Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Medical foundation models are increasingly used as generic visual backbones for downstream diagnosis, yet existing fairness studies in medical imaging mostly focus on a single protected attribute (e.g., age or sex) and rarely examine intersectional subgroups defined by multiple demographics. Compared to single-attribute settings, multi-attribute fairness introduces additional challenges: (i) intersectional subgroup data sparsity, where some demographic combinations contain very few samples, making subgroup-specific adaptation unstable and prone to overfitting; and (ii) sensitive attribute leakage (proxy discrimination), where learned representations retain correlated demographic information and allow models to rely on demographic shortcuts rather than clinically relevant visual features. To address this, we propose Fair Gated-Expert adaptation with multi-attribute gating (FairGE), a fairness-aware adaptation framework for medical foundation models. FairGE integrates two complementary components: a Multi-attribute Gated distribution-aware Mixture-of-Experts (MA-dMoE) module that employs shared experts with multi-attribute gating (e.g., age and sex) to enable intersectional adaptation while avoiding overfitting on small subgroups, and a Multi-attribute Purification (MAP) module that combines a lightweight MLP with multiple adversarial attribute classifiers (one for each sensitive attribute) trained via gradient reversal to prevent the model from encoding sensitive demographic information in the final representation. Experiments on two representative medical benchmarks, HAM10000 and Harvard-FairVLMed, demonstrate that FairGE consistently reduces both marginal and intersectional disparities, improves worst-group performance, and largely preserves overall predictive accuracy. Code will be released.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2301_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/yuchenshao749/FairGE

Link to the Dataset(s)

N/A

BibTex

@InProceedings{ShaYuc_FairGE_MICCAI2026,
        author = { Shao, Yuchen AND Xie, Yutong AND Chen, Jian AND Chen, Qi},
        title = { { FairGE: Gated Expert Routing for Intersectional Fairness in Medical Foundation Models } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16878},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The paper presents a novel framework for evaluating international fairness in medical foundation models. Experiments are conducted on two large, publicly available datasets, and the authors state that the code will be made available. The results demonstrate improved performance both overall and across stratified analyses.

  • 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 work presents a novel application aimed at advancing and more rigorously investigating intersectional bias in medical AI foundational models.
    • A major strength of the paper is its focus on foundation models, which enhances the methodological relevance and supports broad applicability across diverse domains and clinical tasks.
    • The paper also introduces a novel methodological formulation, particularly in Section 2.1, where the multi-attribute gated distribution mixture-of-experts framework is defined.
  • 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.

    Please consider including additional details in the experimental design (Section 3.1). In particular, it should be clarified whether age and sex were considered when defining the train/validation/test splits and whether stratified sampling was applied.

    Reporting the number of subjects in each subgroup in the Results section would further strengthen the validity of the presented findings. For example, in Table 1, the stratified results for younger subjects (age < 60) appear to be stronger, raising the question of whether this group is over- or underrepresented in the dataset. A similar analysis and reporting of subgroup distributions would be valuable for interpreting the remaining stratified results.

  • 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.

    (5) Accept — should be accepted, independent of rebuttal

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

    Strong paper; with clear novelty and contributions; paper also well organized and present results in a clear way. Minor comments only.

  • 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 #2

  • Please describe the contribution of the paper

    The authors propose FedGE, which introduce multi-attribute gated distribution-aware MoE (MA-dMoE) and multi-attribute purification (MAP) modules to debias medical FMs. Experiments on two benchmarks show that FairGE reduces intersectional disparities while preserving overall performance.

