Abstract

Machine learning-based computer-aided diagnostic systems often exhibit compromised fairness due to biases inherent in training datasets, leading to suboptimal and biased performance across diverse populations. These issues are particularly evident in pneumonia diagnosis and lesion segmentation, where disease prevalence and imaging characteristics vary significantly across demographic groups and pneumonia subtypes. To address these challenges, we propose FairPneum, a generalizable framework designed to enhance fairness in learning from biased datasets. FairPneum first introduces a hierarchical embedding disentanglement method that progressively isolates sensitive attributes using adversarial training, while an anti-bias training strategy is employed to leverage counterfactual embeddings to improve robustness against sensitive biases. We also introduce two novel metrics to effectively evaluate the global fairness in continuous performance metrics across different attributes while addressing the dependency between different sensitive attributes. Implementing FairPneum with both convolutional- and transformer-encoders, we demonstrate significant improvements in both performance and fairness on the public COVID-19 chest X-ray dataset and an in-house pneumonia CT dataset. Compared to SOTA fairness learning approaches, FairPneum effectively reduces performance disparities across demographic groups while maintaining or enhancing overall model accuracy. This study highlights a promising approach for developing unbiased and clinically generalizable solutions for pneumonia diagnosis and segmentation. The source code and dataset are available at https://github.com/JXXrrrrR/FairPneum.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0602_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/JXXrrrrR/FairPneum

Link to the Dataset(s)

https://github.com/JXXrrrrR/FairPneum https://www.kaggle.com/c/siim-covid19-detection/data

BibTex

@InProceedings{ChuYue_FairPneum_MICCAI2026,
        author = { Chu, Yuetan AND Jin, Xinran AND Ma, Xinghua AND Luo, Gongning AND Gao, Xin},
        title = { { FairPneum: Improving Fairness in Pneumonia Diagnosis and Lesion Segmentation } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16885},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    1) The authors proposes hierarchical disentanglement. 2) Counterfactual re-entanglement strategy 3) Continuous metrics to evaluate fairness with segmentation.

  • 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) Fairness in medical AI is a clinically significant issue. Biased models can lead to disparate care across demographic groups. 2) The core insight is sound. The hierarchical disentanglement approach is intuitive and easy to understand. 3) Fairness in segmentation, and its metrics are both good addition to the study.

  • 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) The paper lacks many ablation studies that further strengthen the proposed solution. The authors claim that hierarchical approach is better based on intuition and supports this with a visualization, but does not provide a controlled ablation isolating the hierarchical vs. single-layer design choice.

    2) The experiments are limited. The segmentation experiments where FairPneum shows its strongest fairness gains cannot be reproduced with the private in-house dataset.

    3) The authors list “Establishment of a multi-subtype pneumonia dataset” as a contribution, which implies it has value to the community. But a contribution that can’t be accessed has limited practical impact.

  • Please rate the clarity and organization of this paper

    Good

  • 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • 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?

    My decision is based on the major weaknesses: limited experiments and limited ablations.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [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.

    Satisfied with the rebuttal



Review #2

  • Please describe the contribution of the paper

    The paper proposes FairPneum, a fairness-aware framework for pneumonia diagnosis and lesion segmentation. The method combines hierarchical embedding disentanglement with an anti-bias re-entanglement strategy, with the goal of suppressing sensitive-attribute information across encoder stages rather than only at a single latent layer. The paper also introduces LPE and GPE to assess fairness for continuous outcomes such as Dice score, and evaluates the framework on chest X-ray classification and chest CT lesion segmentation.

