List of Papers Browse by Subject Areas Author List
Abstract
Medical image analysis models can exhibit performance disparities across patient subgroups, threatening clinical safety and fairness. Existing methods typically address this issue by optimizing accuracy and fairness metrics for visible demographic attributes (e.g., sex or age) considered in isolation. This strategy not only overlooks potentially more informative latent stratifications, which may reveal deeper sources of model error and inequity, but also fails to scale when multiple demographic attributes are considered simultaneously due to the resulting sparsity of training data within each subgroup. We deal with these issues by introducing the label-free hidden-cohort fairness (LHCF) training paradigm that instead of maximizing fairness over visible demographic attributes, it optimizes fairness across latent subpopulations discovered from image appearance. By clustering images into $K$ appearance‑based cohorts and applying fairness optimization over them, LHCF uncovers underlying sources of model error and avoids the combinatorial sparsity of multi‑demographic attributes, reducing disparities across both single and multiple demographic attributes. We demonstrate on our proposed fairness benchmark, HIDFairBench, that LHCF provides state-of-the-art fairness results on single and multiple demographic attributes without requiring demographic labels for training, improving practicality in real-world clinical settings where metadata may be incomplete or unavailable. Our results position hidden‑cohort fairness as a practical, scalable, and robust alternative to demographic‑based fairness optimization for trustworthy medical image analysis.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4959_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://miladmasroor.github.io/lhcf-project-page/
Link to the Dataset(s)
N/A
BibTex
@InProceedings{MasMil_Fairness_MICCAI2026,
author = { Masroor, Milad AND Nguyen, Cuong AND Wells, Kevin AND Carneiro, Gustavo},
title = { { Fairness Beyond Demographics: Optimizing Performance Across Appearance-Based Hidden Cohorts in Medical Imaging } },
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 proposes a two-step fairness training approach: (1) discovering subgroups based on the latent space, and (2) training with a fairness loss term to regularize these subgroups (e. g. , optimizing for worst-subgroup 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 problem addressed is highly relevant to the MICCAI community.
- Leveraging learned subgroups can unlock model robustness for datasets lacking explicit metadata.
- The proposed method is quite general because it is not restricted to a single fairness objective.
- Notably, the paper demonstrates that using discovered subgroups instead of explicit metadata leads to comparable performance when applying fairness objectives. This is a strong positive indicator of the practical benefits of using such learned subgroups.
- 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 primary weaknesses lie in the presentation of the results and a lack of in-depth discussion. I understand there is limited space, but maybe one of the ablations could be omitted.
- Figure 2 is largely illegible. Readers should not have to zoom in to read the contents. Additionally, the caption should be expanded to better guide the reader through the key takeaways.
- The ablation studies (Figures 3, 4, and 5) would greatly benefit from including the ERM and “Classical” (with metadata) baselines. Currently, it is difficult to assess how these baselines perform across specific subgroups (e. g. , gender, age) because we only have access to their average performance in Table 2.
- I believe the narrative would be stronger if it focused on demonstrating that LHCF is competitive with metadata-based approaches, rather than asserting that it definitively outperforms them. Since Figure 2 does not show statistically significant differences between using the proposed method and metadata, achieving parity with metadata methods is already a solid result. The authors could then highlight the potential gains over ERM in the ablations, using the classical metadata approach as an upper bound.
- Is the subgroup discovery method new or reused from the literature? In both cases, relevant papers should be cited [1]. Similarly, there was prior work doing subgroup discovery followed by bias mitigation, for example [2].
- The paper lacks context regarding the potential impact of these findings in real-world settings.
- Suggestion for future work: Given that the learned subgroup performances do not align with the evaluated metadata, it is quite surprising that the method still leads to improvements (even if marginal) in metadata-based subgroup performances, as shown in Table 2.This dynamic deserves more dedicated exploration.
[1] Eyuboglu, Sabri, et al. “Domino: Discovering systematic errors with cross-modal embeddings. “ ICLR 2022. [2] Ahmed, Faruk, et al. “Systematic generalisation with group invariant predictions. “ ICLR 2020.
- 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 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.
(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 methodology and results are interesting, but the paper would benefit from better presentation to have greater impact. In particular, reframing the claims, include baselines in the ablation, and discuss contributions regarding related work.
