List of Papers Browse by Subject Areas Author List
Abstract
Black box deep learning models for dermatology show performance
disparities across skin types, with darker-skinned patients experiencing lower diagnostic performance. Concept Bottleneck Models (CBMs) offer interpretability through intermediate concept predictions but lack explicit fairness mechanisms. We introduce Fair Curriculum CBM, a four-phase curriculum learning approach for in-training bias mitigation that, instead of relying on concept difficulty, orders training by fairness objectives, including balanced foundation, demographic parity, equalized odds, and performance parity. This progressive structure trains all concepts jointly while gradually introducing fairness constraints and adversarial debiasing directly, rather than treating fairness as a post-hoc adjustment. On SkinCon (3230 images, 6 Fitzpatrick types), Fair Curriculum CBM improves lowest-group F1 by 63% (0.270→0.441, p<0.001) and reduces performance gaps by 44% (0.361→0.203, p=0.003) compared
to difficulty-based curriculum learning, with simultaneous 5.3% overall F1 gain (p<0.001). Paired t-tests comparing fairness-first to difficulty based curriculum approaches show significant improvements on diverse skin types (Types II, V, VI: p≤0.005) without degradation on lighter types. Results from 100 independent runs per model validate that fairness curriculum learning improves rather than constrains performance. Code available at: https://github.com/Matt-Cockayne/FairCBM.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3379_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Matt-Cockayne/FairCBM
Link to the Dataset(s)
SkinCAP dataset: https://huggingface.co/datasets/joshuachou/SkinCAP
BibTex
@InProceedings{CocMat_Fair_MICCAI2026,
author = { Cockayne, Matthew J. AND Ortolani, Marco AND Al-Bander, Baidaa},
title = { { Fair Curriculum Learning for Concept Bottleneck Models in Dermatology } },
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
To me, the main contribution of this work is the introduction of a fairness-driven curriculum learning framework for Concept Bottleneck Models (CBMs), where fairness objectives are progressively integrated during training rather than applied statically or post hoc. In particular, the authors proposed a four-phase curriculum that transitions from balanced sampling to demographic parity, equalized odds with adversarial de-biasing, and finally direct performance gap minimization, while jointly learning interpretable concept representations. This formulation is motivated by the instability of applying fairness constraints and adversarial learning simultaneously, and instead leverages a structured training schedule to stabilize optimization and improve group-level performance. This work demonstrates that the approach can significantly improve lowest-group performance and reduce disparities across Fitzpatrick skin types in dermatology setting, while maintaining or slightly improving overall predictive 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.
I think this research study tackles an indeed clinically important and well-motivated problem, namely performance disparities across skin types in dermatology, which is a well-documented limitation of current AI systems. The focus on Fitzpatrick skin types and lowest-group performance is appropriate, and the use of a dataset with dense concept annotations (SkinCon) allows for more granular analysis than standard classification benchmarks. I also think it is a strength that the work combines fairness and interpretability together, rather than treating them separately, which is particularly relevant in clinical settings where understanding model reasoning is important for trust and adoption. One of the key technical strengths of this work to me is the reformulation of curriculum learning around fairness objectives rather than task difficulty. Instead of the standard easy-to-hard paradigm, the authors design a structured four-phase curriculum that progressively introduces fairness constraints (balanced sampling → demographic parity → equalized odds with adversarial debiasing → performance parity). I think this is an interesting and reasonable idea, especially given that adversarial de-biasing and fairness constraints are known to cause instability when applied abruptly. The notion of using curriculum learning as a stabilization mechanism for fairness optimization is a meaningful contribution too. Another strong aspect of this work is the comprehensive and carefully controlled experimental evaluation. This work also compares multiple baselines, including direct models, standard CBMs, difficulty-based curriculum CBMs, and fairness-aware variants, which helps isolate the effect of the proposed curriculum design. The use of 100 independent runs with paired statistical testing is particularly notable, as it provides stronger evidence than single-run evaluations and helps address variability in training. In addition, the inclusion of multiple metrics (overall F1, per-group F1, lowest-group performance, and gap) gives a more complete picture of both performance and fairness. To me, and from a clinical perspective, the fact that improvements are achieved without degradation on lighter skin types is also an important observation, as it addresses a common concern about fairness-performance trade-offs.
