List of Papers Browse by Subject Areas Author List
Abstract
Machine-learning models applied to skin images often have degraded performance when the skin colour captured in images (SCCI) differs between training and deployment. These discrepancies arise from a combination of entangled environmental factors (e.g., illumination, camera settings) and intrinsic factors (e.g., skin tone) that cannot be accurately described by a single “skin tone” scalar – a simplification commonly adopted by prior work. To mitigate such colour mismatches, we propose a skin-colour disentangling framework that adapts disentanglement-by-compression to learn a structured, manipulable latent space for SCCI from unlabelled dermatology images. To prevent information leakage
that hinders proper learning of dark colour features, we introduce a randomized, mostly monotonic decolourization mapping. To suppress unintended colour shifts of localized patterns (e.g., ink marks, scars) during colour manipulation, we further propose a geometry-aligned post-processing step. Together, these components enable faithful counterfactual editing and answering an essential question: “What would this skin condition look like under a different SCCI?”, as well as direct colour transfer between images and controlled traversal along physically meaningful directions (e.g., blood perfusion, camera white balance), enabling educational visualization of skin conditions under varying SCCI. We demonstrate that dataset-level augmentation and colour normalization based on our framework achieve competitive lesion classification performance. Ultimately, our work promotes equitable diagnosis through creating diverse training datasets that include different skin tones and image-capturing conditions.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3075_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026/supp/3075_supp.zip
Link to the Code Repository
https://github.com/yangwenbo99/disentangle-skin-colour
Link to the Dataset(s)
N/A
BibTex
@InProceedings{YanWen_Automated_MICCAI2026,
author = { Yang, Wenbo AND Rezk, Eman AND Moursi, Walaa M. AND Wang, Zhou},
title = { { Automated Disentangling Analysis of Skin Colour for Lesion Images } },
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 paper proposes a method for generating modified versions of an existing skin lesion by altering its color features (e. g. , luminance, skin tone, camera settings).
- 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 is well written.
- Illumination and color differences can heavily affect the performance of skin lesion analysis models, and finding ways to deal with such variation can be impactful.
- 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 main weakness is the paper validation, which is notably hard when generating counterfactuals, as there are no ground truths. But because validation is not straightforward, it also requires more effort to provide evidence that the proposed method works. For example, the paper’s first results are qualitative (subsection 3.1). To support the authors’ claim that “The figure confirms that the model produces diverse yet realistic skin colour variations,” we would ideally need an assessment from a dermatologist, as it is hard to say that just from one image, and without really knowing the potential variations an image can exhibit in the wild.
Moreover, as skin lesions exhibit different visual characteristics across skin tones [1,2], simply modifying healthy skin color might produce unrealistic skin lesions and provide incorrect priors (regarding the frequency of dermoscopic attributes or their appearance) when training dermatologists or machine learning models. I believe this is very important as it can perpetuate biases in skin lesion diagnosis models. I would suggest that the authors explicitly limit the scope of their solution to environmental factors (e. g. , illumination, camera settings) while preserving skin tone intact.
Finally, it would help to have more comparisons to similar methods. The only comparison happens with citation [16], but only in one of the experiments performed (Table 1).
[1] Gloster Jr, Hugh M. , and Kenneth Neal. “Skin cancer in skin of color. “ Journal of the American Academy of Dermatology 55.5 (2006): 741-760. [2] Karampinis, Emmanouil, et al. “Race-specific and skin of color dermatoscopic characteristics of skin cancer: a literature review. “ Dermatology Practical & Conceptual 13.4 Suppl 1 (2023): e2023311S.
- 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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
I believe the paper should be rejected due to the lack of validation and comparison with other methods. Also, there are potential problems of using the proposed method to simulate skin lesions in underserved populations without proper evaluation, which could perpetuate biases.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
Using a diagnostic model as a proxy for evaluating another diagnostic model has important limitations, as models do not necessarily rely on clinical features to make predictions. To partially mitigate this, subgroup analysis could help increase trust that the proposed method is not detrimental to underrepresented skin tones, and clarify which groups benefit most from the procedure. Such analysis would also provide evidence on whether the potential generation of unrealistic or biologically implausible lesions for specific demographics affects downstream performance. Given these methodological concerns, along with the risk of uncontrolled color manipulation in skin lesion analysis, I recommend rejection.
