Abstract

Skin lesion classification is a crucial clinical task in skin image analysis. Dermoscopy is widely adopted for the diagnosis of skin cancer due to its high image quality and the ability to reveal surface structures. However, dermoscopes require a certain level of expertise to use and thus become less effective when experienced dermatologists are scarce. In comparison, clinical images can be acquired with ordinary cameras, which are more accessible and are therefore increasingly utilized. Nevertheless, clinical images are less diagnostic-effective due to their lower image quality and greater imaging variability. In recent years, although public dermoscopic datasets and deep learning models have been successfully established, little research has been conducted on adapting these resources to clinical images. In this paper, we formulate this research topic as a semi-supervised domain adaptation problem and propose a data-efficient framework for skin lesion classification for clinical images. Abundant diagnostic features extracted from dermoscopic images will be established as informative knowledge vaults to facilitate the transfer of diagnostic capability from the dermoscopic image domain to the clinical image domain. Cluster-level and instance-level adaptation strategies are proposed for this knowledge transfer under different granularities. Comprehensive experiments across diverse datasets and settings demonstrate the superiority of our framework in skin lesion classification and data efficiency, showing its potential for clinical applications.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2731_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

N/A

Link to the Dataset(s)

N/A

BibTex

@InProceedings{ZhaZic_D2C_MICCAI2026,
        author = { Zhang, Zichen AND Shifai, Naweed AND Dzyubachyk, Oleh},
        title = { { D2C: Data-Efficient Skin Lesion Classification for Clinical Images by Representation Alignment Using Dermoscopic Resources } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16885},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The manuscript proposes a framework for data-efficient skin lesion classification from clinical images. The approach is structured into two stages: dermoscopic domain training and subsequent domain adaptation. In the first stage, a ResNet-18 classifier is trained on dermoscopic images to construct feature vaults. In the second stage, multiple domain alignment strategies are introduced to adapt the trained model to the clinical image domain in a semi-supervised manner, leveraging both dermoscopic and clinical data. Experimental results on the ISIC-2017 and ISIC-2018 datasets demonstrate improved AUROC performance compared to baseline methods.

  • 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 approach addresses an important and practical problem. 2.The method demonstrates encouraging classification performance, as evidenced by improvements in AUROC over the baselines. 3.The manuscript is well-written and organized.

  • 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 proposed framework involves a two-stage training procedure. However, the computational complexity, in terms of both training and inference time, is not discussed. A comparison with baseline methods in terms of efficiency would strengthen the practical relevance of the method. 2.The process for selecting hyperparameters (e. g. , parameters a and b, and the balancing coefficients) is not described. 3.The evaluation of classification performance is limited to AUROC. Including additional metrics would provide a more comprehensive assessment of the model’s performance. 4.It is unclear whether the baseline results were reproduced or directly taken from the original papers. This should be explicitly stated to ensure fairness and comparability of the evaluation.

  • 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

    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?

    The method is conceptually interesting and demonstrates encouraging results. However, the evaluation methodology is not sufficiently comprehensive. In particular, the assessment is limited to a single classification metric (AUROC). Furthermore, the study lacks an analysis of computational complexity, such as training and inference time, which is important for evaluating the practical applicability of the proposed approach.

  • 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.

    In the rebuttal, the authors stated that the inference time of the proposed method is comparable to that of existing methods; however, no quantitative results were provided. Consequently, the computational complexity of the proposed approach remains insufficiently discussed. Furthermore, the evaluation methodology relies solely on a single classification metric (AUROC) and should be extended to include additional evaluation metrics.

    Therefore, in its current version, the manuscript does not meet the threshold for acceptance.



Review #2

  • Please describe the contribution of the paper

    The authors propose D2C, a semi-supervised domain adaptation method that transfers knowledge from labeled dermoscopic images to clinical photographs. It uses class-aware feature vaults and four alignment strategies: centroid alignment, contrastive learning, FFT-based mixing, and augmentation consistency. Experiments on merged clinical datasets (PAD-UFES-20 + Fitzpatrick17k) with only 5–10% labels show clear improvements over recent baselines.

  • 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 an important real-world problem, bridging the gap between dermoscopic images and clinical photos used in practice. 2.The use of class-aware feature vaults is a good idea, helping stable feature learning and improving alignment and contrastive learning beyond normal batch methods. 3.The combination of multiple alignment techniques (centroid, contrastive, FFT mixing, consistency) is well designed and effective for handling domain shifts like lighting and color changes. 4.The method is practical for clinical use as it works well with very few labeled target samples (5–10%). 5.The evaluation is strong, with multiple datasets, low-label settings, good baselines, and clear ablation studies.

  • 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 novelty is limited, as most components are already known and the main contribution is their combination. 2.The evaluation is limited to binary classification, with no results for multi-class or imbalanced real-world scenarios. 3.The target datasets are merged without proper analysis, and separate results are not reported. 4.Important baselines (source-only, target-only, mixed training) and statistical tests are missing, making it harder to fully judge the gains.

  • 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 mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure 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

    The paper is clearly written with detailed overview figure. For future work, I recommend: (1) full multi-class evaluation, (2) vault-size ablation, and (3) releasing the merged clinical split.

  • 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 makes a useful contribution to an important area in skin imaging domain adaptation. The results look convincing even when very few labels are available, and the clinical motivation is strong. Although it builds on existing methods, the overall approach works well and meets the bar for acceptance. I support accepting the paper, especially if the authors add multi-class experiments and statistical tests in the final version.

