medical Imaging

Abstract

Many medical AI imaging applications rely on segmentation-classification pipelines. In practice, coarse or imprecise segmentation masks often admit *spurious* background cues and the model learns the associations between these cues and the pathological regions of interest (organs, lesions, etc.). Such *shortcuts* tend to severely affect the accuracy of the downstream task. Prior art relies on fine-grained pixel-level annotations to separate background cues from pathology regions. Unfortunately, high-quality annotations are expensive. We aim to ensure high performance on downstream tasks without relying on expensive annotations. We propose using machine unlearning to mitigate shortcut learning. In contrast to relying on heuristic methods, we propose a principled approach that, for the first time, provides guarantees for shortcut unlearning. We develop a definition of certified unlearning for segmentation tasks and formally connect it to its canonical definition. Further, we introduce an information-theoretic objective function, *global information reduction*, and prove that any -certified machine unlearning operator bounds it relative to the fine-mask reference. Importantly, the bound further limits the accuracy any decision rule can extract from the artefact. We discover existing certified operators collapse at the scale of real medical decoders, and introduce a new machine unlearning algorithm that challenges the SOTA in segmentation at tight budgets. On ISIC melanoma trap sets and SIIM-ACR pneumothorax with a naturally occurring chest-tube shortcut, we compare unlearning with state-of-the-art refinement methods on segmentation, shortcut robustness, fine-mask budget and compute. Unlearning spans the Pareto front on both datasets, no refinement method reaches it, and the certified operator is ahead of exact retraining, the only other method with a guarantee, at every fine-mask budget.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Image Segmentation, Machine Unlearning, Shortcut Learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_pZmGOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses shortcut learning caused by coarse segmentation annotations in medical imaging. The authors formulate segmentation refinement as a machine unlearning problem and propose a certified pixel-level unlearning framework with theoretical guarantees on spurious information reduction. They further develop a one-shot certified unlearning algorithm to correct pretrained segmentation models without full retraining. Experiments on ISIC melanoma and SIIM-ACR pneumothorax datasets show competitive shortcut robustness and favorable trade-offs among segmentation accuracy, annotation cost, and computational efficiency. The overall idea is insteresting.

Critical strengths
  1. The paper addresses shortcut learning caused by coarse segmentation annotations, a practically relevant challenge in medical imaging and formulate segmentation refinement as machine unlearning is interesting and provides a new direction for mitigating shortcut learning.
  2. The paper introduces an information-theoretic objective and derives theoretical guarantees connecting certified unlearning to shortcut reduction.
Critical weaknesses
  1. Definition 2.1 guarantees indistinguishability from a retrained model, while Definition 2.2 compares predictions directly with fine masks. The connection between these guarantees is insufficiently justified, the authors should clarify it.
  2. Removing incorrect foreground constraints is not necessarily equivalent to relabeling those pixels as background under standard segmentation losses.
Questions for authors
  1. How does the standard certified unlearning guarantee in Definition 2.1 imply Definition 2.2, given that the former compares against a retrained model while the latter compares against fine segmentation masks?
  2. How is removing incorrect foreground constraints equivalent to relabeling those pixels as background? Does this equivalence hold under standard segmentation losses?
  3. How do the class-conditional prediction bounds in Equations (5)–(6) guarantee a reduction in spurious mutual information, which depends on the joint distribution of predictions, pathology, and artifacts?
  4. Given the relatively low Dice scores and high variance in some configurations, can the authors provide stronger segmentation baselines, additional statistical analyses, and evaluations with larger models?
Reviewer_dKdiOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The main claim of the paper is to propose a principled approach with theoretical guarantees for shortcut unlearning in medical image segmentation and classification. The intuition for unlearning is to expose the model to a few examples of better annotated images, i.e. with finer segmentation masks that do not overlap with the main source of shortcut (e.g. a chest tube in an X-ray of a patient with pneumothorax): the shortcut is considered unlearned if the model, after unlearning, makes predictions that match closely those of a model trained directly from those finer grained segmentations without being exposed to the (removed) coarser masks. The paper aims to formalize these concepts for segmentation using information theory, and to propose a practical certified unlearning approach.

Critical strengths

The problem of shortcuts is a real issue for (grounded) automated diagnosis models in medical image analysis, that has been demonstrated across various tasks. So the motivation for the paper is sound.

Critical weaknesses
  1. The main weakness of the paper is insufficient clarity, that in my view cannot be addressed through rebuttal, as the changes to the paper that I would request would essentially amount to a new submission.

The paper does not establish a clear link between the theoretical part, that aims to establish a concept of unlearning for segmentation, and the proposed unlearning methods, that are only very briefly described (one paragraph and a half in section 4). Furthermore the abstract positions the method in contrast to existing "heuristic methods", but one of the two proposed methods is called "Heuristic unlearning (ours)".

