acceptodds
Withdrawn from ICLR 2027

Disentangled Counterfactual Diffusion for Identity-Preserving Explanations of Black-Box Medical Image Classifiers

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Complex medical image classifiers, particularly deep learning models, can achieve high predictive accuracy, yet often provide little evidence about which image factors truly drive their decisions, creating substantial safety risks. Saliency maps often highlight regions associated with a prediction but cannot test whether changing those factors would change the prediction; unconstrained generative counterfactuals may also alter patient-specific anatomy and thereby confound the explanation. We introduce Disentangled Counterfactual Diffusion, an identity-preserving framework that explains black-box medical image classifiers through controlled interventions designed to change their predictions. The model factorizes each image into a compact class-associated code and a multiscale individual representation. To construct a counterfactual, it retains the source individual representation, replaces the class-associated code with the target-class code, and performs global DDIM editing. Training combines diffusion reconstruction with class supervision, prototype alignment and separation, adversarial suppression of class information in the individual branch, and cross-covariance decorrelation. For counterfactuals generated through code exchange, additional objectives enforce the classifier prediction, class-code transfer consistency, preservation of individual features, anatomy, and edges, cycle reconstruction, and image realism. By comparing black-box classifier outputs and image differences between an original image and its identity-matched counterfactual, the method identifies class-associated factors that are sufficient to alter the model decision while controlling for subject-specific structure. We evaluate the method across multiple medical image classification tasks. The results show that, compared with existing approaches, our method more accurately identifies the key factors that influence model decisions and provides more precise explanations, offering an actionable route to understanding the decision process of black-box medical image classifiers.

Primary area:
interpretability and explainable AI
Keywords:
Medical image classification, Counterfactual explanations, Disentangled diffusion model, Identity-preserving generation, Black-box model interpretability

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_z3hdOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes DCD, a diffusion-based counterfactual explanation method for medical image classifiers. It separates an image into a class-associated code and a multiscale identity representation, and generates counterfactuals by interpolating the class code toward an opposite-class donor while preserving the source identity features. The method is evaluated on four medical imaging datasets using AOPC and peak degradation.

Critical strengths

The paper addresses an important problem in medical counterfactual explanations: changing classifier-relevant evidence while preserving patient-specific anatomy. DCD obtains the best reported AOPC on all four evaluated datasets and strong PD performance under the selected benchmark protocol.

Critical weaknesses
  1. The central identity-preservation claim is not convincingly validated. Table 4 reports LPIPS and an identity-feature distance only for DCD, without comparison to any baseline. Moreover, the identity distance is computed using the model's own learned identity representation, which is directly optimized to remain stable. Therefore, these results do not establish that DCD preserves identity better than competing counterfactual methods.

  2. Important ablations and closely related diffusion baselines are missing. DCD combines class supervision, prototype losses, adversarial class removal, cross-covariance regularization, exchange consistency, structural preservation, cycle reconstruction, and realism constraints.
    However, I did not find quantitative ablations showing which components are necessary. In addition, the related-work section discusses recent diffusion-based counterfactual methods such as DiffExplainer, DeCoDEx, diffusion-autoencoder approaches, and semantic-abduction methods, but the quantitative comparison does not include several of these closest diffusion-based competitors. The eleven reported baselines are inherited from the CAE/CAML benchmark rather than representing a comprehensive comparison to recent counterfactual diffusion methods.

Questions for authors

Can the authors compare LPIPS and identity preservation against the counterfactual-generation baselines, using an independent identity/anatomy representation rather than DCD's own identity encoder? Can the authors provide ablations for the disentanglement losses, identity conditioning, cycle loss, and realism loss? How sensitive are the results to donor selection and to the use of only six interpolation points? What fraction of test images actually obtain a successful decision-changing counterfactual along the sampled path? Why are several recent diffusion-based counterfactual methods discussed in Related Work not included in the quantitative comparison?

Reviewer_t8N2OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes DCD for explaining image classifiers used in medical domains. The work proposes to disentangle an image into a 8-dimensional class code and four level identity pyramid. A conditional diffusion model such as a U-Net is trained to reconstruct the image from both the class code and identity information. The supervision comes from the procedure of sampling a random donor image from another class, combine the identify of a source image with the class code of the donor class and ask a frozen classifier to predict the donor class. To create separation between the two, the paper uses many loss terms. On OCT, chest X-ray and two brain tumor dataset, authors report that DCD has best AOPC on all four datasets and best PD on three.

