List of Papers Browse by Subject Areas Author List
Abstract
Despite recent advances in deep generative modeling, skin lesion classification systems remain constrained by the limited availability of large, diverse, and well-annotated clinical datasets, resulting in class imbalance between benign and malignant lesions and consequently reduced generalization performance. We introduce DermaFlux, a rectified flow-based text-to-image generative framework that synthesizes clinically grounded skin lesion images from natural language descriptions of dermatological attributes.
Built upon Flux.1, DermaFlux is fine-tuned using parameter-efficient Low-Rank Adaptation (LoRA) on a large curated collection of publicly available clinical image datasets. We construct image-test pairs using synthetic textual captions generated by Llama 3.2, following established dermatological criteria including lesion asymmetry, border irregularity, and color variation. Extensive experiments demonstrate that DermaFlux generates diverse and clinically meaningful dermatology images that improve binary classification performance by up to 6% when augmenting small real-world datasets, and by up to 9% when classifiers are trained on DermaFlux-generated synthetic images rather than diffusion-based synthetic images. Our ImageNet-pretrained ViT fine-tuned with only 2,500 real images and 4,375 DermaFlux-generated samples achieves 78.04% binary classification accuracy and an AUC of 0.859, surpassing the next best dermatology model by 8%. Code, data and pretrained models are available at dermaflux.github.io.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3890_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/dermaflux/dermaflux
Link to the Dataset(s)
Training captions: https://github.com/dermaflux/dermaflux/blob/main/data/curated_dataset_captions.zip
Synthetic dataset: https://huggingface.co/datasets/StathisGln/DermaFlux_synthetic_dataset
BibTex
@InProceedings{GalSta_DermaFlux_MICCAI2026,
author = { Galanakis, Stathis AND Koliousis, Alexandros AND Zafeiriou, Stefanos},
title = { { DermaFlux: Synthetic Skin Lesion Generation with Rectified Flows for Enhanced Image Classification } },
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
1.The paper proposes DermaFlux, a rectified flow-based text-to-image generative framework built on Flux. 1 and fine-tuned using LoRA for dermatological image synthesis. 2.It introduces a clinically-guided captioning pipeline using LLM-generated descriptions aligned with dermatological criteria to enable semantically consistent image generation. 3.The generated synthetic data effectively addresses data scarcity and class imbalance, leading to improved downstream classification performance.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
Major Strengths 1.The manuscript is well-structured and addresses a clinically relevant problem, especially data scarcity in dermatological imaging for skin cancer diagnosis. The motivation for rendering synthetic data generation is clearly articulated, and the paper is drafted within the context of recent advances in diffusion-based generative models by deploying FLUX with LoRA. 2.The authors propose a rectified flow-based generative framework combined with text-to-image conditioning for synthesizing benign and malignant skin lesion images. The use of rectified flows methods represents a suitable alternative to conventional diffusion pipelines, with potential advantages in sampling efficiency and controllability. 3.Experimental results indicate that augmenting limited real-world datasets with the proposed synthetic data yields a performance gain of approximately 6%. Furthermore, the method achieves up to 9% improvement over prior diffusion-based augmentation approaches under comparable training settings, and outperforms existing state-of-the-art classifiers by approximately 8%, demonstrating the practical utility of the approach
- 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.
Major Weaknesses 1.While the use of rectified flow-based models is technically relevant, the overall methodological contribution appears to be limited. The proposed framework closely resembles prior work such as Derm-T2IM (2024), which also employs text-to-image diffusion models for generating binary (benign vs. malignant) skin lesion dataset. The manuscript would benefit from a clearer articulation of what fundamentally differentiates the proposed approach beyond the choice of different generative backbone. 2.The study is limited to binary classification (benign vs. malignant), which restricts its clinical applicability. Given the flexibility of text-conditioned generative models, it would be valuable to evaluate the capability of the proposed framework in generating a broader spectrum (atleast 3-4 classes) of dermatological conditions (e. g. , melanoma subtypes, basal cell carcinoma, squamous cell carcinoma, or other lesion categories/ skin conditions). This would better demonstrate the scalability and generalization capacity of the method. 3.The evaluation primarily focuses on downstream classification accuracy, with limited analysis of the quality and diversity of the generated samples. Incorporating additional quantitative metrics (e. g. , FID, precision/recall for generative models) would provide a more comprehensive assessment of synthetic data quality. 4.Finally, an assessment of clinical realism would further strengthen the study. Incorporating expert evaluation (e. g. , dermatologists) to validate the fidelity and diagnostic relevance of the generated images would significantly enhance the credibility and practical impact of the work.
