List of Papers Browse by Subject Areas Author List
Abstract
Polyp segmentation is critical for early colorectal cancer screening, yet its performance is limited by the quantity and diversity of annotated data. Existing synthesis methods typically rely on masks from public datasets, inheriting structural biases and limiting morphological diversity, while often neglecting style realism, producing synthetic images with inconsistent color and texture. To address these issues, we propose SFD-Polyp, a style-fused conditional diffusion framework that conditions generation on ground-truth masks and injects appearance cues via a Style Integration Module (SIM), enhancing texture and style consistency while preserving mask-aligned semantics. We further introduce Morphology-Preserving Mask Synthesis (MPMS) to generate multiple anatomically plausible mask variants at inference, expanding morphological diversity. Experiments on five public polyp datasets demonstrate that training with the synthesized data consistently improves segmentation across multiple baseline models, highlighting the potential of diffusion-based data generation to enhance generalization in data-limited medical imaging scenarios. Our code is avaliable at: https://github.com/zhaiying0925/SFD-polyp.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4818_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/zhaiying0925/SFD-polyp
Link to the Dataset(s)
N/A
BibTex
@InProceedings{HanLon_Learning_MICCAI2026,
author = { Han, Longfei AND Zhai, Ying AND Li, Haisheng},
title = { { Learning Diverse and Realistic Polyp Data via Style-Fused Diffusion Models } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16892},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper uses a diffusion model of GenAI to synthesize/augment Polyp data for training Polyp Segmentation Networks. This is a type of training data augmentation. Experiments show very good results
- 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.
GenAI for synthesizing (augmenting) polyp segmentation data. Adding texture and style to the data.
- 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.
This is a type of data augmentation 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
In Table 2, what if you train other networks using your synthesized/augmented data set? Would you be able to univerally improve them? Or, what if you make your synthesized data set public? Is the proposed method applicable to other medical imaging problems?
Fig. 2 shows that the synthesized images (S. img) are visually quite different from the original input images. A network trained to perform well on both the original and synthesized data will perform better for the original test data?
- 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 is a more effective data augmentation method that uses GenAI to derive styled training data.
- 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 satisfactory addressed my questions.
Review #2
- Please describe the contribution of the paper
The main contribution of this paper is a style-fused conditional diffusion framework, named SFD-Polyp, for polyp image synthesis, which is designed to improve both the realism and morphological diversity of generated training data for polyp segmentation. Specifically, the paper introduces a Style Integration Module (SIM) to inject appearance cues from real endoscopic images into the diffusion process, and a Morphology-Preserving Mask Synthesis (MPMS) strategy to generate diverse yet anatomically plausible mask variants for conditional image generation. Through these two components, the method aims to produce synthetic polyp images that better match real visual styles while expanding structural variation beyond fixed dataset masks, thereby improving the generalization of downstream segmentation models.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The paper has several strengths. First, it addresses an important and clinically relevant problem, namely the lack of diverse annotated data for polyp segmentation. Second, the proposed framework is reasonably well motivated: it attempts to improve both the visual realism of synthetic images and the morphological diversity of mask conditions, which are both meaningful limitations of existing data generation methods. In particular, the combination of a style integration module and a mask synthesis strategy gives the method a clear overall design. Third, the paper includes experiments on multiple public datasets, different segmentation backbones, as well as ablation and sensitivity studies, which provides a reasonably thorough evaluation of the downstream usefulness of the proposed synthetic data.
- 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 paper suffers from a major methodological inconsistency between training and inference. In Section 2.1, the Style Integration Module injects style features extracted from a real image into the diffusion model during training, while the inference description states that the same module remains active to improve appearance realism, yet the paper never clearly specifies where the style reference comes from at test time. As a result, the actual generation process appears to depend not only on the mask condition, but also on an additional style input, which is not properly defined in the paper. This issue directly affects the theoretical coherence, reproducibility, and fairness of the method. 2.The paper does not explain how the real image used for style extraction is selected during inference. This point is critical. If the style reference is sampled from the test set, then the evaluation is fundamentally compromised by data leakage. If it is sampled from the training set, this should be stated explicitly in the main text together with the sampling protocol. At present, the role and source of the style reference remain ambiguous, which raises serious concerns about experimental validity. 3.The claimed novelty is rather limited. For the Style Integration Module, injecting style information into a diffusion backbone is not, by itself, a new idea, since style injection and feature conditioning have already been widely explored in the broader diffusion and computer vision literature. Here, the main difference seems to be only a weighted fusion design rather than a standard cross-attention formulation. For the mask synthesis module, the use of SLIC superpixels to perturb masks is also based on long-established image processing techniques, and the BFS-based expansion with an overlap threshold is essentially a heuristic variant of morphology-aware mask perturbation rather than a fundamentally new formulation. More broadly, the overall combination of conditional diffusion, mask perturbation, and style conditioning is already closely related to prior polyp synthesis works such as Polyp-DDPM and ArSDM. Therefore, the paper appears more incremental than substantially novel. 4.The evidence supporting the main claims is insufficient. The paper emphasizes that the proposed method produces more realistic and more diverse synthetic images, yet the experimental evaluation focuses almost entirely on downstream segmentation performance using Dice and IoU. While these metrics are appropriate for segmentation, they do not directly validate image realism or diversity. The paper does not provide standard image-generation metrics such as FID, IS, or LPIPS, nor does it include expert visual assessment or any direct comparison of generated image quality. Without such evidence, it is difficult to conclude that the proposed method is actually better at image synthesis than prior methods. 5.The mathematical presentation needs improvement. Some indices are used inconsistently, the notation around Equation 8 is awkward and appears improperly formatted, and some set expressions are not written in standard mathematical form. These issues are not the core weakness of the paper, but they do reduce clarity and make the method appear less rigorously presented. 6.The paper must clarify what exactly the style reference is at inference time. Is a real image randomly sampled from the training set? Is each mask paired with a specific reference image? Or is no explicit reference needed at all? If no reference is needed, then the paper must explain how the fused style feature is obtained during inference. As currently written, the generation pipeline is not closed, and this ambiguity prevents the reader from understanding the actual inference procedure.
