Abstract

Deep learning models for medical diagnosis frequently exhibit substantial performance disparities across sensitive subgroups (e.g., race, sex), even when average accuracy is high. While generative data augmentation offers a route to mitigate this, existing strategies are suboptimal; they typically address only one or two dependency channels between sensitive attributes and image features. We formalize the medical image formation process via a structural causal model, revealing that sensitive attributes actually influence image content through four distinct pathways—a structural complexity neglected by prior works. Based on this insight, we introduce CIPHER (Causal Intervention Pathways for Healthcare Equity and Robustness), a framework designed to systematically intervene on all four causal paths. To achieve this, CIPHER utilizes a diffusion backbone equipped with classifier-free guidance and null-text inversion. This technical design enables the faithful reconstruction of patient-specific anatomy while allowing for the precise, editable synthesis of counterfactuals required to break sensitive dependency chains. We tested CIPHER using chest X-ray and dermoscopy benchmarks across both standard and shifted data distributions. By employing a multi-pathway intervention strategy, our model reduced worst-group disparities by an average of 35.8% compared to disease-conditioned synthesis baselines, while also improving total diagnostic accuracy.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2601_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/fdu-farm/CIPHER

Link to the Dataset(s)

CheXpert dataset: https://aimi.stanford.edu/datasets/chexpert-chest-x-rays MIMIC-CXR dataset: https://physionet.org/content/mimic-cxr/2.1.0/ MILK10K dataset: https://api.isic-archive.com/doi/milk10k/

BibTex

@InProceedings{JiaXin_CIPHER_MICCAI2026,
        author = { Jia, Xinyu AND Guo, Weidong AND Zhao, Wangyuan AND Guo, Yi AND Li, Zeju AND Wang, Yuanyuan},
        title = { { CIPHER: Causal Intervention Pathways for Healthcare Equity and Robustness } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16882},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The paper addresses performance gaps across sensitive attributes despite high average accuracy. Its key contribution is to model medical image formation in a causal framework, treating disease and sensitive attributes as intervenable variables while keeping patient-specific structure fixed. Based on this, the method uses diffusion inversion for counterfactual editing to selectively modify these attributes, generating controlled data that improves fairness and robustness without changing the model architecture.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
    • The paper formulates medical image generation in a causal framework, explicitly separating disease and sensitive attributes as intervenable variables while keeping patient-specific structure fixed. This allows controlled counterfactual data generation without modifying the model or training objective.
    • The four-path design decomposes data generation into generation (from noise) and instance-level editing (from real images), with separate control over disease and sensitive attributes.
  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
    • The baselines are insufficient. Standard augmentation baselines (e. g. , MixUp, CutMix, and recent augmentation methods in the CXR domain [1]) are missing, making it hard to judge the advantage of the proposed method. Also, P1–P4 function more as ablations than true baselines. [1] Prakash, Eva, et al. “Evaluating and Improving the Effectiveness of Synthetic Chest X-Rays for Medical Image Analysis. “ Proceedings of the IEEE/CVF International Conference on Computer Vision. 2025.
    • Since the method relies on conditioning with metadata tokens and synthetic data generation, it is unclear whether the improvements come from the proposed framework or simply from stronger data augmentation.
    • The explanation of the four paths is somewhat repetitive, while the experimental analysis is relatively limited in comparison. (Also, why is Section 2 titled “Analysis”?)
    • Additional qualitative analysis (e. g. , before and after Path 3 and 4 editing) would help better understand the behavior of the method, especially since changes in sensitive attributes (e. g. , age) in CXR images can be subtle and not clearly observable.
  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not provide sufficient information for reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The paper presents an interesting causal and data-centric approach to improving fairness. However, the lack of proper baselines (especially fairness methods) makes it hard to judge its advantage, and the main results are limited and mixed with ablations, so the effectiveness is not clearly shown. More space should be given to analyzing how the intervention-based images actually contribute to fairness in practice.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Reject

  • [Post rebuttal] Please justify your final decision from above.

