List of Papers Browse by Subject Areas Author List
Abstract
The application of machine learning to dermatology has grown substantially in recent years, moving beyond proof-of-concept studies toward potential applications. However, clinical dermatology remains a challenging and still open problem. Diagnostic assessment is often ambiguous, and skin lesions exhibit high variability, compounded by differences in acquisition modality, device quality, and patient demographics. These factors hinder the development of robust models suitable for safe and equitable clinical use. To support translation into practice, it is essential to systematically evaluate how contemporary models generalize across heterogeneous data sources and patient populations. In this work, we benchmark a diverse set of architectures on recent dermatology datasets, spanning dermoscopic images and smartphone-based clinical photographs. We assess the robustness of recent general-purpose and medical vision-language models, as well as foundation models, and compare them against task-specific dermatology classifiers, including embedding-based approaches and convolutional neural networks. Our study provides a comprehensive evaluation of model performance under distribution shifts, modality changes, and demographic variability. By quantifying the gap between current state-of-the-art models and the requirements of clinical deployment, we aim to contribute to the development of reliable, accessible, and clinically applicable AI systems for dermatology.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5997_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/TIC-13/derm-bench
Link to the Dataset(s)
PAD-UFES-20: https://data.mendeley.com/datasets/zr7vgbcyr2/
HAM10000: https://api.isic-archive.com/collections/212/
ISIC18: https://challenge.isic-archive.com/data/#2018
ISIC24: https://challenge.isic-archive.com/data/#2024
DDI: https://stanford.redivis.com/datasets/3r16-5mby7gfer
SD-198: https://derm.cs.sfu.ca/Welcome.html
DermAI-HC: https://arxiv.org/abs/2511.10367
BibTex
@InProceedings{dosEma_Bridging_MICCAI2026,
author = { dos Santos, Emanoel AND Cunha, Kelvin AND Mota, Rodrigo AND Papais, Fabio AND Bezerra, Thales AND Lopes, Natalia AND Medeiros, Erico AND Cruz, Shirley AND Araujo, Jessica AND Borba, Paulo AND Ing Ren, Tsang},
title = { { Bridging Research and Practice: A Systematic Evaluation of Generalist and Dermatology-Specific Models in Clinical Skin Lesion Classification } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16895},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper assesses a wide variety of AI architectures (CNNs, ViTs, VLMs) for dermatological lesion classification (benign vs malignant) with images from various modalities (dermascopic and clinical). They also explore the effects of different methods of data balancing, as data imbalance is a major challenge in dermatology, and explore different ways and prompts to obtain a classification from VLMs. On average, they find that embedding-based methods perform best, but are comparable to CNN- and ViT-based methods. However, they misclassify cases when the morphology does not match the typical class morphology (e. g. , smooth borders on a malignant case). Interestingly, they find that VLMs which have access to patient clinical information perform worse than morphology-only models, which highlights that they are not the solution for this and highlights that there exists a gap for models that can incorporate patient data in a way that results in better performance, especially on atypical lesions.
- 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.
This paper presents a particularly strong benchmarking for dermatopathology imaging. They test a wide-range of models, old and new, with various meaningful configurations and fine-tuning approaches. They also use a lot of data in their testing that covers two different modalities.
- 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 paper uses various public datasets for benchmarking but is missing an overview of the scale, diversity, and coverage of these combined datasets for the benchmarking task. To clarify the strength of the benchamrking which was performed, it is critical to report the total number of images used, the number of images per dataset, the fraction with a “benign” label, and if possible the fraction of images from people of colour which is relevant in dermatology.
- 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
Fig 1 uses light skin tones only, it would be helpful to see false positive and negatives from darker skin tones as well to see if there is a difference.
- 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.
(6) Strong Accept — must be accepted due to excellence
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper uses a very large benchmarking dataset and test many SOTA methods, providing a comprehensive survey of the best approaches for this task.
- 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 an extensive validation of various models in the task of skin lesion malignancy prediction. The authors compare various methods (vision foundation models, fine-tuning CNN and ViT architectures, and VLMs with multiple prompting approaches), using different datasets. The authors also evaluate the impact of different strategies to handle the severe imbalance found in the datasets. These studies are conducted for both dermatoscopy and clinical images, and their combination. Moreover, the selected datasets range from data containing skin cancer images to clinical datasets that comprise multiple types of skin diseases. The overall results suggest that visual models alone, especially fine-tuned ones, achieve better performances than general and medical VLMs.
- 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.
- This paper conducts an up-to-date and needed benchmark in the field of skin lesion analysis.