  • 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.

    The authors have effectively addressed my concerns in their rebuttal.



Review #3

  • Please describe the contribution of the paper

    This paper proposes a domain adaptation method for skin lesion classification using partially labeled data in the target domain. The authors set dermoscopic images as the source domain and clinical images from ordinary digital cameras as the target domain, an approach that has some value.

  • 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 experimental results show that this method is effective. (2) The ways of applying Fast Fourier Transform and contrastive learning between the source and target domains are somewhat interesting.

  • 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.

    In the introduction, the authors emphasize data quantity imbalance as an essential challenge. However, it is not sufficiently clear how the proposed method specifically addresses this issue. In addition, the potential influence of batch-based computation or batch composition is not discussed.

    Moreover, the unlabeled images appear to be used in a fully unsupervised or self-supervised manner. To some extent, this makes the proposed framework resemble existing self-supervised representation learning methods, such as BYOL, which have also shown strong performance under semi-supervised fine-tuning protocols. Therefore, a more explicit discussion of the relation between the proposed method and existing self-supervised / semi-supervised approaches would help clarify the novelty and significance 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 submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure 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.

    (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 method is somewhat interesting and effective.

  • 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 greatly appreciate all reviewers for their thoughtful and constructive feedback. As follows, we will address the raised concerns. R1C1 Computational complexity. Inference time is crucial for clinical use. DivAug requires two forward passes, whereas the rest, including our method, require only one. Thus, our method achieves either faster or equivalent inference time compared to others, given identical classifiers. The first-stage training of a ResNet-18 on the selected ISIC-2017 under the paper settings takes 40 minutes. This stage only needs to be trained once and requires no future retraining. Hence, we believe the cost is sufficiently acceptable. We did not observe significant differences in training time in the second stage compared to others, as the methods differ primarily in their loss function designs. We will add a discussion about this. R1C2 Hyperparameters. They are selected based on experiments and typical values. For instance, the balancing factors are picked from 0.001 to 1, scaled by 10 each time. a is selected from 0.1 to 0.4 and b from 0.6 to 0.9.We will add a proper description of this. R1C3 Evaluation metric. We choose AUROC because it is the most comprehensive metric for skin lesion classification, accounting for data imbalance and different thresholds. Still, we acknowledge the usefulness of additional metrics. We will keep this in mind and apply it for future work. R1C4 Baseline results. All baseline results are reproduced. ReCo, FCAT, and SAT-DA involve skin image datasets but have misaligned dataset and task settings. Hence, we have to reproduce them in our uniform setting. All the other methods were using natural image datasets, making reproduction necessary. We will also add a description of this. R2C1, R2C2, R3C2 Method positioning, novelty. Our method is a semi-supervised domain adaptation approach that emphasizes domain adaptation, as leveraging dermoscopic resources for clinical images is clinically meaningful. The two lesion types in the datasets, nevus and melanoma, are imbalanced in quantity. Hence, the clinical task we address is differentiating melanoma from nevus using imbalanced data, a challenging and practical clinical problem. Since the data imbalance already exists in the datasets, we believe the evaluation results can reflect performance in real-world scenarios. The semi-supervised aspect arises because there are always unlabeled clinical images, which can be leveraged to improve performance. Based on our experiments, L_cdfc and L_fcui have effectively leveraged them, whereas overly complex semi-supervised schemes have adversely affected the performance. Methods such as BYOL are primarily for pure self-supervised representation learning to obtain general backbones, which we believe is not very relevant to our paper. We will add a further discussion for clarification. R2C3 Dataset merging. Too few test data lead to a more biased evaluation. Therefore, we merge the datasets to increase the test data size, enhancing the convincingness of the experimental results. We will add a description of this. R2C4 Baseline issue. The first setting in Sec. 3.3 uses labeled source images and 5% labeled target images, and our full method delivers a significant gain compared to it. The source-only and target-only settings will be inferior to the first setting, as they only use partial data. We believe this issue does not affect the convincingness of our method. R3C1 Data use. The large number of dermoscopic images in the Introduction is mentioned to introduce vault-based computation for L_cara and L_iwca. L_cara aims to align the feature centroids, where the centroid estimates are more accurate with more feature samples. L_iwca uses contrastive learning, which performs better with larger sample sizes. We agree that the description of this part needs refinement. Also, it would be beneficial to perform additional ablation studies for this. We will apply them in future work.




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.

    The paper makes a useful and well-executed contribution to data-efficient domain adaptation for skin lesion classification, with strong clinical motivation and convincing results under low-label conditions. While the novelty is incremental (assembly of existing techniques) and the evaluation has limitations (binary only, missing baselines, no efficiency analysis), the practical value and empirical rigor justify acceptance. The requested revisions (multi-class evaluation, statistical tests, computational analysis, and baseline additions) would strengthen the paper but are not essential for acceptance at this threshold.

  • 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 the reviewers’ concerns.



Meta-review #2

  • 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 valuable, practical contribution to data-efficient domain adaptation for skin lesion classification. Reviewers highlighted the strong clinical motivation of using rich dermoscopic resources to improve the classification of clinical photographs under extreme low-label constraints ($5\text{–}10\%$).



Meta-review #3

  • 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 presents a practical and clinically relevant approach for data-efficient skin lesion classification through domain adaptation between dermoscopic and clinical images. Reviewers appreciated the strong low-label performance, clear experimental validation, and practical motivation. While the methodological novelty is incremental, the overall contribution and empirical evidence are above the acceptance bar.



back to top