    I thank the authors for their thoughtful rebuttal. The causal framework and SCM perspective on bias are interesting. After careful consideration, I keep my original score, mainly because the central question I raised is not yet fully resolved.

    My main concern was whether the observed gains come from the four-pathway causal structure or from sensitive attribute conditioned synthetic augmentation more broadly. The rebuttal argues that matching sample count and generator isolates the causal contribution. This is reasonable but does not fully match the augmentation distribution across methods. Different pathways draw from substantively different conditional distributions. P2, for example, implicitly oversamples minority subgroups through sensitive conditioning. A direct comparison against a non causal but distribution matched synthetic baseline, such as naive subgroup-balanced generation without an explicit SCM, would more cleanly isolate the contribution of the causal structure.

    On baselines, the additional comparisons on the same datasets and metrics are appreciated and helpful. On qualitative validation, the promised supplementary visualizations of Path 3 and Path 4 edits were not included in the rebuttal. Given that some sensitive attribute edits can be subtle, these would meaningfully strengthen the paper.

    The SCM framing is a useful perspective and the breadth of evaluation is appreciated. The points above are why I keep my original score. I would not object to acceptance if the AC and other reviewers find the rebuttal sufficient.



Review #2

  • Please describe the contribution of the paper

    This paper formalizes the medical image generation process as a structural causal model, which illustrates how sensitive attributes influence image generation through four distinct pathways. Experimental results demonstrate that intervening on these four pathways can improve both fairness and diagnostic accuracy in downstream tasks.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    1.The paper is well written and easy to follow. 2.The proposed unified causal model, which captures how sensitive attributes influence the image generation process, is valuable for many real-world applications. The four pathways through which bias can arise, as well as the corresponding intervention strategies to mitigate it, are clearly explained. 3.Experiments across two imaging modalities, CXR and dermoscopy, and under both in-distribution and out-of-distribution settings demonstrate the effectiveness and generalizability of the proposed method.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    1.It is unclear how the method handles multiple sensitive attributes: are they mitigated jointly or one by one? This matters because one-at-a-time debiasing can be problematic in the presence of causal dependence among attributes. In particular, if D is a collider between sensitive attributes (for example S1 -> D <- S2) then conditioning on D while mitigating bias for S1 can induce a spurious association with S2, potentially creating additional or residual bias. The same can happen with C as a collider between sensitive attributes. The paper would benefit from clarifying this point. 2.From the results, both Path 1 and Path 2 show cases in which fairness underperforms the conventional baseline, RandAugment. Is there an explanation for this behavior? 3.In the description of Path 2, the authors state that they “systematically vary S to reshape subgroup composition.” It is unclear what “systematically varying” means in practice. Does this simply refer to balancing S? 4.The results are presented with error bars, but it is unclear how these were generated. Were they obtained through cross-validation or repeated runs?

  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not provide sufficient information for reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    Formalizing bias in medical image generation as a unified structural causal model considering all potential pathways is valuable and effective for understanding how sensitive attributes influence the image generation process and, in turn, for guiding bias mitigation.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #3

