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

Abstract

Accurate dermatological diagnosis naturally necessitates equitable performance across diverse populations, yet a systematic lack of expertly annotated images, especially for underrepresented skin tones and rare diseases, impedes progress toward measurably fair methods. We introduce cgDDI (Controllable Generation of Diverse Dermatological Imagery), a hybrid framework that (1) synthesizes realistic healthy skin samples without disturbing other input properties, (2) maps single-sample rare lesions onto novel skin-tones and locations non-parametrically, and (3) allows for efficient parametric generation with as few as 10 training samples. The framework supports both human and automated segmentation masking, enabling scalability to datasets without pre-made lesion masks. We grow a 656-image dataset by more than 400× and validate across two datasets: biopsy-confirmed Diverse Dermatology Images (DDI) and expert-verified Fitzpatrick17k (F17k). On the DDI benchmark, we achieve malignancy classification accuracy of 86.4% under synthetic-only training and 90.9% state-of-the-art performance with real data fine-tuning, alongside leading fairness metrics. Cross-dataset experiments show +13.9% accuracy improvements on unseen F17k data despite minimal disease overlap. We openly release 266k+ synthetic images, code, and generative models to further support fairness research at https://github.com/hectorcarrion/ControllableGenDDI

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1473_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/hectorcarrion/ControllableGenDDI

Link to the Dataset(s)

https://huggingface.co/datasets/hcarrion/ControllableGenDDI

BibTex

@InProceedings{CarHéc_Controllable_MICCAI2026,
        author = { Carrión, Héctor AND Norouzi, Narges},
        title = { { Controllable Generation of Diverse Dermatological Imagery for Fair and Efficient Malignancy Classification } },
        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 authors propose a Controllable Generation of Diverse Dermatological Imagery (cgDDI), a framework that synthesizes skin lesion images for different conditions and skin tones. cgDDI is a multi-step pipeline that includes: 1) lesion removal from input images (latent diffusion inpainting) to create a healthy skin that will work as a canvas and a control signal for the next steps of lesion synthesis, 2) parametric image generation: a latent diffusion model is finetuned to generate lesion images based on input text created by a finetuned image to text model and the canvas skin image. For underrepresented skin conditions, the framework generates images by layering the segmented lesion of interest over the canvas skin (non-parametric generation). The generated images are validated for data augmentation using a malignancy classification task. The results show that models trained on cgDDI-generated data outperformed other models.

  • 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.The work provides an efficient way to generate realistic synthetic images, specifically, the finetuned components were trained on a small dataset. 2.The authors comprehensively demonstrated the benefits of utilizing the generated images through a classification task using different experiments.

  • 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.Limited novelty: the work is similar to the work presented in: Sagers, L. W. , Diao, J. A. , Melas-Kyriazi, L. , Groh, M. , Rajpurkar, P. , Adamson, A. S. , Rotemberg, V. , Daneshjou, R. , Manrai, A. K. : Augmenting medical image classifiers with synthetic data from latent diffusion models. arXiv (Cornell University) (2023). It would be great if the authors could highlight how cgDDI differs from Sager’s work. 2.Several components in the cgDDI depend on segmentation masks which are difficult to obtain. Using general-purpose segmentation models like SAM is not always suitable for medical images. The authors are encouraged to show some examples where SAM did not perform well to see how image generation is impacted. 3.Implementation details, data splits, and finetuning parameters are not listed (some details are in the GitHub repo, but the reader will not know until they visit the repo). 4.The non-parametric branch of the cgDDI is a relatively basic solution to the underrepresented skin conditions.

  • 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

    To the authors: Out of curiosity, I wanted to know why you think the model needs around 10 samples per skin condition to perform well. The same skin condition can vary significantly, how does this variability affect this number?

  • 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 focuses on improving data diversity, which is an important and highly needed topic. With some clarifications, I think the paper will be a nice addition to the field of equitable diagnosis.

  • 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

    This paper presents cgDDI, a hybrid framework for controllable dermatology image synthesis that combines healthy-skin inpainting, non-parametric lesion mapping, and parametric semantic generation via textual inversion + LoRA, with the goal of improving malignancy classification fairness across skin tones.

