List of Papers Browse by Subject Areas Author List
Abstract
While foundation models like SAM have advanced medical image segmentation, recent evidence raises safety concerns about subgroup disparities in their medical adaptation. Existing fairness strategies face a critical dilemma when adapting these massive models via Parameter-Efficient Fine-Tuning (PEFT): objective reweighting within PEFT’s highly constrained parameter space risks severe accuracy degradation due to feature conflicts, while representation disentanglement often relies on unstable adversarial training that complicates PEFT convergence.
To bridge this gap, we propose AEGIS Anatomy-Embedded Group-Invariant Segmentation, a novel framework adapting generalist models for equitable Head and Neck Squamous Cell Carcinoma (HNSCC) segmentation. AEGIS introduces a Dual-Synergistic optimization strategy that effectively harmonizes representation purification and dynamic loss adjustment. Specifically, bypassing complex adversarial networks, we introduce an Attribute-Conditional Feature Alignment module. Guided by prior attribute metadata, it selectively aligns prototype features: preserving essential biological variations (e.g., gender-specific anatomy) while suppressing non-biological noise (e.g., site-specific scanner variations). Built upon this explicitly disentangled feature space, our History-Aware Dynamic Weighting scheme stabilizes reweighting against batch noise via historical temperature scaling, safely targeting long-tail subgroups while maintaining peak generalist performance.
Extensive experiments on an in-house multi-center HNSCC cohort and external out-of-distribution (OOD) benchmarks demonstrate that AEGIS achieves a favorable empirical trade-off between segmentation accuracy and group fairness.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0597_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/lettuce09/AEGIS
Link to the Dataset(s)
PI-CAI: https://pi-cai.grand-challenge.org/
Harvard-FairSeg: https://github.com/luoyan407/FairSeg
BibTex
@InProceedings{WanSen_AEGIS_MICCAI2026,
author = { Wang, Sen AND Zhan, Zhaoyi AND Xing, Zheng AND Zhang, Minqing AND Jiang, Chloe Meng AND Wang, Zhiliang AND Zhang, Ruimao AND Duan, Xiaohui AND Zhao, Weibing},
title = { { AEGIS: Anatomy-Embedded Group-Invariant Segmentation for Fair Medical Foundation Models } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16879},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes AEGIS, a method for improving the fairness of medical image segmentation models by using a 2D slice-based parameter-efficient fine-tuning (PEFT) paradigm. AEGIS improves fairness across different demographic groups while maintaining segmentation performance. This work provides a robust and scalable solution for addressing unfairness in medical image segmentation, which is crucial for ensuring equitable healthcare outcomes.
- 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 proposed method is innovative and addresses an important issue in medical image segmentation, which is the unfairness across different demographic groups. The use of a 2D slice-based PEFT paradigm is a novel approach that allows for efficient fine-tuning of segmentation models while improving fairness. The experimental results on the three datasets demonstrate the effectiveness of AEGIS in improving fairness while maintaining segmentation performance. This work has the potential to make a significant impact on the field of medical image segmentation and contribute to more equitable healthcare outcomes.
- 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.Some of the references should be the published papers, instead of arXiv preprints. 2.In the introduction, the authors use references [11, 12, 14, 15, 18] to support the claim that foundation models suffer from unfairness. However, some of the references do not involve foundation models. 3.The word “genuine” is too absolute. 4.The authors use \tau=2 as the pixel threshold, but it is not clear how this threshold is determined. Additional ablation studies on the choice of \tau or other explanations would be helpful to understand its impact on the performance of the proposed method. 5.The authors claim that their 2D slice-based PEFT paradigm can support large batches. However, the key factor that determines the batch size is the GPU memory. Besides, I am curious about whether this method can be expended to 3D segmentators, and how to process the 2D results to obtain the final 3D segmentation results? 6.The font size of the figures is too small, which makes it difficult to read the details. It is recommended to increase the font size for better readability. 7.The authors only use DSC and EAGap as evaluation metrics, other metrics regarding the trade-off between accuracy and fairness should also be considered, such as ES-Dice. 8.I am curious about the input bounding box prompt. How is the bounding box generated? 9.GroupDRO is an algorithm designed for classification tasks, what backbone is used for the segmentation task? 10.In table 2, the nnU-Net reports a DSC of 53.9, which is much lower than the results reported in the challenge website. It would be helpful to clarify the reason for this discrepancy, such as whether the same training and testing data are used, and whether the same evaluation metrics are applied. 11.The authors claim that their method achieves Pareto-optimal results, but no theoretical analysis is provided to support this claim.