- The authors pick a variety of models, ranging from a recent dermatology foundational model (Derm Foundation), to generalist and medical VLMs. They also explore the use of DiNO and standard CNN and ViT backbones.
- Multiple datasets are used, with a focus on using both clinical and dermatoscopy data.
- Regarding the VLMs, there is a structured analysis of different prompting mechanisms, as well as the integration of metadata and other clinical knowledge in the prompt. It is unfortunate that this comparison is done included for other models, as metadata has been shown to be useful before.
- An additional experiment is conducted to assess the impact of class imbalance in the performance, in particular across three mechanisms to handle the imbalance.
- Overall the paper is easy to follow, but some experimental details are missing as well as the possibility to release code to reproduce the 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.
The main issues associated with this paper are about some of the methodological choices and lack of details/rational in some aspects of the comparisons.
-
The authors restrain the task to a binary prediction (malignant vs benign). However, most of the selected datasets are multiclasse. This design choice makes it impossible to have a finer analysis of the actual capacity of the various models.
-
Regarding the use of vision foundation models, it is unclear why did the authors pick only Derm Foundation (CNN backbone) and not another model that uses a ViT backbone (e.g, PanDerm). Given the differences between these architectures, this could have been interesting.
-
In the case of the VLMs, there is a similar issue, since the medical VLMs that are picked are generalists. However, MONET is a VLM for dermatology, that could have been included in this comparison.
-
While both the embedding based approaches and the fine-tuning ones achieved good performances without metadata, it would be more fair to also add this information inside the pipeline, to become closer to what it done in one of the prompting mechanisms for VLMs.
-
The fine-tuned transformer backbone seems to be missing from Table 3, or is it DiNOv2? If that is the case, why not fine-tuning DiNOv3 instead?
- Some results for EfficientNet seem to be missing from Table 3.
- It is not clear the rational for the choice of the imbalance handling strategies, as these are not usually applied in the dermatology field. Moreover, it is not clear if SMOTE is applied at the image or embedding level. This is a risk, since SMOTE tends to generate highly unrealistic data, if applied on raw images.
-
- 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 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
-
HAM10000 and ISIC2018 are closely the same dataset.
-
Please consider exploring a dermatology VLM (MONET) and an additional foundation model for dermatology with a transformer backbone.
-
Consider adding a github to ensure reproducibility of your experiments, together with some details on the hyperparameters and training configurations.
-
- 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 paper provides a relevant benchmark on different models for skin lesion malignancy prediction. It is well written, but lacking in some details that hinder the reproducibility of the method, like the hyperparameter range for SVM and XGBoost used together with the embedding models. Regarding the class imbalance, the paper also lacks clarity on wether class weights are used during training in the imbalance setting and whether SMOTE is applied to images or embeddings. The qualitative analysis is also confusing, as it does not provide details on the model that is being assessed. Some relevant and recent models are missing from this comparison, like PanDerm and MONET. Additionally, it would be worth mentioning papers that do similar comparisons.
- 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
This paper presents a systematic benchmark for binary malignancy risk prediction in dermatology across heterogeneous public datasets spanning dermoscopic and clinical images. The authors compare three broad model families: dermatology-specific/general embedding models, end-to-end CNN/ViT models, and vision-language models (VLMs), under single-dataset and merged-dataset settings, and additionally study class-imbalance handling and several VLM prompting variants. The main empirical finding is that dermatology-aligned embedding models and strong vision backbones consistently outperform the evaluated VLMs across most settings, suggesting that domain-aligned visual representations remain more reliable than prompt-based multimodal reasoning for this task.
- 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 addresses an important and timely problem: whether recent generalist and multimodal models actually translate into clinically meaningful robustness for dermatology. A major strength is the breadth of the empirical evaluation, which includes multiple public datasets spanning both dermoscopic and smartphone-based clinical images, as well as merged clinical, dermoscopic, and all-data settings designed to probe robustness under heterogeneous conditions. Another strength is that the paper does not limit itself to a single model family, but instead compares embedding-based pipelines, end-to-end CNN/ViT models, and multiple VLM configurations, including simple prompting, ABCDE-guided prompting, metadata-augmented prompting, and dual-encoder classification. The experimental results are also fairly consistent and clinically relevant: embedding-based approaches achieve the best average F1, CNN/ViT models are competitive, and VLMs underperform substantially, which supports the paper’s central message that domain-aligned visual representations remain critical in dermatology. The additional analysis of class-imbalance strategies and qualitative false positive/false negative examples further improves the practical value of the study.