- 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 components of the method, including demographic parity, equalized odds, adversarial de-biasing with gradient reversal, and error-driven re-weighting, are all already well-established in the fairness literature. Curriculum learning has been explored extensively for stabilizing training, including in the context of CBMs. The main contribution here is the ordering of these existing components into a four-phase curriculum, but it is not entirely clear that this ordering introduces fundamentally new capability beyond a kind of tuned training schedule. I think this work would benefit from a clearer discussion of how this differs in principle from existing in-processing methods such as adversarial fairness, GroupDRO-style re-weighting, o/and prior curriculum-based training strategies. The second concern to me is the scope and generalizability of the experimental validation, which is limited to a single/retrospective dataset (SkinCon) with relatively small sample size (3,230 images) and known class imbalance across Fitzpatrick types. While the use of 100 runs is a strength, the conclusions are still drawn from a single domain and dataset, and it is unclear whether the observed improvements would hold in other dermatology datasets or broader medical imaging settings. The authors also acknowledge missing comparisons with established fairness baselines such as GroupDRO, LAFTR, or CFair, which are standard in the fairness literature. Without these comparisons, it will be challenging to assess whether the proposed curriculum provides advantages over existing in-processing methods beyond the specific baselines included. Another limitation is that the fairness evaluation is largely metric-driven and somewhat narrow in scope. I mean, this work focuses on F1, lowest-group F1, and performance gap, but does not provide deeper analysis of model behavior at the concept level, despite using a CBM architecture. For example, it would be helpful to understand whether fairness improvements arise from more accurate concept predictions across groups, or from downstream classifier adjustments. In addition, while statistical significance is reported, the practical magnitude of some improvements (e. g. , overall F1 gains) is relatively modest, and the large improvements in lowest-group performance may partly reflect instability in baseline models rather than consistent gains. A more detailed analysis of variability and failure cases would strengthen the claims. Moreover, this work somehow overstates the clinical implications of the results given the current experimental setup. The evaluation is retrospective only, based on a single curated dataset with predefined concept annotations, and does not include external validation, clinician assessment, or deployment considerations. Moreover, the method assumes access to high-quality concept labels and demographic annotations, which may not be available or reliable in real-world clinical settings. While the improvements in fairness metrics are encouraging, it is not yet clear how these translate to clinical decision-making or patient outcomes. A more justified discussion of these limitations would improve the overall quality of the work.
- 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?
This work is generally well-structured, with a clear progression from motivation to methodology and experiments, and the four-phase curriculum is presented in a logical manner. The inclusion of a diagram (at page 4) also helps clarify how the curriculum interacts with the CBM architecture. However, some parts of the methodology, particularly the multiple loss terms and phase transitions, are relatively dense and require careful reading to follow. Also, the exposition could benefit from more intuitive explanations of why each phase is necessary and how they interact. I think this work is slightly above the acceptance threshold. It addresses a clinically relevant and important problem, and the idea of structuring fairness optimization through a curriculum is reasonable and practically useful. The experimental evaluation is relatively well, particularly with 100-run statistical validation and consistent improvements in lowest-group performance without degrading majority groups. The methodological novelty is somehow incremental, as it primarily combines known fairness objectives (DP, EO, adversarial de-biasing) within a scheduled training framework. The evaluation is also limited to a single dataset, and comparisons to stronger fairness baselines are missing. With all that, I would be comfortable accepting (weak accept) this work as a useful contribution, but I would not oppose rejection/weak rejection if concerns about novelty or generalizability are raised by other reviewer(s).
- 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 work targets a critical gap in AI-driven dermatology by addressing fairness in diagnostic performance across different skin tones. It introduces a novel curriculum-based training strategy that explicitly optimizes for both predictive performance and demographic parity. The results show improved performance, particularly for underrepresented skin tone groups, demonstrating the effectiveness of the proposed approach in mitigating bias.
- 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 primary strength of the paper lies in its introduction of a novel fairness protocol for AI models in dermatology. The empirical evaluation is comprehensive, incorporating rigorous statistical testing that provides strong support for the effectiveness of the proposed curriculum learning approach. Additionally, the use of a large, diverse, and representative dataset—SkinCon—further strengthens the validity of the findings.
- 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 experiments demonstrate the advantages of the proposed curriculum learning approach, a clearer comparison with the fairness performance (with respect to skin tone) of other state-of-the-art models on the same classification task would strengthen the evaluation. Such a comparison would provide better context for assessing the relative usefulness and impact of the proposed method.
- 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?
The work is both timely and relevant, addressing an important gap between the development of AI models and their reliable deployment in clinical settings.
- 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 novel four-phase curriculum learning framework for training Concept Bottleneck Models (CBMs) by integrating fairness constraints. Unlike conventional curriculum learning approaches that focus on sample complexity, this method is primarily driven by fairness objectives and incorporates adversarial debiasing to mitigate bias throughout the training process.
- 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.
- Novelty: While the individual methods employed are already established in the literature, their integration into the proposed framework is highly well-structured and soundly motivated by previous research. Furthermore, the effectiveness of this specific combination is convincingly validated through comprehensive ablation studies.
