List of Papers Browse by Subject Areas Author List
Abstract
Training models on heterogeneous medical data often results in performance disparities across demographic groups. In skin disease diagnosis, these disparities are particularly challenging since sensitive attributes can correlate with clinically relevant factors, making traditional suppression-based fairness interventions potentially harmful to diagnostic accuracy. This limitation motivates us to revisit this problem through a multi-objective optimization perspective, where each group defines a performance objective within a shared backbone. When aggregate performance is optimized under shared parameterization, the learned representations may under-express group-specific characteristics due to shared capacity allocation. To mitigate such effect, we propose a group-specific interference masking framework that enables conditional parameter utilization without altering the backbone architecture. The method identifies representation components that lack stable alignment with a given group and selectively modulates them without introducing additional trainable parameters. Experiment results demonstrate that the proposed approach consistently reduces inter-group performance gaps while maintaining competitive overall accuracy. Our code is available at https://anonymous.4open.science/r/GIM-FAC9/.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1968_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Gracellgg/Group-Interference-Marking
Link to the Dataset(s)
Fitzpatrick17k dataset: https://github.com/mattgroh/fitzpatrick17k
ISIC 2019 dataset: https://challenge.isic-archive.com/landing/2019/
BibTex
@InProceedings{XuGel_GroupConditioned_MICCAI2026,
author = { Xu, Gelei AND Fan, Haoran AND Shi, Yiyu},
title = { { Group-Conditioned Representation Modulation for Fair Skin Disease Diagnosis } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16886},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The main contribution of this article is the introduction of a multi-objective optimization perspective on fairness in skin disease diagnosis, diverging from traditional methods that suppress sensitive attributes which often contain clinically crucial diagnostic signals. The authors propose Group-specific Interference Masking, a framework that addresses the suboptimal representation of under-expressed demographic groups caused by shared capacity allocation in standard joint training
- 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 major strength is Group-specific Interference Masking (GIM). GIM calculates a group-specificity score to identify and selectively mask convolutional channels that lack stable alignment with a given demographic group during fine-tuning. This method is highly parameter-efficient, as it operates entirely on a pretrained backbone, keeps all weights shared across groups, and introduces absolutely no additional trainable parameters or inference overhead. Through comprehensive experiments on the Fitzpatrick-17k and ISIC 2019 datasets, the authors demonstrate that GIM successfully reduces inter-group performance disparities while maintaining highly competitive overall diagnostic accuracy, thereby achieving a significantly improved accuracy-fairness trade-off compared to existing bias-mitigation baselines
- 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 core concept of conditional representation modulation via channel masking is not a groundbreaking paradigm shift. The authors themselves compare their work to existing baselines like SCP, which also performs channel pruning to remove disparity inducing channels, and FairPrune, which prunes parameters whose importance differs across groups. The primary technical novelty here is limited to the specific calculation of a “group-specificity score” using prototype cosine similarity to generate the binary masks. The framework’s experimental validation is entirely restricted to VGG-11 and ResNet-18, so evaluating a new computer vision mechanism exclusively on older, low-capacity models is a major limitation for a top-tier modern conference. The method exhibits severe sensitivity to the mask ratio (r). The methodology is formulated and tested strictly around binary demographic groups (g∈{0,1}). To achieve this, the authors artificially binarize complex variables, compressing the Fitzpatrick 17k dataset into just “light skin” (types 1-3) and “dark skin” (types 4-6), and limiting the ISIC 2019 dataset to binary sex
- 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.
(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?