  • 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 is strongly motivated by the need to ensure fairness in pre-trained medical FMs by adopting debiasing strategies.
    • The proposed MA-dMoE and MAP modules are quite interesting, described clearly, and well-adapted to the fairness task. Their combined effect on improved fairness metrics is demonstrated through ablation study.
    • The experiments are sufficient, demonstrating FairGE’s performance compared to baselines on two well-established baselines and extensive ablation study on core hyperparameters.
  • 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.
    • Restricting a higher multi-attribute routing weight for age vs sex is possibly a source of selection bias, as this assumes that FairGE performs best under settings where disparities are known. For example, it would be quite interesting to see how FairGE performs with alpha_age=alpha_sex=0.5 and when alpha_age < alpha_sex.
    • What is the compute required to train MA-dMoE and MAP across L=6 layers? Table 4 could be expanded to show scaling curve with respect to AUC-gap and ES-AUC metrics.
    • Expanding upon the prior point: Compute is a consideration if the proposed method’s success requires re-training large chunks of the medical FM.
    • Can the method scale to multi-attribute (e.g., age-sex-race)? In such cases, sample sizes decrease dramatically. How would the method handle this?
    • The paper claims to report accuracy metrics but no results are provided. If only AUC is being reported, the abstract should be modified: “overall predictive accuracy” to “overall AUC” or equivalent.
  • 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
    • Tables 1 & 2 should bold the best performance per metric. For intersectional AUC, baselines achieve higher AUC than FairGE (e.g., Base for A0-M in Table 1).
    • Could the authors comment on why AUC-gap decreases from L=2 to 6 and then increases after? Is the observation that attribute-specific features start dominating after some initial layers in the Transformer? Exploring attention maps could be interesting.
  • 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?

    The paper is quite interesting and well-motivated. There some minor concerns about experiments and scalability.

  • 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

    This paper proposes FairGE, a lightweight fairness-aware adaptation framework for frozen medical foundation model visual encoders. The framework consists of two modules: (1) Multi-attribute Gated distribution-aware Mixture-of-Experts (MA-dMoE), which inserts shared expert adapters into the last few transformer blocks and uses per-attribute routers with learned aggregation weights to enable intersectional demographic adaptation; and (2) Multi-attribute Purification (MAP), which applies a Purifier MLP combined with multiple adversarial attribute classifiers via gradient reversal to suppress sensitive demographic leakage in the final representation. The approach is validated on HAM10000 (dermoscopy, age–sex) and Harvard-FairVLMed (fundus, race–gender), demonstrating consistent improvements in intersectional subgroup AUC, reduced AUC-gap, and improved equity-scaled AUC while largely preserving overall predictive accuracy.

  • 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.

    S1.Timely and well-motivated problem formulation. Intersectional fairness — where bias concentrates in subgroups defined by combinations of demographic attributes — is a critical but under-studied problem in medical AI. While most existing work (FairMedFM [NeurIPS 2024], FairCLIP [CVPR 2024], FairSDE [AAAI 2026]) addresses single-attribute fairness, this paper explicitly targets multi-attribute intersectional settings. The two challenges identified — subgroup data sparsity and sensitive attribute leakage — are genuine and clearly articulated. S2.Principled and elegant architectural design. The MA-dMoE module provides a neat solution to the sparsity problem: rather than training separate adapters per intersectional subgroup (which would overfit on small subgroups), the design shares expert parameters across all groups and only differentiates through per-attribute routing gates with learned aggregation weights. This is parameter-efficient and avoids the combinatorial explosion of subgroup-specific modules. The complementary MAP module with per-attribute adversarial classifiers via gradient reversal addresses the orthogonal problem of demographic leakage in a principled manner. S3.Comprehensive and well-structured experimental evaluation. The paper reports a thorough set of metrics (overall AUC, AUC-gap, ES-AUC, DEOdds) across two distinct benchmarks covering different modalities (dermoscopy and fundus), different tasks (skin lesion classification and glaucoma detection), and different attribute combinations (age–sex and race–gender). The consistent improvement pattern across both datasets is convincing. FairGE achieves the best fairness–accuracy trade-off compared to six baselines spanning traditional, debiasing, and CLIP-based fairness methods. S4.Thorough ablation studies. The ablation in Table 3 clearly isolates the contribution of each component (MA-dMoE, group-gating, MAP, single vs. multi-attribute gating), and Tables 4–5 provide useful design guidance on the number of adapted blocks and the attribute weight ratio. The progressive improvement from Base → w/o MA-dMoE → w/o group-gating → w/o MAP → FairGE-SA → FairGE is clean and informative. S5.Strong quantitative results. The reduction in AUC-gap from 11.98 to 3.19 on HAM10000 (a 73% relative reduction) while simultaneously improving overall AUC from 90.56 to 92.07 is noteworthy. Many fairness methods sacrifice accuracy for equity; FairGE achieves both. The worst-group AUC improves substantially (82.18 → 89.64 for A1-F on HAM10000), demonstrating real clinical value.