- 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
The paper addresses an important problem and demonstrates empirical improvements; however, the novelty over existing text-to-image diffusion-based approaches is limited. Expanding the scope beyond binary classification and providing a more comprehensive evaluation of generative quality would significantly strengthen the contribution.
- 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 paper addresses an important and clinically relevant problem, with challanges like data scarcity and class imbalance in dermatological imaging. The proposed use of rectified flow-based generative modeling combined with text-to-image conditioning and LoRA fine-tuning is technically sound and well-motivated. The incorporation of clinically meaningful textual descriptions for guiding image synthesis is a notable strength, and the experimental results demonstrate consistent improvements in downstream classification performance.
However, the overall novelty is somewhat limited, as the proposed framework closely resembles existing text-to-image diffusion-based methods (e.g., Derm-T2IM), with the primary distinction being the choice of generative backbone. Additionally, the evaluation is restricted to binary classification, which limits clinical applicability, and lacks a comprehensive assessment of generative quality (e.g., diversity, realism metrics, or expert validation).
Overall, while the paper demonstrates a valid approach for generating synthetic skin lesion data and better empirical results, the incremental methodological contribution and limited evaluation scope influenced the final score.
- 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 have somehow justified the current limitations. Althought the novelty is yet very limited in terms of binary data generation but the use of FLUX diffusion model is still unexplored which is the strong aspect of this work.
Review #2
- Please describe the contribution of the paper
The main contribution is the development of a rectified flow-based generative framework that synthesizes clinically grounded skin lesion images from natural language descriptions. It is useful for improving skin lesion classification performance, particularly in data-scarce settings, by providing high-quality synthetic images that enhance the training of classifiers.
- 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 a significant challenge in medical image analysis, which is the scarcity of large, well-annotated datasets for skin lesion classification. Even though it is not a new idea, developing a method to generate synthetic data that can augment real datasets is a valuable contribution.
-
The paper successfully adapts a state-of-the-art generative model for dermatological image synthesis. The Flux.1 backbone is relatively unexplored in the field compared to diffusion models.
-
The authors construct a ~500k image-text dataset derived from public dermatology sources, which is a substantial effort and provides a valuable resource for training the generative model. Also, the synthetic captions are structured around clinical attributes.
-
The paper carries out experiments to show that the generated synthetic images improve the classification performance of both ResNeXt and ViT classifiers for binary skin lesion classification.
-
The paper is very well written, and the methodology is clearly described. The results are presented in a clear and comprehensive manner.
-
- 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 synthetic captions were generated based on the ABC components of the ABCDE rule. While this is a reasonable approach, this rule is designed to deal with pigmented lesions, in particular melanoma. The datasets described in Table 1 contain a variety of lesion types, including non-pigmented lesions such as basal cell carcinoma, squamous cell carcinoma, among others. Therefore, the generated captions may not be fully representative of the diversity of lesions in the training data, which could limit the generalizability of the synthetic images.
-
Also, the dataset mixed clinical and dermoscopic images, which have different visual characteristics. The paper does not clarify whether the synthetic images were generated separately for each modality. For example, in Fig. 3, it seems that the generated images are dermoscopic, but it is not clear if the model can also generate clinical images.
-
Only 100 captions were evaluated by clinical experts for medical plausibility. Even though it is a time-consuming task, given the large size of the dataset, it would be important to evaluate a larger sample of captions to ensure that they are consistently clinically plausible and accurately reflect the lesion characteristics. This raises concerns about systematic errors or hallucinations in the generated captions.
-
The paper focuses on binary classification (benign vs malignant), but it does not specify the types of lesions included in each category. For example, it is not clear if the malignant category includes only melanoma or also other types of skin cancer. This information is important to understand the clinical relevance of the results and the potential impact of the synthetic data on different lesion types.