- 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 does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure 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.
(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?
Although this paper has some weaknesses in manuscript writing and method description, the idea is interesting and the target task is important.
- Reviewer confidence
Very confident (4)
- [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 explained all my concerns, so I recommend to accept this paper.
Review #2
- Please describe the contribution of the paper
This study focuses on the fairness-awareness of the foundation models (FMs) optimized for the medical segmentation problem. The paper proposes an optimization strategy named as ‘Dual-Synergistic’ based on the MedSegX architecture to tackle down some problems in the existing Parameter-Efficient Fine-Tuning (PEFT) adapted FMs. The methodology especially focus on two common issues rised with FMs such as SAM when fairness-awareness is critical in medical segmentation task: 1) Considering even noisy features rather than anatomic-only ones, 2) Optimization with a limited data distribution ending up with under-represented diversity. The proposed framework, named as ‘Anatomy-Embedded Group-Invariant Segmentation’ (AEGIS), addresses these issues by introducing two mechanisms called as ‘Attribute-Conditional Feature Alignment’ (ACFA) and ‘History-Aware Dynamic Weighting’ (HADW). These mechanisms are designed as loss terms and plugged into the main objective function in addition to the segmentation task that is optimized by Binary Cross-Entropy (BCE) and Dice losses. ACFA is responsible to be salient on biological anatomy related features (gender specific etc.) while suppressing the noisy features (site specific etc.). HADW then considers purified/anatomical features and focuses on the worst-performing sub-groups temporally to reach the larger data analysis during training steps. Experimental analysis is performed on three datasets with one in-house (The Head and Neck Squamous Cell Carcinoma) and two cross-datasets (Harvard-FairSeg and PI-CAI) using off-the-shelf models such as nnU-Net, MedSAM, MedSegX. An ablation study is also conducted to demonstrate the contribution of the framework components (ACFA and HADW).
- 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 study addresses an important concern in FMs regarding awareness of the fairness by introducing several mechanisms (ACFA and HADW) in a framework named as AEGIS. Isolation of anatomical features represented with sensitive attributes by ACFA module is a moderately novel approach and purifies the optimization for a fairer model construction. HADW is then applied to purified features for dynamically adjusting the loss weighting by considering data from worst-performing sub-groups and reduces the disparities across members inside sub-groups. These mechanisms allow for smaller disparity gaps while retaining segmentation performance as higher as possible. Proposed framework is extensively tested on three different use-cases in comparison with existing/well-known models. The method outperforms those off-the-shelf models mostly in subgroup disparities without sacrificing on segmentation performance evaluated by Dice score. The ablation study also verifies the positive performance contribution of the components. The paper is well organized and the methodology is explained clearly.
- 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.
In the paper, the major flaw is the misleading analysis for Pareto Frontier (PF). According to Figure 2 (a-c), it is claimed that AEGIS outperforms other systems in terms of utility-fairness trade-off performance, which is indicated as performance-fairness PF, but there is only one operating point/model for each system. Contrary; given a system, PF is constructed as a subset of non-dominated points/models among all points/models generated by that system. So, each model should be a different realization of the system (AEGIS, MedSAM etc.) regularized by some fairness hyper-parameter (lambda) or system threshold etc., and surviving/non-dominated points should create the frontier. The comparative analysis should be then performed between system frontier curves based on specialized metrics/tools (hypervolume, uniformity/fairical etc.). This part must be revised in the paper accordingly by dropping the claim or changing it to reflect true Pareto optimality analysis. Otherwise, the current plot is just a visualization of Table 1 and misleading about the superiority of the AEGIS on Pareto optimality.
That would be also a good extension to test the proposed framework with different FMs and sensitive attributes higher than two members (non-binary) such as race (Harvard FairSeg has this attribute with three sub-groups as Asian, Black, White). The current analysis is only based on MedSegX and sensitive attributes with two sub-groups. A comprehensive analysis with these factors would make the framework more generalizable under different conditions.
Qualitative examples would be supportive to perform visual analysis on predicted segmentation masks across the systems in consideration.
The code and the in-house HNSCC cohort data (if possible) should be also made public for reproducibility purpose. Although the method is clearly explained, the reproducibility is lack because of unavailable data.
It is intuitive but that would be better to put sub-captions (a, b,..) inside Figure 2 to match what a-c is in the text.
- 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 main factor about weak rejection is the misleading Pareto Frontier (PF) analysis. The other parts of the paper look fine and the minor points regarding qualitative demonstration, source code sharing etc. may be fixed easily. However, PF analysis should be carefully handled to finalize the study well. This is the only major flaw in the paper.