7.There is also a serious logical problem in the mask synthesis pipeline. The paper claims that the proposed mask synthesis strategy increases morphological diversity by generating anatomically plausible mask variants through superpixel-guided expansion. However, after smoothing, the generated mask is intersected again with the original mask to produce the final condition. This is highly problematic, because if the final mask is constrained back to the support of the original annotation, then it cannot truly introduce meaningful outward shape variation beyond the original boundary. In that case, the method may generate only restricted sub-variants, rearrangements, or mild boundary perturbations of the original mask, which directly conflicts with the paper’s repeated claim that it substantially enriches structural diversity.
- 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?
My overall score is mainly driven by the gap between the paper’s intended contribution and the actual level of technical clarity and experimental support provided in the current version. The paper addresses an important and clinically relevant problem, and the general motivation of improving both appearance realism and morphological diversity in synthetic polyp data is meaningful. The experimental setup also includes multiple datasets, different segmentation backbones, and ablation studies, which shows a reasonable amount of effort. However, I have several major concerns that substantially weaken the paper. Most importantly, the proposed Style Integration Module is not clearly defined at inference time: the paper does not explain where the style reference image comes from during generation, which creates a serious inconsistency between training and inference and also raises concerns about possible data leakage. In addition, the claimed novelty appears limited relative to prior polyp synthesis methods such as Polyp-DDPM and ArSDM, and the mask synthesis strategy seems more like a heuristic perturbation scheme than a fundamentally new formulation. The evidence for the key claims is also not fully convincing, because the paper mainly reports downstream segmentation performance but does not directly evaluate the realism or diversity of the generated images. Finally, there is a logical weakness in the mask generation pipeline itself, since the final intersection with the original mask seems to restrict the claimed structural diversity. Taken together, these issues make me feel that, although the paper is interesting and potentially useful, the current version does not yet provide a sufficiently rigorous, clearly justified, and convincingly novel contribution.
- Reviewer confidence
Very confident (4)
- [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.
The main weaknesses remain: the mask-diversity claim is still overstated, direct generation-quality evaluation is missing, and the novelty over prior conditional diffusion-based polyp synthesis methods is limited. Therefore, I maintain my rejection recommendation.
Review #3
- Please describe the contribution of the paper
Acceptable performance is demonstrated in Tables 1 and 2.
- 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.
Positives 1.This paper uses diffusion-based generation to tackle the problem of data scarcity, which is extremely important. 2.It incorporates SIM and MPMS to ensure higher fidelity and variation in results. 3.It shows reliable segmentation enhancement in various datasets and architectures.
- 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.
Technical Problems 1.Less novelty: Similar to current conditional diffusion + style transfer algorithms. 2.It does not include any metric evaluating image generation (FID, SSIM, human evaluation, etc.). 3.High dependency on the availability of ground-truth masks, thus reducing scalability in practical applications. 4.No theoretical guarantee for anatomical accuracy provided by MPMS except for IoU constraint. 5.Insufficient comparison with other diffusion-based approaches for medical image synthesis. 6.No study of the effect of hyperparameters such as diffusion iterations, noise schedule, and α on model performance. Include few more performance metrics and results. 7.It claimed generalization ability is indirect and depends only on segmentation enhancement.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
8/10
- 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
Author Feedback
We sincerely thank the reviewers and Area Chair for their detailed and constructive comments.
R2/AC – Style reference during inference. The style reference used during inference is sampled only from the training set, and no test-set images are used. Specifically, we construct a style bank from all training images, the same mask can be paired with different training style to produce diverse appearances. The style feature only provides global color or texture cues, while lesion geometry is governed throughout denoising. The protocol is reproducible and free from data leakage. We will clarify this clearly.