- Reviewer confidence
Very confident (4)
- [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 describe a generic fairness/robustness method that works by 1) Clustering training data, and then 2) Applying generic fairness/robustness mitigation techniques relying on group labels using the cluster labels instead of (e.g. demographic) group labels. The authors find that their proposed scheme outperforms the base methods (using existing demographic group labels) across three public datasets and various overall as well as fairness-specific performance metrics.
- 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.
A generic method that is directly applicable in many tasks and settings, and that appears to perform strongly.
It is somewhat surprising and generally interesting (and reassuring) that the clustering+mitigation scheme mostly outperforms the same mitigation methods applied to existing demographic group labels. Besides (rather marginal) improvements in settings in which group labels are available, this particularly opens the door to applying such mitigation methods in the many scenarios in which group attributes are unavailable, rendering bias mitigation schemes vastly more widely applicable. The practical impact of these findings may be quite significant.
- 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 arguably most relevant branch of prior work is not acknowledged or cited; possibly the authors are unaware of it? They are not the first to propose the “unsupervised discovery of groups + subsequent pseudo-supervised robustness mitigation” scheme. Some relevant prior works in the same line of research include but are surely not limited to:
- https://proceedings.neurips.cc/paper/2020/hash/e0688d13958a19e087e123148555e4b4-Abstract.html (2020)
- https://proceedings.mlr.press/v139/creager21a.html (2021)
- https://arxiv.org/abs/2108.02943 (2021)
- https://arxiv.org/abs/2403.13870 (2024)
While this significantly limits the novelty of the present manuscript, I think sufficient contributions remain here: 1) The more general framing described here (as opposed to a very specific instance of this scheme) is useful on its own. 2) The evaluation showing superiority over group-supervised versions of the same mitigation methods are new and interesting.
Still, I believe the authors should clearly acknowledge these earlier works and possibly integrate at least some of them as baselines in their benchmark since they are really very close cousins.
- 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
JTT might be another relevant baseline to include since it is also unsupervised and generally a very popular method. (https://arxiv.org/abs/2107.09044)
Could the authors clarify why they chose to construct a new benchmark as opposed to e.g. using the existing MEDFAIR benchmark (which they cite)?
As opposed to many prior works in this area (see references above under Weaknesses), the authors choose to cluster globally, i.e., not stratified by class labels. While valid in principle, this can result in clusters with only a subset of the target class labels, which can present challenges for many mitigation methods. Could the authors elaborate a) on the reasoning underlying this choice, and b) whether they encountered any challenges due to this? Was this maybe a reason why more “traditional” fairness methods such as TPR/FPR disparity regularization (e.g. https://github.com/aida-ugent/fairret) were not included in the comparison?
No reference or definition is provided for the “Performance-Scaled Disparity (PSD)”.
The authors write that the Brier Score “captures calibration and identifies under- and overperforming cohorts”. The BS captures a combination of calibration and discrimination, i.e. it is not a pure measure of either of them. (See e.g. https://link.springer.com/chapter/10.1007/978-3-319-23528-8_5 as a somewhat random reference on the topic.) In addition, average BS may differ strongly between the different classes. This means that differences in BS between clusters are not necessarily due to differences in model performance; they may simply reflect differences in class balance.
The authors write that “Overall, backbones inducing moderate hidden-cohort granularity, e.g., ResNet18 (K=7) and MedCLIP (K=6) selected by BIC, yield better accuracy and fairness than CLIP (K=10) and DINOv2 (K=5).” - do the authors truly believe that these K numbers are the driving factor between the differences in how these embedding backbones perform in their analyses…? This honestly seems like grasping at straws (for an explanation) to me. It is fine to say that we do not (yet) know why a certain backbone works better for this than another one, if that is the case.
“As an intermediate alternative between Classic and LHCF, DAC clusters samples using embeddings augmented with demographic attributes and then optimizes fairness over these clusters.” - How exactly is this “augmentation” implemented? I cannot find any more details on this. There are some potential issues here: 1) By including such demographic labels, the clustering problem becomes mixed (real+integer), which is generally challenging to handle well with GMM. 2) The impact of demographic labels on the results will in general depend on the weighting / scaling of these added dimensions, no?