  • 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 addresses an important and clinically relevant problem: subgroup performance disparity in medical imaging models. This is timely and well-aligned with MICCAI’s interest in robust, deployable medical AI.
    • The strength of the work is the hierarchical treatment of bias, which is well motivated in this application because the manuscript argues that age, region, and subtype can remain entangled with radiological appearance across feature scales.
  • 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 main weakness is that the methodological novelty appears moderate. The core ingredients overlap with related fairness-representation papers, such as Fairness by Learning Orthogonal Disentangled Representations and Fairness via Representation Neutralization.
    • A second weakness is that the baseline comparison is not fully clean. The manuscript states that the embedding debiasing block and reconstruction block were inserted after each encoding layer for all compared fairness methods, while only FairPneum used the proposed optimization strategy. This means the experiments more directly support superiority within the authors’ harmonized architecture than superiority over faithful original implementations of the cited baselines.
    • Third, the segmentation fairness evidence is limited by the test-set scale. The CT study contains 57 independent test studies, while fairness is analyzed across four sensitive attributes. This is a fairly small basis for subgroup fairness claims, especially for multi-level attributes such as subtype, and it limits confidence in the stability of the reported subgroup differences.
    • Fourth, the fairness-metric contribution should be positioned more carefully. Prior medical segmentation fairness work has already evaluated subgroup disparity using Dice-based dispersion measures such as standard deviation (SD) and skewed error ratio (SER), and FairSeg introduced equity-scaled segmentation metrics. Against that background, LPE/GPE are better interpreted as an alternative continuous-outcome formulation rather than the first meaningful fairness metrics for segmentation.
    • Fifth, while LPE is reasonably intuitive, GPE appears somewhat heuristic. The paper defines lower LPE/GPE as fairer, and that interpretation is internally consistent in the manuscript, but the correlation- and significance-weighted aggregation in GPE is not yet fully justified relative to simpler subgroup-disparity metrics already used in the literature.
    • Finally, the method depends on sensitive attribute labels for adversarial supervision. This is a practical limitation for real-world deployment, since such labels may be incomplete or unavailable in many clinical datasets.
  • 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 has provided an anonymized link to the source code, dataset, or any other dependencies.

  • 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

    This is a relevant paper on an important topic, and I appreciate the attempt to address fairness in both diagnosis and lesion segmentation. My main suggestion is to position the novelty more cautiously. The method appears to combine several existing fairness-learning ideas in a sensible and clinically motivated way, but the manuscript should make clearer whether the main contribution is the hierarchical integration, the re-entanglement strategy, the continuous fairness metrics, or the combination of these elements. I also encourage the authors to clarify the comparison protocol more explicitly, since modifying all baselines with the proposed architectural blocks changes the interpretation of the benchmark.

  • 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?

    I lean positive because the paper addresses an important problem, is technically coherent, and presents a reasonably broad empirical study across tasks, modalities, and backbone families. The paper is weaker on core novelty than on practical relevance and completeness. In my view, the contribution is best understood as a meaningful integration of existing fairness-learning ideas for hierarchical medical imaging, with promising but still somewhat limited fairness evidence in the CT setting.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [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 rebuttal addressed several of my main concerns. The authors clarified the baseline comparison protocol and the intended interpretation of GPE, which make the experimental design and fairness-metric contribution easier to interpret.

    I still think the novelty is moderate, and some limitations remain, especially the relatively small CT test cohort for subgroup fairness evaluation. However, the paper addresses an important problem, is technically coherent, and presents a reasonably broad evaluation across tasks, modalities, and backbones. Overall, I now support acceptance.



Review #3

  • Please describe the contribution of the paper

    This paper proposes a plug-and-play fairness learning framework called FairPneum, for lesion segmentation and pneumonia classification. Its core contributions include (1) An adversarial hierarchical disentanglement module that is applied following each encoder layer which separates sensitive attributes from task-relevant features, (2) A re-entanglement strategy to prevent bias from moving onto the next encoder layer, employing first a zero substitution and then counterfactual embedding swapping sequentially to regularise the decoder towards attribute invariant predictions, and (3) Introducing fairness metrics for continuous variables as outcomes, like Dice Similarity Coefficient, namely Local Performance Equality (LPE) using first-order Wasserstein distance and Global Performance Equality (GPE) using correlation-weighted LPE aggregation. Experiments on a public chest X-ray dataset (MIDRC) and an in-house pneumonia CT dataset demonstrate improved fairness across age, sex, geographic region and pneumonia subtype groups under two backbone encoder architectures.

  • 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 with a concrete clinical motivation, provides coherent explanations of the proposed methods and results. 2.Hierarchical disentanglement is well-motivated by Figure 1b showing that bias persists across encoding layers. 3.Counterfactual embedding swapping is a novel and intuitive training regulariser and the ablation (FairPneum⁻) validates its contribution. 4.LPE and GPE fill a genuine gap regarding establishing a fairness metric that handles continuous scores in segmentation. 5.Two backbone architectures (SwinT and MedNeXt) tested, supporting generalizability claims.