- Reviewer confidence
Very confident (4)
- [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.
As the authors address the main flaw regarding their Pareto Frontier claims by promising to update the manuscript accordingly, I would like to change my decision from ‘weak reject’ to ‘accept’. My other comments were about improving the paper for a future extension.
Review #3
- Please describe the contribution of the paper
The paper proposes a fairness-aware framework AEGIS for adapting medical foundation models using PEFT. The core contribution is a dual-stage optimisation strategy that combines by introducing a feature alignment module aims to disentangle inter-group feature conflicts, and a dynamic reweighting schema to prevent the model from ignoring harder long-tail demographics.
The method is employed MedSegX (ViT-B) and evaluated mainly on the Head and Neck Squamous Cell Carcinoma (HNSCC) segmentation on in-house and external Harvard-FairSeg and PI-CAI datasets. The paper claim that AEGIS alleviates the performance-fairness trade-off typically observed in PEFT-based fairness optimisation.
- 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 the fairness-aware degradation problem in medical foundation models, particularly in the context of PEFT for HNSCC segmentation via adaptation of large pre-trained segmentation models.
- The proposed modules, representation alignment and loss reweighting, are conceptually intuitive and align with the problem-specific solution.
- Avoiding adversarial training reduces optimization instability, particularly under PEFT settings.
- Experiments on in-house and OOD datasets, and ablation study, validate the evaluation analysis.
- 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.
-
Fairness objective is not clearly or formally defined. What notion of fairness the method is truly optimizing?
-
Fairness metric choice is limited. The evaluation primarily focus on the EA Gap. The paper does not sufficiently justify why this metric is preferred over standard fairness metrices in medical image segmentation fairness.
-
The proposed method builds on the MedSegX. However, the baseline MedSegX already achieves comparable or better performance on overall Dice score, EA Gap for minority subgroups (e.g., Female and Young). This raises concern over whether the observed fairness EA gap reduction stems from genuine fairness optimization or from degraded baseline subgroup performance introduced by the proposed modifications. A clear analysis and justification is required.
Minor weaknesses: Hyperparameters are fixed without sufficient justification or sensitivity analysis. Page 3, τ is set to 2; page 4, β = 0.9; λalign = 0.1, λf air = 1.0, and λbg = 0.05, and T =0.1, the lower temperature value plays a significant role.
-
- 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 mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure 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?
I recommended Weak Reject. The lack of clarity in the fairness objective, non-standard fairness metrices, and insufficient justification of the EA Gap performance gain. Use of non-standard fairness metrices weakens the interpretability and comparability of the fairness claim. There is required a clear fairness objective they are optimizing and explain how it relates to standard fairness notions in medical imaging.
The major concerns are already mentioned.
- 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.
I thank the authors for their detailed responses. I noted a mathematical discrepancy between the text definition of the Equal Accuracy Gap ((\Delta = \max_k |\text{Dice}k - \text{Dice}{\text{overall}}|)) in Section 3.2 and the actual values printed in the table (quantitative results provided in Table 1).
The printed values for (\Delta ) across all datasets consistently reflect the Max-Min Gap ((\max_k \text{Dice}_k - \min_k \text{Dice}_k)) rather than deviation from the overall mean. For instance, in the PI-CAI site baseline, the table reports a gap of 9.8% (exactly (58.1\% - 48.3\%)).
I recommend to fix this discrepancy in the final version.
Author Feedback
We thank the reviewers and AC for the careful reading.
R1/R2:Pareto-front claim. A:We thank the reviewers for identifying this issue.We agree that “Pareto Front/Pareto-optimal” is too strong for Fig.2(a-c),because each method is one operating point rather than a frontier obtained by sweeping a fairness-control parameter.We will revise the title,caption,and conclusion to“performance-fairness trade-off analysis”and remove the formal Pareto-optimality claim. Frontier-level tools such as hypervolume/uniformity are thus not applicable to Fig.2(a-c).This is a terminology correction,not a change to the empirical conclusion:AEGIS remains in the favorable high-Dice/low-gap region, reducing the maximum EA Gap while preserving competitive/best overall Dice.