Review #2
- Please describe the contribution of the paper
This paper proposes a novel method that consists of 1.a model that performs monotonic decolorization of the image, 2.Post-processing for color correction, 3.A data augmentation and normalization method that considers unlabelled testing data for adaptation purposes.
- 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 is well structured and easy to understand 2.The good ablation study of latent space exploration for blood perfusion, color temperature and camera setting. 3.Details for the
- 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 complete flowchart workflow is not provided. 2.There are a large number of skin datasets available only two datasets are used for the method evaluation. 3.Only a single baseline is used; there are Style-GAN methods or other generative models that can be compared with your method.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
1.Provide a complete end-to-end flow chart, diagram, or model workflow of the process that captures the complete flow of the tasks
- 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
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
Authors does not provide appropriate justification.
Review #3
- Please describe the contribution of the paper
The authors propose an unsupervisedly learned skin-colour disentangling framework that adapts disentanglement-by-compression to learn a structured, manipulable latent space for skin colour captures in images (SCCI) from unlabelled dermatology images.
The proposed framework is a modification of the disentanglement-by-compression framework presented in [25], by compressing different SCCI characteristics into an organised latent space of independent dimensions.
The authors modify the framework in [25], which is based the principle of information bottleneck to low-level vision feature disentanglement, by (1) Introducing a randomized decolourization mapping function (instead of a naïve linear grayscale conversion) to prevent skin tone leakage to the framework’s colour synthesis network that impedes the model from achieving a robust and generalizable representation learning. (2) Introducing a geometry-aligned post-processing step to selectively reject unreliable changes to the reconstructed RGB skin lesion images that encourage unintended colour shifts due to localised patterns (such as ink marks and scars) in images. This is intended to also contribute to better representation learning by suppressing rarer artefacts (ink marks, scars) in skin images. This can enable the generation of subsets of skin lesion images, for the purpose of data augmentation, that are within distribution and free of these rare artefacts.
- 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 present three use cases for their framework: (1) As an educational visualization tool that enables physicians in training to utilize the SCCI transfer capability of the framework to visualize what a single skin condition under a range of SCCI. (2) As a data augmentation strategy to improve the performance of a model on the downstream task of skin lesion classification, by sampling from (or modifying) the SCCI latent space of the framework. (3) As a data normalisation strategy for downstream skin lesion classification tasks.
The introduction of the colourless image in the bottleneck of the model, as shown in Fig. 3, is well-configured to constrain the embedding space to learn a lower dimension colour representation of skin lesion images.
The authors presented a well-structured ablation study results in Table 1.The qualitative ablation study images shown in Fig. 4 present an effective and explainable visualisation of the impact of the proposed randomized decolourisation mapping function and the post-processing block.
Fig. 1(a) is a good visualization of the SCCI transfer between source and target images.
Fig. 1(b) is a good visualization of the continuity of the 5 different dimensions of the latent space, though none of the 5 dimensions were assigned to specific SCCI characteristics.
The potential impact of the proposed framework is contribution towards the improvement of health equity by enabling the generation of training set augmentations that mimic the distribution of different populations. The work is a step towards achieving equitable diagnosis, where models for skin lesion classification can be reliably adapted across different clinical settings and populations.
- 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 authors define SCCI as characteristics (such as intrinsic patient skin tone, skin blood perfusion, illumination condition during imaging, camera white balance, and other device-dependent characteristics) of skin images that affect the generalizability of computer vision across populations of dermatology datasets. The authors also state that the latent space in their architecture that encodes independent SCCI characteristics (such as intrinsic patient skin tone, illumination condition, etc) has 256 dimensions, and that each dimension encodes a specific SCCI characteristic. This should mean that the latent space encodes 256 mutually independent SCCI characteristics. However, the authors only show five dimensions from the latent space, which encodes the “Skin Blood Perfusion”, “Camera White Balance”, “Colour Temperature of illumination”, and two other unmentioned SCCI characteristics.
The authors mentioned that they computed variance of each of the 256 dimensions of the latent space and identified only 5 “active” dimensions which are presented in Fig. 1(b). What does “active” mean in this context? Does it mean that the framework encodes only 5 SCCI characteristics out of a potential 256 independent SCCI characteristics? Does it mean that the remaining 251 dimensions are redundancies?
What informed the choice of 256 dimensions for the latent space size? What will be the impact of reducing the size of the latent space to 128 dimensions on the output of the framework? The authors should have performed an ablation study to test this. Such a study could have informed an optimal latent space size.