  • 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.Since the in-house test dataset is small and as significance is computed at the slice level rather than the patient level, slices from the same patient are not independent, which may inflate statistical confidence in the subgroup comparisons. 2.The GPE metric weighs each attribute by its Pearson correlation with the outcome. If all attribute’s pearson correlations are near zero, the weighting term |ρ|(1−p) collapses towards zero for all attributes, and GPE becomes numerically dominated by the stability constant ε. In this edge case the metric loses its intended interpretation. The authors do not discuss this failure mode. 3.Counterfactual pairing criterion is ambiguous. For multi-label sensitive attributes, it is unclear what the counterfactual partner “from the same task class but with different sensitive attributes” for a selected sample would be. It is not specified whether “different” requires all attributes to differ, atleast one to differ or only the primary attribute being targeted. 4.It is stated that Mycoplasma subtype is predominantly found in children and COVID-19 subtype in older adults. If there are very few elderly patients with Mycoplasma subtype, the counterfactual Eᵇ pool for elderly is small, then how do you pair sample x in a “one-to-one manner” to another sample x’ to get counterfactual pair? 5.Adding extra modules to the baseline along with adversarial training takes more time to train and more computing resources to run. The paper does not report how much extra time or memory FairPneum requires compared to the baseline. 6.For the segmentation experiments (Table 2), the paper marks results with asterisks (*) to show which improvements are statistically significant. But for the classification experiments (Table 1), no such significance testing is reported. This makes it difficult to know whether the accuracy improvements in Table 1 are real or just by chance.

  • Please rate the clarity and organization of this paper

    Good

  • 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • 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

    Kindly check for spelling and grammatical errors like Figure 2(b) embedding disentanglemen(?). Also, this paper does not cite Kusner et al. (2017) on counterfactual fairness despite using the term “counterfactual,” which may reflect an incomplete engagement with the fairness theory literature rather than any integrity concern.

  • 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?

    FairPneum addresses a clinically important and underexplored problem: demographic fairness in pneumonia segmentation, with a technically sound and well-ablated method. The hierarchical disentanglement motivation is convincing, LPE and GPE are genuinely novel contributions to the fairness metrics literature, and dual-backbone evaluation supports generalizability. Despite the weaknesses, the work is clear enough to be understood and the methods are detailed enough so that other researchers can reproduce them.

  • 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.

    The rebuttal partially addressed Q3/Q4, but key methodological concerns remain unresolved, including slice-level significance testing, unconvincing GPE behaviour in near-zero correlations, and asymmetric statistical reporting in Tables 1 and 2.



Author Feedback

We sincerely thank reviewers and AC for their constructive feedback and positive comments on the clinical significance (R1/R2/R3), intuitive motivation (R1/R2/R3), fair segmentation contribution (R1/R3), and backbone generalizability (R3). We are encouraged that R2 and R3 gave accept recommendations, while R1 considered the paper marginally below the threshold mainly due to ablations and reproducibility. We address the major concerns below.

R1: Hierarchical design Our compared DFA is a representative single-level disentanglement baseline, as it learns one bias/intrinsic representation split in a single latent space. Therefore, our comparison with DFA provides a method-level comparison between single-level and hierarchical disentanglement. The existing ablation further isolates the contribution of counterfactual re-entanglement. Fig.1(b) further motivates why sensitive information should be suppressed at multiple levels. We will clarify this distinction in the revision.

R1: Dataset and reproducibility We have obtained hospital approval to release the in-house CT dataset, including images and lesion annotations, together with source code upon acceptance. Thus, the segmentation experiments will be reproducible. FairPneum also does not rely solely on this dataset; it is also evaluated on the public MIDRC dataset, with consistent fairness improvements across modalities and backbones.

R2&AC: Method novelty Our novelty is not adversarial debiasing alone, but its stage-wise encoder-level debiasing to medical imaging, where sensitive information can be distributed across feature levels and segmentation predictions depend on multi-level encoder features. Compared with RNF, which debiases the classification head, FairPneum directly debiases multi-level encoder features and is applicable to encoder-decoder segmentation. Compared with Orthogonal Disentangled Representations, which learns a global sensitive representation split, FairPneum performs stage-wise disentanglement and combines it with anti-bias re-entanglement.