  • 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.

    W1.Limited diversity of foundation model backbones. The paper does not specify which pretrained medical foundation model encoder is used, and apparently evaluates on only one backbone. Given that the framework is proposed as a general-purpose adapter for “medical foundation models,” it is important to validate on multiple backbones (e.g., BiomedCLIP, LVM-Med, DINOv2-medical, MedMAE) to demonstrate generalizability. The dMoE paper (Oh et al., 2025; ICML 2025) already demonstrates integration with multiple architectures — FairGE should do the same. This is a notable gap for a methods paper. W2.Only two datasets with limited attribute diversity. While HAM10000 and Harvard-FairVLMed are standard benchmarks in the fairness literature, both are relatively small (~10K images each) and cover only binary classification. The attribute combinations tested are limited to age–sex and race–gender. Extending to more attributes (e.g., socioeconomic proxies, insurance type in CheXpert/MIMIC-CXR) or to multi-class tasks would strengthen the claims. In particular, the Harvard-FairVLMed experiment uses race–gender but the paper primarily motivates with age–sex — a more consistent treatment would be desirable. W3.The multi-attribute gating weight α is manually tuned, not learned. Table 5 shows a grid search over (α_age, α_sex) with the constraint α_age ≥ 0.6.This manual tuning is somewhat ad-hoc and may not generalize to settings with more than two attributes (e.g., age × sex × race). The paper should discuss whether α can be learned end-to-end (e.g., via a softmax over learnable parameters) and how the approach scales to J > 2 attributes. W4.Missing statistical significance analysis. All results appear to be from single runs without error bars or confidence intervals. Given the small subgroup sizes (especially in intersectional groups), variance across random seeds could be non-trivial. Reporting mean ± std over multiple runs would strengthen the claims. W5.The adversarial purification (MAP) convergence and effectiveness deserve deeper analysis. Gradient reversal is notoriously difficult to tune (the λ_j coefficients) and can suffer from training instability. The paper does not report: (a) how λ_j values were selected; (b) the adversarial classifiers’ accuracy at convergence (i.e., how well demographic information is actually suppressed); or (c) any t-SNE or feature-space visualization showing the purification effect. Without this evidence, the claim that MAP “suppresses sensitive demographic information” remains insufficiently supported. W6.Comparison with concurrent/related MoE-based fairness methods is incomplete. Fair-MoE (Wang et al., MICCAI 2025) and FairMoE (Liang et al., MICCAI 2025) are highly related concurrent works that also use MoE for fairness in medical imaging. While the paper cites Fair-MoE [21], a direct experimental comparison on a shared benchmark (Harvard-FairVLMed) would be highly informative and is feasible since both methods target the same dataset. The paper would also benefit from a clearer conceptual comparison: Fair-MoE uses MoE for bias filtering in VLM encoders via a fairness-oriented loss, while FairGE uses MoE for demographic-aware adaptation with adversarial purification — these are philosophically different approaches and the distinction should be highlighted.

  • 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 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

    The paper focuses on classification; extending FairGE to segmentation (where dMoE has already been validated) would be a natural and impactful direction. Minor: Equation numbering and referencing could be improved — Eq. (3) for pooling feels unnecessary. The connection between MA-dMoE and the optimal control perspective from the original dMoE paper is not discussed. It would enrich the paper to comment on how multi-attribute gating relates to the control-theoretic interpretation. For Harvard-FairVLMed, the DualFairVL and CMAC-MMD baselines are missing (Table 2). It would be helpful to clarify why these were omitted.

  • 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.