  • 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.The paper addresses a genuinely important and underexplored problem: improving dermatology AI under skin-tone imbalance and data scarcity, especially for rare conditions. 2.The proposed framework is well motivated and combines complementary generation strategies for different scarcity regimes. The compatibility with automated masks is also a useful practical feature. 3.Releasing 266k+ synthetic images, code, and models could be useful for the community.

  • 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 DDI dataset contains only 656 images, meaning each fold in the five-fold cross-validation yields roughly 130 test samples. Results stratified by disease rarity (Table 3) are unreliable as the sample size is extremely small. For example, the “very rare” category covers only 6 test samples. 2.The paper does not report subgroup-level sample counts, which should be a requirement when fairness across subgroups is one of the central claim. 3.The paper does not compare against standard data augmentation (e.g., geometric transforms, color jitter, MixUp) applied to the same real training data. It cannot determine whether the performance gains are attributable to the proposed generative framework specifically, or simply to having more training samples regardless of source. 4.Table 4 contains several settings where synthetic data hurts performance relative to the real-data baseline, for example DDI→F17k. The paper should discuss why synthetic data can be harmful in specific settings.

  • 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 usefulness of the application and the synthetic images were released.

  • Reviewer confidence

    Somewhat confident (2)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #3

  • Please describe the contribution of the paper

    This paper aims to address the issue that dermatological AI systems trained on skin-tone-imbalanced datasets exhibit degraded performance on darker-skinned populations. Existing solutions, such as large-scale data collection or reliance on proprietary generative models, do not scale well to underrepresented skin tones or rare diseases with extremely limited samples. The authors propose cgDDI, a hybrid synthetic data generation framework combining three components: (i) latent diffusion inpainting to generate healthy skin images by removing lesions while preserving anatomical and acquisition characteristics; (ii) non-parametric lesion mapping that transplants real lesions onto synthetic healthy canvases, enabling augmentation even for single-sample diseases; and (iii) parametric semantic generation using textual inversion and LoRA fine-tuning, regularized via a prior-preservation loss constructed from the generated healthy images. Starting from 656 DDI images, the method produces over 266k synthetic images balanced across Fitzpatrick skin tones. Experiments show improved classification performance and fairness metrics on DDI, as well as gains in cross-dataset transfer to F17k. The authors release the full dataset, code, and models.

  • 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 directly targets fairness disparities in dermatology AI, a well-established issue with real-world implications. The motivation is clear and aligns well with MICCAI priorities.
    • A key conceptual contribution is leveraging generated healthy skin images, matched in acquisition conditions and skin-tone distribution, as the prior-preservation set for LoRA fine-tuning. This is a non-obvious and practically useful design choice that enables effective few-shot generation.
    • The lesion transplantation strategy is simple yet effective, particularly for extremely low-data regimes where parametric models are insufficient. This component is well-motivated and empirically useful.
    • The public release of 266k images balanced across skin tones is a significant contribution, especially given that comparable prior efforts rely on private datasets.
    • The paper follows established evaluation protocols, reports multiple fairness metrics, and includes cross-dataset and ablation experiments. The breadth is appropriate for the venue.
    • The authors acknowledge and attempt to mitigate leakage from conditioning on test images, which is an important and often overlooked issue.
  • 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 reported improvement in EOM (≈+17 points) is substantial, especially alongside gains in overall accuracy. The paper does not provide a detailed breakdown (e.g., per skin tone or disease category) to explain the mechanism behind this improvement. As a result, it is difficult to assess whether the gains reflect true fairness improvements or dataset-specific effects.
    • The F17k evaluation is based on a small test subset (~43 samples), making reported gains sensitive to small variations. Confidence intervals or statistical significance analyses are not provided for these results.
    • The interpretation of cross-dataset improvements as evidence of “generalizable lesion features” is not fully substantiated. Alternative explanations, such as improved coverage of skin-tone or acquisition variability, are not ruled out.
    • No dermatologist review or expert assessment of generated images is provided. Given the scale of the released dataset and its intended downstream use, this is a notable omission.
  • 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 paper addresses an important clinical fairness problem and proposes a practical, well-engineered solution. The combination of synthetic data generation techniques is effective, and the public release of a large, balanced dataset is a meaningful contribution to the community. The primary concerns are empirical rather than conceptual: several key claims (particularly around fairness and generalization) are insufficiently analyzed, and parts of the evaluation rely on small sample sizes. Additionally, the absence of clinical validation for generated data limits confidence in its downstream utility. Overall, the work is solid and useful, but would benefit from stronger empirical validation and clearer interpretation of results.