R2&AC: Baseline comparison The compared fairness methods are not defined by fixed architectures; their core contributions are fairness objectives or representation-learning strategies, e.g., Lin et al.’s marginal ranking loss with subgroup update and PnD’s expert mixture with adaptive gating. In original papers, they also used general backbones such as DenseNet, rather than unique architectures. Therefore, using the same backbone is necessary for fair comparison, as it removes confounding effects from architecture and parameter count. Importantly, the proposed hierarchical disentanglement and re-entanglement are not used by other methods. The shared components only provide a common backbone, while each method keeps its own fairness mechanism. Hence, the gains mainly reflect the proposed strategy rather than an architecture. We will revise the manuscript to make this rationale explicit.

R2&R3&AC: Evaluation metrics Existing metrics such as SD and SER summarize group-level dispersion or error ratios. LPE instead compares the full performance distribution, while GPE further considers dependencies among sensitive attributes. GPE is intended for correlated attributes. When all correlations are uniformly low, GPE is outside its intended use case; in that case LPE is sufficient since attribute dependencies are negligible. We will clarify this in the revised version.

R3&AC: Counterfactual pair “Different sensitive attributes” means that at least one sensitive attribute differs. Each sample is paired with at most one mini-batch partner from the same task class. With a batch size of 16, eligible partners are usually available, but our method does not require every subgroup combination to appear in each batch. If no eligible partner exists, the sample is not swapped and only the debiased transformation is used. Thus, small subgroup pools only reduce swap frequency in that iteration rather than invalidating the method.




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.

    The paper proposes a fairness-aware framework for pneumonia diagnosis and lesion segmentation combining hierarchical embedding disentanglement with counterfactual re-entanglement, and introduces LPE and GPE as continuous-outcome fairness metrics. Several substantive concerns are raised across the reviews. The methodological novelty is moderate given overlap with prior fairness-representation work. More fundamentally, because the proposed architectural blocks are inserted into all competing methods while only FairPneum uses the proposed optimization, the reported gains may reflect superiority within the authors’ harmonized architecture rather than over faithful baselines. The GPE metric also requires justification, particularly its behavior in low-correlation regimes and its positioning relative to existing subgroup-disparity measures, and the counterfactual pairing criterion is underspecified under multi-label attributes and small subgroup pools. The rebuttal should address the major reviewers’ concerns including these points.

  • 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.

    This paper proposes FairPneum, a fairness-aware framework for pneumonia diagnosis and lesion segmentation that combines stage-wise embedding disentanglement with a counterfactual re-entanglement strategy, together with LPE and GPE as fairness metrics for continuous outcomes. The reviewers agree that the problem is clinically important, the hierarchical motivation is well grounded, and the evaluation across two modalities and two backbone families is reasonably broad. The rebuttal clarified the baseline comparison protocol, the positioning relative to prior fairness-representation work, the intended scope of GPE, and the counterfactual pairing criterion under multi-label attributes. The methodological novelty remains moderate, as the contribution is best read as a stage-wise integration of existing fairness-learning ideas adapted to encoder-decoder medical imaging, and the authors are encouraged to position it in those terms and to add a brief discussion of the intended-use regime of GPE and of the patient-level versus slice-level statistical analysis in the camera-ready version.



Meta-review #2

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

    Reject

  • Please justify your recommendation.

    The rebuttal doesn’t fully address reviewers’ concerns.



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.

    The authors have responded to most of the reviewers’ comments in a generally convincing and satisfactory manner. In particular, the revised manuscript addresses several concerns raised in the initial round of reviews and improves clarity in multiple sections.

    However, some concerns remain regarding the methodological novelty of the proposed approach. Moreover, the reported performance gains appear limited, and in several cases the method does not consistently outperform prior work. This is further compounded by the relatively limited scope of the experimental comparison, which includes only a small number of baselines (three methods for classification and two for segmentation), which is insufficient to fully contextualize the proposed approach within the broader state of the art.

    That said, despite these limitations, the overall response from the reviewers has been generally positive. In particular, there is consensus regarding the clinical relevance of the work, the introduction of a new dataset, and the proposed fairness-oriented metrics for segmentation.

    Given these considerations, I support acceptance of the manuscript.



back to top