-
The evaluation focuses almost entirely on classification performance, without assessing the intrinsic quality of generated images. While improved classification accuracy suggests that the synthetic images are useful, it does not directly demonstrate that they are realistic or clinically meaningful. Additional evaluations, such as human expert assessment of image quality or quantitative metrics for image realism, would strengthen the paper’s claims about the quality of the generated images.
-
The training and evaluation datasets are aggregated from similar public sources. It raises concerns about potential data leakage or overlap between training and test sets, which could artificially inflate performance metrics. The paper should clarify how they ensured that there is no overlap between the images used for training the generative model and those used for evaluating the classifiers.
-
The paper does not provide details about the computational resources required for training DermaFlux, which is important for reproducibility and for other researchers who may want to adopt or build upon this work. Also, there is no code or model release mentioned, which limits the ability of the community to validate and extend the findings.
-
The paper does not compare DermaFlux against other types of generative models beyond diffusion-based methods. For example, it would be interesting to compare against GANs or VAEs that have been used for skin lesion synthesis in prior works (for example, Bissoto et al. 2018).
-
The authors report that in both experiment scenarios, the datasets are balanced between benign and malignant skin lesions. However, in real-world clinical datasets, there is often a significant class imbalance, with benign lesions being much more common than malignant ones. It would be important to evaluate the performance of classifiers trained with DermaFlux-generated data under more realistic class imbalance conditions to better understand their practical utility in clinical 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 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.
(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 presents a useful and well-executed approach for improving skin lesion classification using synthetic data, and the experimental results are compelling. The dataset construction and clinically guided captioning are also valuable contributions that can benefit the community. However, there are some limitations and areas for improvement, particularly regarding the evaluation of the generated images and the potential for data leakage. The paper would be stronger with additional assessments of image quality and more clarity on the dataset splits.
Overall, I believe the paper makes a meaningful contribution to the field of medical image synthesis and classification, but it would benefit from addressing the identified weaknesses to strengthen its claims and impact.
- 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.
I appreciate the authors’ detailed response to the reviewers’ comments. I’m still concerned about some topics, such as using ABC rule for caption generation of non-pigmented lesions, and the lack of evaluation of the generated images. However, I think the experiments indicate that the generated images are useful for improving classification performance, which is the main goal of the paper. I also appreciate the authors’ willingness to address these concerns in future work and to release their code and dataset. Therefore, I suggest accepting the paper, but I encourage the authors to consider these points in their future research.
Review #3
- Please describe the contribution of the paper
The paper introduces a rectified flow–based generative framework (DermaFlux) for synthesizing skin lesion images. A large-scale dermatology dataset is constructed by aggregating public datasets and generating attribute-level captions. The method adapts a pretrained Flux.1 backbone via LoRA for efficient domain specialization. Extensive experiments demonstrate that the generated synthetic data improve downstream skin lesion classification and outperform diffusion-based augmentation 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 work addresses data scarcity and class imbalance in dermatology, which are practical bottlenecks in clinical deployment. 2.The use of attribute-level captions provides a more clinically grounded supervision signal compared to generic text prompts. 3.Results indicate clear gains in low-data regimes, especially useful for real-world scenarios where labeled medical data are limited.
- 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 approach mainly combines existing components (rectified flows, LoRA, synthetic captions) without introducing a fundamentally new modeling framework. The novelty lies more in integration than in algorithmic contribution. 2.While captions are partially verified, there is no systematic expert evaluation of the generated images. It remains unclear whether the synthetic lesions are clinically reliable beyond visual plausibility. 3.The quality of generated images is assessed primarily through downstream classification performance. Additional generative metrics or human evaluation would strengthen the claims. 4.The comparison is limited to a specific diffusion-based method, inclusion of more recent or competitive generative baselines would improve fairness. 5.The use of automatically generated captions may introduce noise or bias, and the paper does not analyze its impact in detail.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper addresses an important problem in medical imaging and presents a well-executed framework for synthetic data generation. The experimental evaluation is solid and demonstrates consistent improvements, particularly in low-data settings. However, the methodological novelty is limited, as the approach mainly integrates existing techniques. In addition, the lack of direct clinical validation of generated images weakens the overall impact.
- 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.
I decide to retain my original rating and acceptance recommendation.