While the proposed GIM framework presents a parameter efficient approach to addressing shared capacity allocation in fair skin disease diagnosis, its contributions are ultimately too incremental. The core mechanism of channel masking heavily overlaps with existing baselines like SCP and FairPrune , with the primary technical novelty restricted to the prototype cosine-similarity scoring. Furthermore, the experimental validation is severely limited by modern standards, as the authors rely exclusively on outdated, low-capacity architectures like VGG-11 and ResNet-18, leaving the method’s scalability to contemporary models unproven. The framework also demonstrates extreme hyperparameter brittleness, achieving its optimal accuracy-fairness balance at a very narrow mask ratio of r=2% before predictive performance begins to degrade at just r=4%. Finally, the evaluation oversimplifies complex demographic variables into binary attributes and completely lacks an intersectional fairness analysis, which undermines its practical clinical relevance and the authors’ unsupported claims of multi-group extensibility
- 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 main contribution of this paper is the introduction of Group-specific Interference Masking (GIM), a novel framework designed to reduce performance disparities in skin disease diagnosis across demographic groups without sacrificing overall diagnostic accuracy. The core components of this contribution include: A Multi-Objective Optimization Perspective: The authors shift away from traditional “suppression-based” fairness interventions, which often attempt to hide sensitive attributes like skin type. Instead, they reframe fairness as a multi-objective optimization problem where each demographic group represents a distinct performance objective within a shared model backbone. Group-specific Interference Masking (GIM): This framework enables conditional parameter utilization. It identifies specific convolutional channels that lack stable alignment with a particular demographic group and selectively suppresses them using binary masks during fine-tuning. Channel Utility Scoring: The paper proposes a method to quantify “group-conditional utility” for individual channels by calculating group-specificity scores. These scores determine which channels are consistently useful for a group and which ones create representational interference. High Parameter Efficiency: A significant technical contribution is that GIM achieves group-conditioned adaptation without altering the backbone architecture or introducing any additional trainable parameters. The intervention operates entirely through fixed binary masks applied to activations and gradients. Empirical Validation on Clinical Benchmarks: The authors demonstrate that their approach consistently reduces inter-group performance gaps on the Fitzpatrick-17k and ISIC 2019 datasets. Notably, the method maintains competitive overall accuracy and provides the most substantial benefits to groups that originally exhibited the lowest baseline 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 major strengths of the paper “Group-Conditioned Representation Modulation for Fair Skin Disease Diagnosis” include its innovative framing of fairness, its parameter-efficient technical approach, and its strong empirical results that prioritize clinical accuracy. 1.Novel Problem Formulation: Fairness as Multi-Objective Optimization A core strength is the authors’ departure from traditional fairness methods that aim to suppress sensitive attributes. They argue that in dermatology, attributes like skin type are clinically relevant because they encode vital diagnostic signals, such as UV susceptibility. Instead of “attribute blindness,” they frame fairness as a multi-objective optimization problem where each demographic group is a separate performance objective within a shared backbone. This allows the model to capture group-specific characteristics rather than forcing a “compromise representation” that may degrade performance for all. 2.Original Technical Framework: Group-specific Interference Masking (GIM) The GIM framework introduces a highly original way to manage model capacity without altering the underlying architecture. Channel Utility Scoring: The authors developed a method to identify which convolutional channels provide stable, group-aligned behavior. By calculating group-specificity scores based on prototypes, they can quantify a channel’s utility for a specific demographic. Conditional Modulation: Rather than removing channels globally, GIM uses these scores to create binary masks that selectively suppress low-utility channels for a given group during fine-tuning. This enables group-conditioned expressiveness while allowing groups to still share high-utility parameters. 3.Exceptional Parameter Efficiency A significant practical strength is that the method requires no additional trainable parameters and does not change the backbone architecture. The entire intervention operates through fixed binary masks applied to intermediate activations and gradients. This makes it a lightweight and “practical direction” for improving fairness in existing clinical models with fixed capacity. 4.Strong Demonstration of Clinical Feasibility The paper directly addresses “the cost of fairness”—the tendency for fairness interventions to reduce overall diagnostic accuracy. Maintaining Accuracy: Experimental results on the Fitzpatrick-17k and ISIC 2019 datasets show that GIM consistently reduces inter-group performance gaps while maintaining or even improving competitive overall accuracy. Targeted Improvement: The method is particularly effective for groups that originally showed lower baseline accuracy, suggesting it successfully helps those groups better express their statistical structures. 5.Rigorous Evaluation and Insightful Ablations The work is supported by a comprehensive evaluation against numerous state-of-the-art baselines (e.g., FairPrune, SCP, HSIC) and uses robust metrics like Equal Opportunity (EOpp), Equalized Odds (EOdds), and the FATE score to measure the accuracy-fairness trade-off. Furthermore, their ablation studies provide deep insight into the mechanism: Mask Ratio: They demonstrated a clear trade-off where a small degree of masking (e.g., 2%) optimizes both accuracy and fairness. Swap Masking: By showing that applying one group’s mask to another leads to performance degradation, they provide empirical evidence that the learned masks truly encode group-dependent information rather than arbitrary channel selections.
- 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.