R2/R3/AC – Generation quality evaluation. We agree that Dice/IoU do not directly measure image realism or diversity. Our work is positioned as task-oriented data synthesis for improving polyp segmentation, where the key question is whether generated samples improve downstream task under limited annotations. Related polyp synthesis works also commonly emphasize downstream performance, since synthetic images are ultimately used as training data. We will clarify this evaluation scope, and discuss FID/LPIPS or expert assessment as complementary future evaluations.
R2/R3/AC – Logical problem in mask synthesis. We thank the reviewer for careful observation. We agree that the final intersection prevents unrestricted outward expansion beyond the original mask. This is intentional: in polyp segmentation, the training set already contains polyps of different sizes, and arbitrarily enlarging each lesion may introduce anatomically implausible supervision. MPMS is therefore designed to generate morphology-preserving variants, rather than unconstrained new shapes. These variants still provide useful augmentation by introducing local contour or sub-region variations while maintaining clinical plausibility. We will revise the manuscript to clarify that MPMS generates anatomy-constrained mask variants.
R2/R3/AC – Novelty. We agree that Polyp-DDPM and ArSDM are closely related baselines. Their generation mainly relies on existing GT masks as fixed structural conditions. In contrast, our method augments the mask condition by deriving morphology-preserving variants from the training masks. The novelty lies in how constraints are coupled with diffusion synthesis. Specifically, SIM is not merely a weighted fusion block, but converts real endoscopic appearance into a controllable style condition for rendering the same lession structure under different realistic styles. Rather than directly reusing the fixed GT masks, MPMS synthesizes new variant image-mask pairs from existing annotations without extra manual labeling. Thus, SLIC, BFS expansion, and smoothing serve as a constrained mask-condition generator, not independent heuristics. We will revise the manuscript to clarify this clearer.
R1 – Generalization of Synthesized images. The synthesized data is intended as a general augmentation source rather than being tied to a specific segmentation network. Experiments on two backbones show consistent improvements, suggesting good generalization (Table 2). The code will be released upon acceptance. Regarding Fig. 2, S. Img is not intended to reconstruct the input image, it is a new sample generated under mask and style conditions. Therefore, visual differences are expected and help increase appearance diversity. Since real and synthetic samples are jointly used during training, the model still learns from real data while benefiting for other senarios.
R2/R3 – Reproducibility, formulation, and limitations. We will carefully revise mathematical notation, equation formatting, and set definitions for clarity. We acknowledge that the SFD-Polyp relies on available masks, and does not provide a formal anatomical guarantee for MPMS. Its goal is not to replace annotation, but to maximize limited annotations through conservative, plausible variants. Broader hyperparameter sensitivity and diffusion-specific comparisons will be discussed as 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.
While Reviewer 3 gives a strong Accept, the concerns raised by Reviewer 2 are fatal and cannot be adequately addressed during rebuttal. Specifically: 1.Methodological inconsistency: The Style Integration Module requires a style reference image during inference, but the paper never specifies where this reference comes from. This creates ambiguity about reproducibility and raises potential data leakage concerns if test-set images are used. 2.Logical flaw in mask synthesis: The paper claims to generate diverse masks, but the final step intersects the synthesized mask back with the original annotation, which contradicts the claim of generating structurally novel masks beyond existing data. 3.Missing direct evaluation of generation quality: For a generative model paper, the lack of FID, IS, or human evaluation metrics is a significant gap. 4.Limited novelty: The method largely combines existing techniques (style injection, superpixel-based mask perturbation, conditional diffusion) and is closely related to prior polyp synthesis works such as Polyp-DDPM and ArSDM.
Given these fundamental issues, the paper does not meet the bar for early acceptance.
- 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.
Reviewer #2 provided a detailed, fatal critique identifying that the style reference during inference is undefined (risking data leakage), the mask synthesis logic is self-contradictory, and the method lacks direct generation quality metrics (FID, LPIPS). While Reviewer #3 gave an Accept (5) and Reviewer #1 a Weak Accept (4), their reviews were overly brief and lacked depth, whereas Reviewer #2’s thorough analysis reveals fundamental flaws that cannot be resolved through rebuttal. The weak consensus and unresolved methodological defects lead to rejection.
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.
Based on the outcome of the rebuttal process, two reviewers assigned an “Accept” score, while one reviewer assigned a “Reject.” Reviewer 2’s (the “Reject” reviewer) feedback focused primarily on the assessment of generation quality and the perceived lack of novelty. However, the paper’s primary focus lies in data augmentation, rather than generation quality itself, and its effectiveness is demonstrated through the performance gains achieved in downstream tasks. Given that the overall sentiment of the reviews is positive, the recommendation for this paper leans toward “Accept.”
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.
two accept and one reject.
i read through and would accept subject to minor