  • Please describe the contribution of the paper

    The paper introduces an alternative view to mitigating bias in medical image classifiers using synthetic data. By carefully considering the relationships among demographic (sensitive) attributes, disease, and medical images, the authors present different types of bias as distinct pathways in a structural causal model (SCM). Eliminating various biases (e.g., population bias) involves interventions on specific pathways, ultimately amounting to different control/editing mechanisms for an image generation model (e.g., different input conditions). This proposed framework is named CIPHER. CIPHER specifically employs a latent diffusion model for controllable image generation via textual guidance. To demonstrate the benefit of CIPHER in mitigating bias, the authors study its impact on fair image classification across 3 datasets corresponding to 2 applications: CheXpert Plus and MIMIC-CXR for chest X-ray disease classification, and MILK10K for dermoscopy disease classification. Experimental results reveal that intervening on each of the four pathways presented results in significant differences in downstream overall performance and fairness.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
    • Novel formulation: the causal framework presented in this work offers a paradigm for mitigating bias in medical image classifiers using synthetic data. Traditionally, previous works, such as Moroianu, S.L. et al. (2025), train image generation models with controls for both disease and demographic (sensitive) attributes, and generate balanced synthetic datasets (i.e., equal proportions of sensitive attributes) that are then used to train downstream image classifiers. However, by more carefully considering the data generation process and its relationship to different types of bias, the proposed framework, CIPHER, offers a novel approach to generating synthetic images to combat bias.
    • Evaluation on multiple domains: the experiments in the paper apply the proposed framework to 3 datasets corresponding to 2 different applications: CheXpert Plus and MIMIC-CXR for chest xrays disease classification, and MILK10K for dermoscopy disease classification. Validating CIPHER across different domains helps elucidate the generalizability of the proposed framework. References: [1] Moroianu, Stefania L., et al. “Improving Performance, Robustness, and Fairness of Radiographic AI Models with Finely-Controllable Synthetic Data.” arXiv preprint arXiv:2508.16783 (2025).
  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    Unclear methodology and experimental protocol: In general, Sec 3.Methodology and Sec 4.2.Ablation Experiments are difficult to follow and have many missing details. This makes it difficult to appreciate the benefits of CIPHER relative to related work, such as that of Moroianu, S.L. et al. (2025). The authors should make it clearer how the different pathways relate to the strategies employed in prior work, or whether each pathway is a unique contribution. I summarize some key points that are unclear or missing below:

    • Data generation for different pathways: it was really difficult to follow what intervening on a pathway means practically for downstream image classifier training. Can authors please clarify how input conditions were sampled to generate the synthetic dataset for each pathway i.e., how are the conditions c_s and c_d sampled for Path 1 and Path 2? Are sensitive attributes combined together or studied separately? Does intervening simply imply creating a balanced synthetic dataset?
    • CIPHER: What exactly is the “CIPHER” method in relation to the four proposed pathways? Is the final synthetic dataset an equal combination of images from each of the four pathways?
    • Image classifier training details: How are the image classifiers (DenseNet-121) trained, i.e., what were the optimizers, losses, batch sizes, learning rates, etc.? What do different runs entail to generate the error bars? References: [1] Moroianu, Stefania L., et al. “Improving Performance, Robustness, and Fairness of Radiographic AI Models with Finely-Controllable Synthetic Data.” arXiv preprint arXiv:2508.16783 (2025).
  • Please rate the clarity and organization of this paper

    Poor

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not provide sufficient information for reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    Does Path 2: Sensitive-focused generation constitute balanced sampling across sensitive attributes, making it comparable to how previous work (e.g., Moroianu, S.L. et al., 2025) has traditionally mitigated bias using synthetic data?

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    In general, I find the paper’s methodology and experiments very unclear (see major weaknesses section). Therefore, it is unclear what the paper’s technical contributions.

  • Reviewer confidence

    Somewhat confident (2)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    The authors have sufficiently addressed my concerns. If accepted, I hope they can incorporate their rebuttal into the final version to help improve the manuscript clarity.