Author Feedback
We thank the reviewers R1, R2, R3 and meta-reviewer for their careful evaluation and positive assessment. We are encouraged that they recognized the clarity and technical soundness of the paper, the relevance of addressing data imbalance in medical imaging, and the promise of our generative framework as an alternative to diffusion-based approaches. They also highlighted the value of our clinically structured dataset and strong low-data performance gains.
Novelty (R1, R2, R3): Our contributions extend beyond the choice of backbone architecture. Our method uses natural language to generate skin-lesion images, enabling detailed control through fine-grained low-level attributes. In contrast, DermT2I is constrained by the CLIP encoder and its 77-token limit, restricting the level of detail and control expressible in text. We also introduce a large-scale dataset with fine-grained descriptions based on established dermatological standards (ABC rule). This differs from the more limited textual supervision used in DermT2I.
Binary classification (R1, R2): We focus on the benign-vs-malignant setting because our primary clinical motivation is to support early screening and reduce unnecessary referrals to dermatologists. We investigate whether synthetic augmentation can improve reliable discrimination in this clinically relevant setting. Our ROC analysis demonstrates that threshold adjustment can substantially reduce unnecessary consultations at a sensitivity comparable to state-of-the-art models. Moreover, our text-conditioned framework can be extended to multi-class lesion generation by conditioning on skin lesion subtypes. However, this remains future work, as some datasets contain only benign/malignant labels, while others provide finer subtypes requiring expert-defined clustering for multi-class evaluation.
Evaluations (R1, R2, R3): We believe that downstream classification performance is a meaningful and widely adopted way to evaluate synthetic medical images in augmentation settings, as also adopted in DermT2I. The main objective of our method is to generate meaningful synthetic samples that improve skin-lesion classification in low-data regimes. The observed improvements in both classifiers suggest that the generated images preserve diagnostically relevant cues. We agree that additional analyses, such as large-scale expert studies on the generated images and captions, or experiments under class-imbalance settings, could provide further insight into image realism and clinical fidelity. Nevertheless, during development, we consulted clinical experts regarding the plausibility of the generated samples.
Caption generation (R2, R3): After consulting our experts, we selected the ABC template as a practical generic caption framework, considering time and annotation costs. Our goal is not to generate exhaustive dermatological reports but simple, structured conditioning signals that can be reliably observed and described by a VLM, and later used as input to DermaFlux. We plan to enrich these captions with more human-annotated descriptions, including details about non-pigmented lesions.
Dermoscopic-vs-Clinical (R2): The current formulation does not explicitly encode the image type, as it is not included in the input captions. In fact, DermaFlux generates both modalities because both are represented in the training data.
Data leakage (R2): To verify the absence of data leakage, we compared image checksums across splits, confirming that no images were shared.
Baselines & Code (R2, R3): The comparisons focus on recent diffusion-based dermatology methods, as they currently represent the dominant paradigm for high-quality image generation. We will update the related work with GAN-based methods, clarify the training resources used, and release our code and dataset.
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 tackles an important problem in dermatology, namely data scarcity, and shows improvements in classification performance using synthetic data. The idea is practical and the results are encouraging.
At the same time, the novelty seems moderate, as the method mainly builds on existing approaches. The evaluation also focuses mostly on classification, with limited analysis of the quality and realism of the generated images. Some details about the dataset and setup are not fully clear.
Overall, the work is promising, but these points need to be clarified in the rebuttal.
- 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 addresses data scarcity in dermatology imaging through text-conditioned synthetic lesion generation. The reviewers noted the practical value of the framework and the observed gains in low-data classification settings. The rebuttal clarified the relation to prior text-to-image methods, dataset construction, data leakage checks, and the scope of the binary classification setting. Some limitations remain regarding direct expert evaluation of generated images and broader generative baselines, but the updated reviewer feedback supports acceptance.
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 practical value and strong empirical results in addressing data scarcity and class imbalance in dermatological image analysis. Reviewers highlighted the value of the large-scale image-text dataset and the clinically guided captioning pipeline constructed using attribute-level descriptions.
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.
Although the methodological novelty is incremental and the evaluation would benefit from stronger generative quality and expert clinical validation, all reviewers converged on an accept post-rebuttal, recognizing consistent classification gains, usefulness in low-data settings, and the authors’ satisfactory responses and commitment to release code/data.