It seems it is only three of the coordinates in the latent space that mainly encode physically meaningful trajectories (like blood perfusion, camera white balance). No test was performed to quantify what physically meaning trajectories (if any) were encoded by the other coordinates in the latent space. These other coordinates may also just be derivatives or linear combinations of the three main trajectories (blood perfusion, white balance, etc.). This means these other coordinates may be highly correlated with the main coordinates. The authors should have quantified the level inter-correlation between the latent space coordinates, to estimate their level of mutual independence.
The independence of the coordinates in the latent space is assumed and not verified. The authors should have estimated the correlation between the coordinates to quantify their level of mutual independence. The independence of the dimensions in the latent space was not well validated or quantified.
To perform colour transfer at test time using a previously unseen skin lesion RGB image, does the image pass through the randomized decolourization block before being fed to the colour synthesis network? If yes, does the decolourization block produce the same colourless image if the same RBG image is fed to the block several times (i.e. does the decolourization block produce predictable, reproducible colourless image for a given skin RGB lesion image). If, at test time, the decolourization block produces unpredictable colourless image inputs to the synthesis network, wouldn’t that make the reproduced coloured image output unpredictable? The authors need to address the potential unpredictability of the colourless image (x) at test time, which may lead to an unpredictable reproduced colour image. If, at test time, a single RGB skin lesion image is fed to the network 100 times, will the 100 resulting colourless images (x) be identical, will the 100 reproduced coloured images be identical (while keeping the colour embedding space fixed)?
Training the colour synthesis network on the training and test dataset is not ideal. With this configuration, the authors will not be able to evaluate the performance of the model on skin lesion images previously unseen by the network. Is there not some form of data leakage when you feed a single testset image y_i to the colour synthesis model at test time, but the model has seen a decolourized version of this testset image y_i at training time?
The authors could have split the dataset into training, validation, and test sets. The training and validation set should have been used at training time. Then at test time only the test set should be used for evaluation. This would enable the evaluation of the model’s performance on completely unseen data.
However, with the current configuration, the synthesis model has seen a version of all the data (training and test set). This prevents the evaluation of the true generalizability of the colour synthesis network.
What do the authors mean by “mostly” monotonic decolourization? Their monotonic decolourization mapping function applies some randomization to achieve decolourised images from RGB images. Authors should explain why they describe this process as “mostly” monotonic. The term “mostly” suggests partial achievement of the monotonic decolourisation goal. Authors should expatiate on this.
Minor issues: On first mention, the authors define the acronym SCCI as “skin colour captured in images”. However, on the second mention they define it “skin colours captured in images”. Authors should use a consistent acronym definition, “skin colour captured in images” is the recommended definition.
The acronyms FST (Fitzpatrick skin tone scale) and MST (Monk skin tone scale), on page 2, should be written in full.
- 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.
(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 propose a framework that disentangles five physically meaningful characteristics that influence the skin colour captured in images of skin lesions into a low-dimension feature space using the concept of information bottle. The framework is achieved by making two useful modifications, (1) introduction of a randomized decolourization mapping function, and (2) introduction of a post-processing step to reject image artefacts that cause unintended colour shifts, to the recent framework in [25].
The authors present three useful use cases of their framework: (1) educational visualization, (2) data augmentation strategy, and (3) data normalisation strategy.
On the other hand, the paper’s major weakness are centered around (1) The lack of justification for the latent space size, (2) The relatively low (five) number of physically meaningful skin image characteristics encoded in the latent space, (3) The lack of validation of the level of independence of the 256 dimensions of the latent space, (4) The lack of a separate test dataset never seen by the model during training, and (5) The potential unpredictability of the randomized decolourization mapping function if it is used at test time.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The authors partially addressed my concerns about the characteristics of their colour embedding space and my concerns the determinism of decolourization during testing.
Clarification on the use of trajectory building sample photos (TBSP) to obtain physically meaningful curves in the latent space is noted. However, the use of TBSP seems like a handcrafted approach. The reliance on capturing TBSPs in order to identify physically meaningful trajectories in the latent space does not appear to be optimal. It introduces a scenario where errors in how the researchers capture the TBSPs (for example, if they do not ensure that only one physical parameter is changing between the sequence of photos) affect what trajectory will be traced out in the latent space.
Author Feedback
We thank reviewers for the thoughtful comments and for recognizing the relevance of the problem (all), the clarity of writing (all), and the quality of ablation study (R2, R3) and visualization (R3). Below, we address the concerns.