While the paper presents a novel and efficient approach to fairness in skin disease diagnosis, several weaknesses or limitations can be identified from the sources: Sensitivity to the Mask Ratio Hyperparameter: The effectiveness of the GIM framework is highly dependent on the selection of the mask ratio (r). The authors acknowledge a clear trade-off where a small amount of masking improves fairness and accuracy, but “excessive masking can restrict shared representational capacity and reduce overall accuracy”. This requires careful manual selection using validation data for each specific task. Primary Focus on Binary Groupings: Although the authors state that their formulation is extensible to multi-group settings, the methodology and all reported experiments focus exclusively on binary cases (e.g., light vs. dark skin, male vs. female). The performance and complexity of the method in more granular, multi-group demographic scenarios are not demonstrated. Heuristic Selection of Target Layers: The method applies modulation at specific points, such as the “final convolutional layer” for VGG-11 or the “layer4 block” for ResNet-18.The paper does not provide a generalized rule for selecting these layers across different or more modern architectures, suggesting the selection may be heuristic or require trial and error. Dependency on the Initial Backbone: The channel utility scores are calculated based on a frozen pretrained model. This implies that if the initial model is severely biased or has failed to learn relevant features for a specific group during pretraining, the resulting prototypes and masks may be suboptimal. Limited Diversity in Evaluation: The empirical validation is restricted to two datasets (Fitzpatrick-17k and ISIC 2019) and uses relatively older backbones like VGG-11 and ResNet-18.It is unclear how the method translates to more complex architectures (e.g., Transformers) or a wider variety of medical imaging tasks. Fixed Model Capacity Constraints: The framework is intentionally designed for fixed model capacity. While this is a strength for efficiency, it could be a limitation in cases where a model lacks the fundamental representational power to handle diverse demographic groups, where adding parameters might be necessary for a truly fair solution.
- 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
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?
Key Strengths Novel Conceptual Framing Original and Efficient Methodology: The Group-specific Interference Masking (GIM) high parameter efficiency Superior Empirical Performance High Reproducibility Overall, the paper provides a compelling and innovative approach to fairness that is particularly well-suited for the unique constraints of dermatological diagnosis. It is a high-quality contribution that should be accepted.
- 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 authors propose a solution for the problem of group-biased model performance in skin disease diagnosis. As they describe, previous solutions to this problem improve fairness not by improving disadvantaged groups but by degrading the performance of favored groups. Their method is based on a previous work that structured channel modulation can meaningfully affect group-wise performance. According to the authors, channel-level modulation preserves computational efficiency while avoiding excessive parameter fragmentation in comparison with weight-level masking, and offers finer resolution in representation adjustment than layer-level control. This paper’s main contributions are 1) the proposed group-specific interference masking framework; 2) the code that is that that will be made available; 3) the experiments in public datasets. They propose a method for group-conditioned parameter utilization. This is done by per-group score calculation for each channel of a certain set of network features, and based on these scores, some channels are suppressed and disregarded at the final classification for a given group. The core of their approach is the score calculation. This is done by extracting the average or the normalized channels of a given set of output features and utilizing a log-softmax function to calculate an average score for each channel and for each group in the dataset. Based on this score and a threshold, binary masks are utilized to suppress specific feature channels for a specific group. The authors demonstrate their method using precision, recall, and macro-averaged F1 score, and fairness evaluation metrics such as EOpp, EOdds, and FATE. The experiments are conducted on publicly available datasets (ISIC2019, and Fitzpatrick-17k) and follow standard procedures from previous works, such as reference 13.They compare their model with several baseline methods and achieve competitive results, with the best average precision, recall, and F1-score, and best fairness scores in the Fitzpatrick-17k dataset, and best average precision and F1-score, and competitive fairness score results in the ISIC2019 dataset.
- 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 authors propose an efficient and effective way to utilize group-specific features for fair skin diagnosis. This paper is well written, and the results are well presented; the methods are clearly explained and are evaluated with traditional public datasets. The authors indicate that they will make their code available. The method is tested against several fair diagnosis baselines.
- 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 authors could describe more clearly what their contribution is relative to reference 13.2) Reference 13 examines the impact of applying the channel modulation in several layers of the backbone network. In this work, the authors opt to apply the masking of features in the last layer, and not much thought is put into why this choice was made. The authors could include additional explanation on why they chose to apply their method in the last layer and how implementing it in other layers’ features could modify the final 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 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 authors explain the methodology in detail and present results in a satisfactory manner. They claim that it will provide a link to their code. The proposed method contributes to efficiently increasing fairness in skin diagnosis while maintaining performance.
- 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
N/A
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.
- The proposed group-specific interference masking offers a lightweight and parameter-efficient way to improve group-wise fairness while preserving overall diagnostic performance.
- The empirical evaluation is generally strong. The results suggest a favorable accuracy-fairness trade-off, especially for under-performing demographic groups.
- The authors should better clarify the novelty relative to prior channel modulation/pruning methods, evaluate the method on more modern backbones, and discuss extension beyond binary demographic groups.