Author Feedback

We thank the reviewers and AC for their thoughtful feedback. Reviewers acknowledge that our causal analysis for fairness is intriguing, valuable, and novel (R1–3), clearly demonstrated (R2), that CIPHER is versatile and effective (R1–2), and that evaluations across three datasets are thorough (R2–3). As highlighted by R3, our paper offers a fresh and carefully crafted perspective on mitigating bias in medical classifiers, which we believe will be of broad interest to the MICCAI community. Below, we address reviewer concerns regarding baselines, technical contributions, and experimental protocols. 1.R1 on baselines. Prior work has shown that synthesis-based methods consistently outperform simple baselines (“Generative models improve fairness of medical classifiers under distribution shifts”); accordingly, we focus our comparisons on synthesis-based methods. For completeness, we additionally report results for MixUp, CutMix, JTT (Just Train Twice), and SELF (Selected Last-layer Finetuning), confirming our findings (order: CIPHER / MixUp / CutMix / JTT / SELF). CXR-ID: age gap 0.068/0.077/0.079/0.083/0.085; race gap 0.027/0.031/0.033/0.034/0.037.CXR-OOD: age gap 0.068/0.085/0.081/0.089/0.094; race gap 0.036/0.058/0.062/0.049/0.055.Dermoscopy: age gap 0.104/0.162/0.109/0.116/0.106; gender gap 0.020/0.032/0.028/0.041/0.032; skin-tone gap 0.046/0.118/0.095/0.125/0.115.We further emphasize that CIPHER differs fundamentally from existing synthesis-based methods such as Prakash et al., which rely on label-conditioned generation, equivalent to the P1 variant in our ablations. We compare thoroughly against P1 and demonstrate superior performance, supporting our claims regarding the importance of causal structure. 2.R3 on method contributions. In causal theory, an intervention severs the dependency from causes to the intervened variable. Our analysis identifies four pathways through which the sensitive attribute S influences the observed outcome Y, and we show that each can be blocked by learning the appropriate joint distribution and causal structure. For example, our first generative model intervenes on P1 (Fig. 2) by synthesizing disease (D)-conditioned images; training a classifier on these images causes it to disregard within-disease variation, so predictions on Y are no longer affected by manifestation bias. This effect is difficult to achieve by hand-crafted means but is learned implicitly from data (Table 1: both smaller and larger chest cavities would be correctly predicted as “Pneumothorax”). The remaining pathways can be understood analogously by referring to Sec. 2.Because the causal graph in Fig. 1 has not been previously discussed in this context, existing methods heuristically exploit synthesis-based augmentation via a single mechanism and lack a complete causal view (e.g. Moroianu et al. conflate P1 and P2). CIPHER addresses all pathways, and our ablations consistently confirm its advantages. 3.R1 on the effectiveness of causality over strong data augmentation. The benefit of data augmentation depends jointly on the quantity and quality of synthesized samples. We carefully control for both (see Sec. 4.1) by augmenting with an equal number of synthesized samples generated by the same pretrained model across all methods. The observed reduction in subgroup gaps by CIPHER under this fixed data budget directly demonstrates the effectiveness of the causal structure, not merely the augmentation volume. 4.R2 on handling multiple sensitive attributes. The current framework mitigates one sensitive attribute at a time. Extending the SCM to support joint interventions over multiple attributes is a natural and straightforward extension, which we plan to include. 5.R1 on qualitative edits. We will provide additional visualizations via supplementary online content. 6.R2–3 on implementation details. Error bars denote standard deviations over five independent runs. Code, data, and pretrained models will be made publicly available.




Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    This manuscript has received mixed reviews, with major concerns centered on insufficient baselines (e.g., missing standard augmentation methods like MixUp/CutMix and fairness baselines), unclear experimental methodology, and difficulty isolating whether improvements come from the causal framework or simply stronger data augmentation. During the rebuttal period, the authors are required to address these limitations.

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    The paper presents a novel causal framework for fairness in medical imaging, formalizing bias pathways through a structural causal model with four distinct intervention strategies. Reviewers acknowledge the strong formulation, clear exposition, and validation across multiple modalities (CXR and dermoscopy). While Reviewer #1 raises baseline concerns, Reviewer #2 and #3 find the work valuable and methodologically sound. The rebuttal sufficiently addresses reproducibility and clarifies the causal contribution. Recommendation: Accept.



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    All the reviewers agree that the proposed method has technical merits. The main concern of R1 is isolating the effects of data augmentation, which could be potentially addressed in the revised manuscript. Though additional results have been included in the rebuttal, these results have been ignored in this evaluation and the decision is primarily based on the technical contributions of this work.



Meta-review #3

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    I agree with R1 that there are a few technical issues that worth further discussion and validation, but it seems there is a consensus that the proposed method is generally innovative and thought provoking. I believe this paper might stimulate a meaningful and vibrant discussion at MICCAI.



back to top