Critical strengths

The motivation in the medical domain is clear. The motivation that flipping a decision by any means (usually the anatomic features) is not good enough as a counterfactual. The paper demands both a valid decision change while preserving identity features. The loss terms although many individually make sense and have a motivation associated with them.

Overall AOPC when compared to CAML gains about 0.1 and 0.08 on OCT and Brain Tumor2. The limitation section is upfront with the authors acknowledging the factorization is not causally identifiable, the classifier cannot be a black box and the siz path points are a coarse estimate of classifier boundaries.

Critical weaknesses

The paper is extremely hard to read and seems to jump into terminologies without properly introducing them or giving enough context

The evaluation with baseline approaches is incomplete. Some of the comparisons seemed to have been reused from baseline paper and DiffExplainer is the only diffusion based method. The paper is missing comparison with DiME, Decodex and ACE and other cited approaches.

The idea of separating class vs identity information is not new and comes from CAE and CAML. The only change comes from replacing the decoder with a diffusion model and adding regularizers for the model to work. In my view the paper is well-engineered solution combining pieces known to work and I think the authors need to be more upfront about what is new.

There is a delicate balance between counterfactual vs adversarial example. While counterfactural need to be semantically meaningful, adversarial need not be. With the current setup, it is very easy for the generator to flip the frozen classifier by exploiting the classifier-specific feature representation and the current evaluation using deletion would fail to capture it. Can you check if DCD can flip an independent classifier trained with different seed or architecture? That would be true test of counterfactual-ness.

Despite introducing so many loss terms, the paper totally misses the mark with no ablation of these terms. The weights for these loss terms are not reported. Additionally, other choices like 8-dimensional code, number of levels in identity feature pyramid, DDIM initialization are not justified. At minimum, replacing the diffusion vs a CAE style decoder with the same loss terms will help.

The identity distance measurement is somewhat circular. It is trained to match the distance and measuring that very distance is not the right way. LPIPS would be a better measurement but it is reported only for DCD so we have no way to say that DCD preserves identity better than baselines. The paper says some of this is future work but it is cheap enough to do it now.

The donor is sampled randomly but no sensitivity analysis is done to account for this randomness

For being a counterfactual paper, the paper does not use metrics like flip rate, proximity. Size of changed area or FID. Minimality is preferred in Counterfactual generation yet there is no mention of it.

Brain Tumor2 presents an excellent opportunity to evaluate the localization ability of this setup due to presence of tumor segmentation masks using metrics like pointing game of IoU.

Finally, the paper has no statistical significance, std-dev or multi-seed runs reported. Which is very important for a paper that heavily relies on randomness both in the architecture and method.

Questions for authors

Can you add comparisons with other diffusion based counterfactual methods like DiME, DeCoDEx and ACE? Were the baselines rerun with your ResNet-50 classifiers or taken from CAML?

Can you provide ablations for the loss terms and report their weights?

Can you check if DCD counterfactuals can flip an independent classifier trained with a different seed or architecture?

Can you report LPIPS for the other counterfactual baselines too?

What is the flip rate of DCD, and how are images that have no point on the path handled in AOPC and PD?

Reviewer_6SbVOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The proposed Disentangled Counterfactual Diffusion (DCD) generates counterfactual explanations by separating class-associated codes from identity features. Then, it uses a conditional diffusion model to edit the class code while preserving source structure. It reports improved deletion-based faithfulness metrics on four medical imaging datasets.

Critical strengths

The framework tries to address the important problem of changing classification predictions while preserving source anatomy.

Critical weaknesses
  1. The contribution from DCD is insufficiently validated. No ablation study isolates the benefits of diffusion, multiscale conditioning, or the numerous regularizers.
  2. The evaluation does not establish reliable, anatomy-preserving counterfactual generation. Counterfactual success and failure rates handling are unspecified. The preservation metrics lack baseline comparisons. The identity distance replies on the representation explicitly constrained during the training.
Questions for authors
  1. Which components of DCD drive the improvements? Can you provide controlled ablations separating the effects of diffusion decoding, multiscale source conditioning, and the disentanglement objectives?

  2. What are the counterfactual success rates, and how are failures handled? Do the preservation gains hold against baselines using an independent anatomical assessment?

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

Related papers

Discussion (0)

Sign in to comment.