The font size in all figures is far too small.
- 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?
Well-written and robust paper with interesting findings and potentially significant impact. The major weakness is the lack of consideration of prior works from the cluster->mitigation branch of the literature.
- Reviewer confidence
Very confident (4)
- [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 main contribution is a label-free hidden-cohort fairness (LHCF) framework for medical imaging. The method first clusters image embeddings into appearance-based latent cohorts, then applies fairness-aware training across those cohorts instead of across observed demographic groups. The paper argues that this can reduce the sparsity and scalability problems of demographic and intersectional fairness training while still improving fairness on visible demographic groups.
- 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.Important problem.
The paper targets a real limitation of current medical imaging fairness work: many methods rely on demographic labels that are missing, unreliable, restricted, or too sparse for stable intersectional optimization.
2.Broad empirical scope within the submission.
The paper evaluates across three public datasets, multiple fairness methods, several backbones, multiple visible demographic partitions, and intersectional groups. The ablations on the number of clusters, backbone choice, and DAC vs. LHCF strengthen the overall empirical story.
3.Consistent empirical trend
The strongest gains appear in the harder intersectional settings, which is exactly where visible-group approaches tend to suffer from sparsity. This makes the empirical findings more compelling.
4.Compatibility with existing fairness methods.
LHCF is presented as a wrapper around existing fairness objectives rather than a one-off architecture, which improves the generality of the idea.
- 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.Potential pipeline ambiguity / leakage concerns.
The manuscript is not sufficiently explicit about whether pretraining, GMM fitting, BIC model selection, hidden-cohort assignment, and checkpoint selection are all performed in a fully split-aware way. This matters a great deal in fairness evaluation, and the paper should make the train/val/test protocol completely explicit in the rebuttal.
2.Alternative explanation not fully ruled out.
The hidden cohorts may be capturing label prevalence or task difficulty rather than fairness-relevant latent patient structure. The paper needs stronger controls such as class-conditional clustering, label-predictability of cluster ID, semantic cohort analysis, or matched random-cluster controls.
- 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
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?
The paper tackles an important real problem in medical imaging fairness. The core idea is meaningfully novel in this context, and the method is simple and broadly compatible with existing fairness pipelines.
Some parts of the train/val/test protocol and clustering pipeline do require clearer specification, and stronger controls would help rule out alternative explanations such as clustering by difficulty.
Overall the paper contributes a useful idea with enough empirical support to merit inclusion at the conference.
- 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 thank the reviewers and meta-reviewers for their positive evaluation of our work and for the constructive feedback. We are encouraged that the novelty, clinical relevance, and fairness-aware modelling without demographic labels were well recognized. We carefully revised the manuscript to improve clarity, contextualization, and presentation. To better motivate the practical significance of our approach, we expanded the discussion of the real-world impact of LHCF in both the Abstract and Introduction, emphasizing its applicability in clinical settings where demographic metadata may be missing, noisy, incomplete, or unavailable due to privacy constraints. We also clarified the implementation details of Demographic-Aware Clustering (DAC), including how demographic information is incorporated into the clustering pipeline through metadata-augmented embeddings. In addition, we updated the description of the Brier Score to clarify that it captures a combination of calibration and discrimination. To address concerns regarding split-aware evaluation, we will clarify the training protocol and document the details of pretraining, GMM fitting, BIC-based cohort selection, hidden-cohort assignment, and checkpoint selection procedures in an online code repository link that will be indicated in the paper. We additionally improved the presentation of the manuscript by correcting LHCF/LCHF typos throughout the paper, adding ERM and Classic FairDi baselines to Fig.~5, and incorporating several relevant references suggested by the reviewers where space permitted. Regarding Fig.~2, we clarified that the compact visualization is primarily due to MICCAI page limitations, while the figure quality remains high and can be clearly inspected when zoomed. We again sincerely thank the reviewers and meta-reviewers for their insightful comments and constructive suggestions, which helped improve the clarity and presentation of the paper.
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 reviewers acknowledged its novelty in fairness-aware medical image modeling without demographic labels, solid experimental validation across multiple datasets, and high clinical relevance. The weaknesses raised—including incomplete citations, figure readability, missing baselines, and unclear descriptions—are all minor and fully addressable in revision.