  • 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-Reviewer for their decision to early accept our work and their constructive feedback. We are encouraged that the reviewers describe the work positively, addressing an “important and underexplored problem” (R3) with “complementary generation strategies” (R2), and “comprehensively demonstrated the benefits” of our method (R1) via a “well-motivated” solution (R3). We now respond to concerns:

A. Novelty: cgDDI’s novelty rests primarily on: (1) the first use of healthy-image prior-preservation loss (identified by R3 as “a non-obvious and practically useful design choice that enables effective few-shot generation”), which mitigates the semantic drift inherent to LoRA fine-tuning of latent diffusion; (2) non-parametric lesion-mapping augments ultra-rare single-sample DDI diseases, where other parametric methods fail; (3) full Fitzpatrick I-VI model coverage with rich fairness and generative-quality metrics; (4) generation and release of pixel-perfect in-clinical-distribution healthy samples, which serve multiple uses downstream. These 4 aspects are unique to cgDDI. We will incorporate minor edits to further highlight novelty.

B. Sample size & statistical reliability: We adopt the PatchAlign DDI seeds & 5-fold protocol; reporting standard deviations in Tables 2 and 4 across these folds. The Exp. 2 Equality of Opportunity Metric (EOM) gain is +17 points (std 1.9), are above fold variance. The “very rare” stratum and the small F17k test subset are inherent challenges of expert-verified real-world dermatology data distribution (Sec. 2). Per-tone and per-disease subgroup counts are present in our shared repository. We will add an explicit subgroup-count table to the camera-ready paper.

C. Mechanism of the fairness improvement: The EOM measures consistency of true-positive rate (TPR) across skin tones. Table 2 contains the per-tone breakdown: in Exp. 2 the medium-skin-tone accuracy rises from 74.6% to 86.4% (+11.8 points) while light- and dark-tone accuracies are largely maintained, closing the medium-tone deficit. This matches cgDDI’s explicit oversampling of medium tones (Sec. 3.4), targeting their relative under-representation in malignant DDI cases. We will clarify this in camera-ready.

D. Comparison to standard augmentation: PatchAlign baseline’s training pipeline already applies standard augmentations (random crops, rotations, color jitter & flips); cgDDI’s gains are on top of these. Standard augmentation structurally cannot operate along sensitive attributes. Thus our data generation strategies are orthogonal and complementary to conventional augmentation. We will highlight this in camera-ready.

E. Automated mask quality: We agree SAMv3 masks may be imperfect and are aware of specialized dermatology segmentation models that may improve upon SAMv3; benchmarking such alternatives is outside the scope of this work. Our goal is to show that automatic masking is compatible with cgDDI such that human labelling is not a hard requirement. We will upload additional SAMv3 failure examples as a multimedia appendix or to the code repository.

F. Cases where synthetics hurt + cross-dataset interpretation: F17k and DDI share only one disease (CTCL); synthetics underperform only on direct cross-domain transfer where the source lacks the target’s diseases. Mix→target settings (synthetics spanning both) recover and surpass baselines, indicating the failure mode is class coverage, not generative quality. We agree with R3 that the +13.9% F17k gain is also consistent with improved skin-tone and acquisition-condition coverage; we will note this alongside the lesion-feature interpretation.

We thank the reviewers again for helping us strengthen cgDDI.



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 paper addresses an important problem in dermatology AI, focusing on improving performance and fairness under skin-tone imbalance and limited data. The proposed cgDDI framework is well motivated and combines multiple generation strategies in a practical way. The release of synthetic data and code is also valuable, and the experimental results show improvements in classification performance across settings.

    There are some limitations. The novelty over prior generative approaches could be clearer, and parts of the evaluation rely on small sample sizes, making some conclusions less certain.

    Overall, the strengths outweigh these concerns, and the paper meets the bar for acceptance.



back to top