It is not clear from the methodology

  • what are the assumptions supporting the theorems vs. what are the results proven; how the proposed method checks those assumptions and thus follows the result
  • different methodological subparts are not adequately linked together
  • how the theoretical analysis supports the practical unlearning approach is not clear

The experimental section & results also lacks clarity.

  1. In addition, I am not fully convinced of the soundness of the mathematical derivations. Some notations introduced early in the paper seem to be abuses of notation, such as U(Y) to give one example, where U is an unlearning operator and Y a segmentation mask (rather than a model); it is not clear how those abuses of notations affect the soundness of the mathematical derivations

  2. Hypotheses, such as disjoint spatial support between the shortcut and the region to be segmented, are also debatable (couldn't the chest tube touch the air in the pleural cavity in the case of pneumothorax?).

  3. The paper claims that the proposed methods span the Pareto front in the experiments, a claim that is challenged by the numbers reported in the tables.

  4. In terms of baselines, the paper does not make a substantial effort to compare to baselines trained directly from the finer annotations (possibly with SSL large scale pre-training to adjust to the few-shot constraint).

Questions for authors

Follow-up minor question assuming a new submission to another conference with revised material:

At an intuitive level, I would like to question the theoretical framing of shortcut unlearning by the paper - Definitions 2.1-2.2 state that the predictions of the "unlearned" model trained on a set of segmentations bar some unlearned coarse segmentations should be undistinguishable from those of a model directly trained without those few coarse segmentations.

But - if the model was originally trained on 1000 coarse masks containing the shortcut; and then unlearns 50 coarse masks in favor of 50 finer grained masks for 50 of those images (e.g., a few-shot, annotation costly scenario matching the motivation in the introduction); that still leaves 950 coarse masks with shortcuts that take part in the training of that model. Wouldn't such a model therefore still be biased? Hence I am not sure as to why we would want to certify that our predictions are close to those of such a model.

Reviewer_89tNOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper formulates correction of coarse segmentation masks as removal of pixel-level training constraints. It develops a spurious-information objective and bounds its change under certified unlearning, subject to a disjoint-support data model. A one-shot Gaussian perturbation followed by rescaling and recovery is proposed as a certified operator. Experiments compare unlearning, fine-tuning, and refinement methods on synthetic images, ISIC melanoma trap sets, and SIIM-ACR pneumothorax data.

Critical strengths

Connecting annotation refinement with removal of unwanted training influence is an interesting direction. The paper also evaluates a naturally occurring chest-tube shortcut as well as dermoscopic artefacts, and reports segmentation, robustness, annotation budget, and computation rather than relying on one metric.

The discussion of disjoint support and of the insufficiency of the marginal pixel-level definition (Appendix B.3) is useful. These distinctions could support a more carefully scoped theoretical contribution.

Critical weaknesses
  1. The certificate proved for Algorithm 1 does not use the reference in Definition 2.1. Appendix C explicitly compares the output with retain-only retraining followed by the same noise, rescaling, and recovery, rather than with unperturbed retain-only retraining. Closeness to this modified reference does not by itself establish the stated guarantee relative to the fine-mask model or transfer that model's low shortcut dependence. The certificate also requires a fixed, forget-data-independent discrepancy bound Delta that holds almost surely for every covered dataset and deletion request. The experimental configuration does not provide its numerical value or establish such a bound. Shared seeds alone do not establish it. Thus the reported epsilon and delta values are not a demonstrated certificate for the advertised reference.

  2. The bridge from the geometric theory to learned segmenters is unresolved. Equation (1) defines Y^(r) as a deterministic dilation of the true mask, whereas Lemma D.3 uses Y^(r2) for a trained model's predictions. Properties of a dilated annotation, including containment of the true foreground, need not hold for those predictions. In addition, the proof of Proposition 3.2 selects an r0 after which both alpha(r) and rho(r) decrease, but justifies its existence only from alpha(r) tending to zero. That does not establish eventual monotonicity of rho(r). The synthetic experiment further reports I(Z;S|Y) in Table 3 rather than the foreground-conditioned quantity I(S;Y|Z=k) used in the theorem. These gaps prevent the current theory and experiments from establishing the central shortcut-reduction claim.

Questions for authors
  1. What numerical Delta was used in each experiment, and what argument makes it a valid almost-sure, forget-data-independent bound? Can the claimed certificate be established against the unperturbed retain-only training distribution in Definition 2.1?
  2. Which results concern dilated ground-truth masks and which concern learned predictions? Please supply the missing assumptions or derivation linking them, including the monotonicity claim for rho(r).
  3. Can the synthetic experiment evaluate the theorem's actual foreground-conditioned information quantity and its stated reference, instead of substituting I(Z;S|Y)?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.