- Extendability: Although the paper primarily evaluates the framework on a dermatology task using a vision-based Swin-Tiny model, the proposed approach demonstrates significant versatility. Even though it is not explicitly stated by the authors, the framework could be straightforwardly adapted to different architectures and various domains, including tasks involving tabular data.
- Robust validation: The framework is effectively validated through ablation studies, reflecting a rigorous methodological workflow. The robustness of the results is ensured by averaging performance across multiple runs for each approach and applying paired t-tests. These results clearly demonstrate the synergy of the integrated techniques, leading to a significant performance gap between the proposed method and alternative approaches.
- 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.
- Single-Dataset Evaluation: The framework is validated on a single dataset, which somewhat limits the generalizability of the effectiveness of the proposed approach. However, it is appreciated that the authors acknowledge this limitation in Section 6.
- 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
-
Missing Definitions and Clarifications: Section 3.2 lacks explicit definitions for certain components of the loss functions, specifically L_concept and L_binary (although the latter is presumed to be Binary Cross-Entropy). Additionally, the authors should explicitly define the λ parameters in the text (which Figure 1 suggests are weighting coefficients) and the variable T used in Equation 1.Finally, it is highly recommended to briefly describe the “GAP” metric presented in the tables to ensure clarity for all readers.
-
Unreferenced Table and Column Details: Table 2 is never referenced in the main text. Furthermore, it is not clear what specific data or metrics are reported under the columns “Difficulty Curriculum” and “Fairness Curriculum.” The authors should ensure the table is properly cited and its contents are explained.
-
Dataset Preprocessing and Split Details: Section 4 lacks necessary information regarding the image preprocessing steps applied to the dataset. Additionally, the specific role of the validation set is unclear (e.g., was it used for hyperparameter tuning or early stopping?). Currently, the text gives the impression that all results are directly derived from the test set, which should be clarified to confirm the methodological soundness of the training process.
-
Visual Presentation of Figure 1: Figure 1 is highly appreciated as it effectively illustrates the entire workflow. However, the use of thick dashed lines makes the diagram visually cluttered. I suggest replacing them with thinner lines or using colored background panels to separate the areas, which would enhance overall readability.
-
- 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?
The proposed methodology is solid, and the experimental evaluation demonstrates positive results. The authors successfully integrate multiple techniques to improve both the fairness and the interpretability of the model. Notably, they achieve this without sacrificing the overall predictive performance. While there are some minor issues regarding clarity and missing definitions in the text, these do not undermine the core contribution and can be easily addressed by the authors in the camera-ready version.
- 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
Thank you to all reviewers for the constructive and encouraging feedback. We are grateful that the clinical motivation, fairness methodology, and statistical validation were positively received. Below we clarify the main concerns and how these will be addressed in the final version.
Contribution scope and novelty: We agree that DP/EO/adversarial debiasing are established individually. Our contribution is not a new standalone fairness loss, but a structured fairness curriculum for CBMs that schedules when and how these components are introduced (sampling strategy, fairness objectives, and adversarial activation) to avoid instability from abrupt joint optimisation. The submitted ablations support this design choice, as removing phases or adversarial warmup substantially degrades fairness/performance, while static fairness without curriculum underperforms. This will be further clarified In the limitations of the manuscript. Comparisons to additional fairness baselines and dataset scope: We accept and acknowledge this limitation in the paper. Our study targets the intersection of fairness and interpretability in CBMs, which requires datasets with both dense concept and protected-attribute labels. Availability of this joint supervision data constrains comparable benchmark settings in dermatology. Our claim is calibrated to this data availability. We demonstrate consistent gains over direct, standard CBM, difficulty-based curriculum CBM, and static-fairness CBM under matched runs, rather than claiming superiority over all fairness methods. Generalisability and clinical interpretation: We agree this is a single-dataset study and should be interpreted accordingly. Therefore, the manuscript does not claim deployment readiness or patient-outcome impact. It reports methodological evidence that a fairness-optimised curriculum design improves lowest-group performance and reduces group performance gaps in this controlled setting. External validation and broader clinical assessment remain necessary future work that will be outlined in the manuscript. Clarity requests: Thank you for the detailed suggestions to improve clarity. We will make wording clarifications to improve readability. We will also explicitly define the loss components and symbols, define the ‘gap’ metric in text, clarify validation-set use for model selection/checkpointing, and ensure table references/column meanings are explicit.
We appreciate the reviewers’ feedback and believe it has helped us improve the clarity and positioning of our contribution. The scope of this research will be expanded in future work, which will approach cross-domain generalisation and clinical validation.
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 agree to accept this paper, especially given the novelty of the methodology proposed. I also encourage the authors to revise their paper based on the thoughtful comments from the reviewers,