R1: Validation
We agree that validation is challenging and visual examples alone may be insufficient. While dermatologist assessment would be valuable, it is still limited to a small number of examples. Hence, we use a diagnostic model as an efficient proxy to evaluate diagnostic quality across many more images (Table 1). Its high performance indicates our method does not harm diagnostic quality.
R1: Race-specific Lesion Characteristics
We acknowledge that lesion appearance can differ across races and that modifying the surrounding skin colour may introduce biases. However, insufficient training data for under-represented populations also introduces biases. Prior work (e.g., [16]) establishes that biases removed by skin-tone transfer significantly outweigh those introduced (further confirmed by our results in Table 1).
R1 & R2: Comparison with Other Existing Works
Fair comparison with (Style)GAN-based lesion augmentation methods is difficult since their objectives and evaluation protocols differ—most are for generating synthetic samples or modifying lesion attributes (size, type, shape). In contrast, our method adapts skin colour and environmental factors across sites while preserving the lesion’s essential appearance. As [16] most closely aligns with an intended application, we adopt it as the primary comparison along with its evaluation protocol.
R2: Reproducibility
Space limitation precluded a full flowchart, but Sec. 2.1 describes model’s data flow and Sec. 2.3 details the application procedure. The code repository (including all details for reproduction) will be released upon publication.
R3: Shape and Length of the Color Embedding (CE)
Training encourages the model to set the number of active CE entries to be minimal while still capturing relevant colour properties (proof in [25]). CE length only needs to exceed the number of ground-truth factors. We set it to 256—sufficiently large, yet still computationally efficient.
Active entries are identified via per-entry variance across the dataset. By model design, low-variance entries have negligible visible effect and are thus inactive.
The 5 active entries (Fig. 1(b)) are basis-like latent coordinates rather than explicitly named physical SCCI factors. Physical factors (in Fig. 2) correspond to trajectories in the latent space rather than individual coordinates. They are obtained by varying CE along meaningful curves in the latent space (Sec. 3.2); more trajectories can be identified given additional TBSPs captured under controlled conditions.
R3: Independence of CE Entries
Under ideal conditions, CE entries are statistically independent (mathematical proof in [25]). In practice, this holds approximately (Sec. 3.4), as verified by low correlation (typically ≤ 0.1) and by rows 5–6 of Table 1. R3: Data Leakage?
Unlike the normalization method, we used the unlabelled test set when training the augmentation method. While some may view this transductive setting as leakage, similar setups are established and accepted in prior arts—e.g., document embeddings ([28], p. 5) are often trained on both training and test documents before supervised classification. This also matches the deployment scenario in Sec. 2.3.
The three-set setup suggested by R3 was considered during development but omitted due to space. The Acc. difference on the portion not seen by the colour model is around 1/1000. Other
R2: For fair comparison, we used the same datasets as [16]. R3: During testing, randomized decolourization can be replaced by a deterministic valid g_α, resulting in negligible difference. R3: Setting α, β ≥ 0 makes g_α strictly increasing at the cost of flexibility.
[28]: F. Incitti, et al. “Beyond word embeddings: A survey.” Info. Fusion, 2023
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
After reviewing this manuscript alongside the expert reviews, I note that the evaluations are mixed, with reviewers recognizing both the relevance of the problem and several potentially interesting aspects of the proposed framework. However, there are important concerns regarding validation, limited comparison with related generative approaches, insufficient justification and analysis of the latent space, possible data leakage in the current evaluation protocol, and the clinical realism and fairness implications of simulated skin-tone variations. Given that some of these issues may be clarified or partially addressed by the authors, I recommend allowing the paper to proceed to rebuttal for further consideration.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The paper is recommended for acceptance based on its solid application of information-bottleneck principles to unsupervised skin-colour disentanglement and its high practical relevance for health equity. Reviewers commended the clear writing, structured ablation studies, and insightful visualizations of the latent space use cases (educational visualization, data augmentation, and normalization).
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
The paper addresses an important problem in skin lesion analysis and proposes an interesting color-disentanglement framework. However, significant concerns remain regarding validation, potential bias from simulated skin-tone changes, limited comparison to related generative methods, and the transductive use of test data. The rebuttal clarified some technical points but did not sufficiently resolve concerns about clinical realism, fairness implications, and evaluation reliability.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
The rebuttal doesn’t fully address reviewers’ concerns.