R1/R4:Fairness objective and metrics. A:We respectfully disagree that our evaluation is unsupported or based on one non-standard metric.Medical segmentation fairness has no universally adopted metric analogous to Dice/HD95;metrics should match the fairness notion and task.Our notion is group-wise segmentation accuracy parity:reducing group-conditioned Dice disparities while preserving overall Dice.This objective is implemented in Eq. (1):L_task preserves utility,while L_align and L_fair reduce inter-group feature conflicts and emphasize historically underperforming subgroups.EA Gap, defined in Sec.3.2,directly measures the largest subgroup-level Dice deviation targeted by AEGIS.The evaluation is not “only EA Gap”:Table 1 reports overall Dice,subgroup Dice,and EA Gap jointly.ES-Dice is a useful complementary composite metric derived from overall performance and group disparities;since its ingredients are already reported,it requires no extra experiments.Our decomposed reporting is more transparent because it shows which subgroup improves and whether utility drops,rather than collapsing them into one score.
The results do not indicate fairness via uniform degradation:HNSCC preserves overall Dice while reducing all gaps;PI-CAI improves both overall Dice and site gap;Harvard-FairSeg shows marginal overall-Dice change while improving the lower-performing older subgroup and reducing the age gap.Thus, gains come from improving/preserving underperforming subgroups with little or no overall-Dice loss,not from uniformly degrading all groups.
R1/R4:Implementation details. A:GroupDRO use the identical MedSegX-PEFT backbone.The box prompt follows MedSegX:a GT-derived tight box with up to 20-pixel perturbation during training and no perturbation during evaluation for all prompt-based methods.This is an oracle protocol for controlled benchmarking,not a deployment assumption.AEGIS follows the MedSegX 2D protocol:3D MRI volumes are sliced into 2D axial images;Dice is computed on 2D predictions and averaged over samples/subgroups.We did not use the official PI-CAI 3D detection/diagnosis pipeline;leaderboard numbers are not directly comparable to our Dice-based OOD segmentation evaluation,explaining the apparent nnU-Net discrepancy.τ=2 affects only foreground-prototype construction.β=0.9 stabilizes EMA,T=0.1 focuses on harder groups,and λ_align/λ_fair/λ_bg balance fairness terms with L_task.These values were chosen from validation performance/training stability,then fixed for all reported test evaluations.
R1/R2:Scope,presentation,and reproducibility. A:More foundation models,non-binary attributes,and qualitative masks would strengthen generality.Under rebuttal rules we do not add new results,but the submission already covers one foundation-model backbone,three datasets,in-domain/OOD settings, and gender/age/site attributes. We will update arXiv-only references where possible,separate FM-specific from broader bias evidence,replace “genuine”,enlarge figure fonts, and improve readability.Fig.2 already includes sublabels (a-f);we will make them more visible.We will release all code upon acceptance;private HNSCC data may be requested by email subject to institutional approval and a data-use agreement.
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.
Disparate scores from reviewers, who also differentially report utility (R4: MedSegX performs as well or better on test cases, R1: “improve[s] fairness and maintains seg performance”). Rebuttal with author’s response is sensible here to attempt to clarify.
- 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 reviewers are in agreement. However, please fix the error noticed by R4.I have determined this to be acceptable conditional on a fix in the camera ready.
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.
After reading the rebuttal, I recommend Accept. The paper gives a concrete PEFT framework for fairness-aware medical segmentation. Its two main parts are easy to identify: ACFA changes the feature space used by the segmentation model, and HADW changes how subgroup errors are weighted during training. The experiments on the in-house HNSCC cohort and the two external datasets show smaller subgroup gaps while keeping Dice in a similar range, and the ablations indicate that both parts contribute.
The main concerns raised in the reviews were about the way the results were presented rather than about a clear failure of the method. In particular, the Pareto-frontier analysis was misleading because the paper showed one operating point per method, not a frontier over multiple settings. The rebuttal addressed this point by agreeing to revise the claim and presentation. There is also a remaining notation issue: the reported EAGap values appear to follow a max-min subgroup gap, while the equation in the paper defines a deviation from the overall Dice. This should be fixed in the camera-ready version.
Overall, the method is technically reasonable, the experiments support the main empirical claim, and the remaining issues are concrete manuscript corrections. I would therefore support acceptance, with the expectation that the authors revise the Pareto-frontier language and correct the fairness-gap definition in the final version.
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.
This paper presents a fairness-aware framework AEGIS for adapting medical foundation models. They demonstrate performance using metrics that reviewers found insufficient and by overstating claims for the Pareto-Front performance analysis. During the rebuttal, the authors clarified these two points, especially: 1/ they proposed to change the claims for the Pareto-Front / Pareto-optimal issue ; 2/ they provide clarification on their choice of fairness objective and metrics. Additional points were mentioned by the authors, providing details on the implementation. Therefore, I recommend acceptance of this paper.