    (5) Accept — should be accepted, independent of rebuttal

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

    This paper addresses a genuinely important and under-explored problem — intersectional fairness in medical foundation models — with an elegant, principled, and well-validated solution. The two-module design (MA-dMoE for demographic-aware adaptation + MAP for attribute purification) is complementary and well-motivated. The experimental results are strong, with consistent and substantial improvements across two benchmarks and a thorough ablation study. The main limitations — single backbone, limited dataset diversity, manual α tuning, and missing statistical significance — are addressable and do not undermine the core contribution. The problem formulation alone — shifting the community’s attention from single-attribute to intersectional fairness — has significant value. I recommend acceptance.

  • 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



Author Feedback

We sincerely thank the reviewers for their positive and constructive feedback. We appreciate their recognition of the importance of intersectional fairness in medical foundation models, the value of our MA-dMoE and MAP modules, and the strength of our experimental evaluation. Below, we address the remaining comments and clarify the planned revisions.

Reviewer #1. Q1.More details for experiments. We used stratified sampling when constructing the train/validation/test splits, and the distributions of the sensitive attributes, including age and sex, were considered to avoid severe subgroup imbalance across splits. Q2.Analysis of the number of subjects. The stronger performance for younger subjects (age<60) in Table 1 is related to the subgroup distribution in HAM10000, where younger and older subjects are approximately 2.5:1.More younger samples provide more stable training and evaluation. We will report subgroup sample sizes and conduct an analysis in the final version.

Reviewer #2 Q1.Attribute routing weights As described in the paper, since age exhibits larger disparities than sex in our main results (i.e., the performance of the Base model in Tables 1 and 3), we consider settings with alpha_age > alpha_sex​. We also evaluated the equal-weight setting alpha_age=alpha_sex=0.5, where our FairGE still reduces the AUC-gap to 4.33, outperforming all baseline methods. This suggests that FairGE remains effective even without prior assumptions about which attribute contributes more strongly to the disparity. Q2.Computational cost. The computational cost of FairGE is modest in our setting. Since we adopt parameter-efficient adaptation with linear probing and only train the last L=6 adapted blocks together with lightweight routing and purification modules , all experiments can be conducted on a single NVIDIA RTX 3090 GPU. Q3.Scalability to more attributes FairGE can naturally scale to more than two sensitive attributes. The reason is that each attribute has its own attribute-specific routers, while the experts are shared across all groups. The final routing combines the active single-attribute routing distributions, rather than learning separate parameters for each intersectional subgroup. Thus, FairGE mainly requires sufficient samples for each individual attribute rather than every intersectional subgroup. Q4.Minor issue. We will revise the wording to avoid ambiguity.

Reviewer #3 Q1.Experiments for more backbones. In this work, we use the visual encoder of InfiMed as the medical foundation model backbone. Besides, we have also verified the effectiveness of FairGE on CLIP-style backbones, where we observed consistent improvements in fairness metrics,e.g., the AUC gap narrowed to 3.76%. Due to the page limit, the main paper focuses on the current backbone. Q2.Only two datasets. Besides HAM10000 and Harvard-FairVLMed, we also evaluated our FairGE on another skin disease dataset, i.e., BCN20000, and observed consistent effectiveness, for example, the AUC gap decreased from 11.13% to 5.19%. Q3.Gating weight α In this study, manually selected routing weights α already achieve strong performance. We agree that learning \alpha end-to-end is a promising direction. Q4.Reporting mean ± std We will include statistics from three runs in the final version. Q5.MAP effectiveness a) λ_j values. In our experiments, we selected λ=0.01 based on validation performance. b) How well demographic information is actually suppressed. The age prediction accuracy decreased from 0.901 to 0.732, and the sex prediction accuracy decreased from 0.566 to 0.506, indicating that MAP effectively suppresses demographic information. c) Visualization. We will consider adding a compact feature-space visualization in the final version if space permits. Q6.More MoE-based fairness methods. We also compared our FairGE with Fair-MoE on a CLIP backbone. FairGE achieved a higher overall AUC (0.814vs. 0.773) and a smaller race-related AUC gap (2.88%vs.5.39%).



Meta-Review

Meta-review #1

  • Your recommendation

    Provisional Accept

  • 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.

    All the three reviewers strongly agree that the paper deals with an important problem (intersectional fairness) and makes a solid technical contribution (via gated distribution-aware mixture of experts). The experimental results are also convincing.



back to top