- 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 of the paper is limited reproducibility. While the paper outlines the overall experimental setup, it does not provide sufficient methodological detail or publicly available code to enable faithful reproduction or to assess whether the comparison across model families is fully fair. Key implementation details, such as hyperparameter tuning, exact prompt design, model selection procedures, repeated runs, and performance variance, are either missing or insufficiently described.
A further weakness is the way the VLM family is instantiated and interpreted. Many VLM baselines are evaluated mainly through zero-shot or prompt-based inference, where the model is instructed to act as an “expert dermatologist,” but this does not establish real clinical expertise. Therefore, the weaker VLM results may partly reflect limited dermatology adaptation and possible domain mismatch from general-purpose pretraining, rather than a definitive limitation of VLMs for this task.
- 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
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?
I lean toward Weak Reject. The paper addresses an important and clinically relevant problem and presents a broad benchmark across dermoscopic, clinical, and merged dermatology datasets. The empirical results are also fairly consistent, with dermatology-aligned embedding models and strong vision backbones outperforming the evaluated VLMs. However, the current manuscript falls short of the acceptance bar mainly because reproducibility is limited: important implementation details and publicly available code are missing, making it difficult to faithfully reproduce the results or assess whether the comparison across model families is fully fair. In addition, the comparison itself is not fully controlled, since embedding-based models use frozen backbones with supervised downstream classifiers, CNN/ViT models are fine-tuned end-to-end, and many VLM baselines are evaluated mainly via prompt-based or dual-encoder inference. Therefore, the results more directly support the weakness of lightly adapted VLMs in this setting rather than a definitive conclusion about the broader suitability of VLMs for dermatology. A further concern is that many of the LLM/VLM baselines rely on general-purpose pretraining and prompt-based “expert dermatologist” framing, which raises questions about domain mismatch and clinical reliability. Overall, I find the study relevant and potentially useful, but in its current form the evidence is not yet rigorous enough to fully support the strength of its claims.
- 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 sincerely thank all reviewers for their time, effort, and constructive feedback. We greatly appreciate the valuable insights provided. Below, we address key comments and clarify potential misunderstandings.
1.Scale, diversity, and coverage of datasets We will provide a detailed description of all datasets in the public repository, including references to the original databases where distributions can be verified. Due to manuscript space limitations, we included only an overall summary. In the revised version, we can clarify key attributes, i.e., total number of images and label distributions.
2.Figure 1 and skin tone diversity Thank you for this suggestion. We agree that broader skin tone representation is important and will revise Figure 1 to include examples from darker skin tones.
3.Binary prediction despite multiclass This is an important observation. We focused on benign vs. malignant classification because lesion malignancy risk has the strongest impact in primary care and early consultations, where prioritizing high-risk patients is often more clinically relevant than fine-grained lesion categorization for non-specialist physicians. Nevertheless, we agree that multiclass evaluation provides complementary insights and plan to include these results in the public benchmark repository which will be linked in the revised manuscript.
4.Selection of vlms and foundations We appreciate this suggestion. We included Derm Foundation due to its relevance as a dermatology-specialized model; however, we agree that transformer-based architectures would enrich the comparison and offer insights into architectural differences. We consider this an important benchmark extension and plan to continuously expand evaluations in future works.
5.Missing details in tables/text We will revise table content to explicitly clarify the use of the correspondent backbone and correct missing values.
6.Limited reproducibility We appreciate this important concern. Due to manuscript space constraints and restrictions on supplementary materials, some methodological details may have been insufficiently described. We will release the repository containing implementation details, including hyperparameters, prompts, model selection procedures, and experimental settings, and aim to include the repository link in the camera-ready version which was omitted for blind review.
7.Interpretation and evaluation of general VLMs We agree with the reviewer’s observation. Our goal was not to claim an inherent limitation of VLMs for dermatology, but rather to evaluate commonly adopted settings (e.g., frozen backbones, zero-shot prompting, limited fine-tuning) relative to specialized dermatology-trained models. The results highlight potential risks of applying general-purpose VLMs in clinical settings without domain adaptation. We agree that weaker performance may reflect domain mismatch rather than fundamental limitations. To avoid potential misunderstandings, we will explicitly discuss this point in the conclusion.
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 reviewers agree that this paper addresses a timely and important clinical problem by providing a comprehensive benchmark for skin lesion classification. The study is well-regarded for its extensive comparison across different model architectures (CNNs, ViTs, and VLMs) and its evaluation of both dermoscopic and clinical images.