Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Coronary stenosis is a major risk factor for ischemic heart events leading to increased mortality, and medical treatments for this condition require meticulous, labor-intensive analysis. Coronary angiography provides critical visual cues for assessing stenosis, supporting clinicians in making informed decisions for diagnosis and treatment. Recent advances in deep learning have shown great potential for automated localization and severity measurement of stenosis. In real-world scenarios, however, the success of these competent approaches is often hindered by challenges such as limited labeled data and class imbalance. In this study, we propose a novel data augmentation approach that uses an inpainting method based on a diffusion model to generate realistic lesions, allowing user-guided control of severity. Extensive evaluations show that incorporating synthetic data during training enhances lesion detection and severity classification performance on both a large-scale in-house dataset and a public coronary angiography dataset. Furthermore, our approach maintains high detection and classification performance even when trained with limited data, highlighting its clinical importance in improving the assessment of stenosis severity and optimizing data utilization for more reliable decision support.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/3613_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)

https://github.com/medipixel/DiGDA

BibTex

@InProceedings{SeoSum_DiffusionBased_MICCAI2025,
        author = { Seo, Sumin and Lee, In Kyu and Kim, Hyun-Woo and Min, Jaesik and Jung, Chung-Hwan},
        title = { { Diffusion-Based User-Guided Data Augmentation for Coronary Stenosis Detection } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15967},
        month = {September},
        page = {151 -- 161}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The authors propose a novel data augmentation method that integrates an inpainting strategy with a generative model to synthesize coronary angiograms. The proposed data augmentation method can generate high-quality angiograms in a controllable manner based on user-defined percentage diameter stenosis.

  • 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 conditions that make sense for ControlNet. In particular, the condition of %DS can be defined by the user, making the data generation process more controllable and flexible.
    2. The re-annotated ARCADE validation set, which contributes to the further development of the prediction of coronary stenosis.
  • 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 descriptions of quantitative evaluation in section 3.3 are not clear. In Figure 4(a), the authors showed mAP50 metric for the experiemnts with different model size. What does the model size mean? Shouldn’t the model size be the same for all the experiemnts with different setting?
    2. Also in Section 3.3, the authors showed the mAP50 metric for the experiments with different data size, but why is the performance comparison over data size necessary?
    3. Lack of model (condition) ablation study. The authors only ablate the effect of different data setting, it is not clear the individual effect of conditions.
    4. Lack of comparison with other state-of-the-art synthetic data augmentation methods. It is not clear what the advantages of the proposed method are compared to the other similar works for data generation.
  • 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.

  • 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/2025/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 — could be rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    I recommend that the paper be rejected, but subject to rebuttal. The main concern is that the evaluation is not complete and confident enough.

  • 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 recommend accepting the paper. The author did a good job of responding to all points.



Review #2

  • Please describe the contribution of the paper

    The paper proposed a generative model based data augmentation approach that generate realistic lesions with user-guided control of severity.

  • 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 study proposed a novel data augmentation method to address limited annotation and imbalance data distribution problems in medical imaging. The conditional input of the generative model is desirable to obtain controllable synthetic image.

  • 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 limitation lies in that the generated lesion only has variation in diameter. The lesion length and other morphological factors have not been considered. Therefore, the diversity of generated images is somehow limited. There is a lack of comprehensive comparision experiment, comparing the proposed method with other SOTA detection/classificaiton models.

  • 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.

  • 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/2025/en/REVIEWER-GUIDELINES.html
    1. How were the text prompt processed and its effectiveness in the final performance remains unknown.
    2. The performance of the generative model (Multi-ControlNet）was not reported
    3. The symbols of the features should be marked on the corresponding places in the figure for easier understanding.
  • 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 — could be rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The study is of novelty in introducing inpainting based data augmentation; The experiments are not comprehensive and model implementation and module effectiveness needs further illustration.

  • 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

    The primary contribution of the manuscript is the development of a diffusion-based user-guided data augmentation pipeline for coronary stenosis detection and severity classification in coronary angiography (CAG) images. The pipeline integrates:

    A ControlNet-based diffusion model with dual conditioning (original image with lesion bounding box and vessel segmentation mask) to generate realistic synthetic angiograms with user-specified stenosis severity (%DS), addressing data scarcity and class imbalance. A one-stage YOLO-based detection and classification framework that localizes lesions (≥50% DS) and classifies severity (50-70% DS, ≥70% DS), trained on augmented datasets. A re-annotated public ARCADE dataset with precise lesion locations and %DS values, to be released for research. Validated on a large in-house dataset (7,894 images) and ARCADE (1,200 images), the approach improves detection (mAP50: 0.717 vs. 0.688 baseline) and classification performance, especially for severe stenosis, enhancing clinical decision-making in data-scarce settings.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    The use of a ControlNet-extended Stable Diffusion model with dual conditioning (bounding box and segmentation mask) for user-guided stenosis synthesis is innovative. Unlike prior augmentation methods (e.g., mask-based synthesis, ref 5), it allows precise severity control (%DS) via user-adjusted vessel contours, targeting specific regions to avoid artifacts. The pipeline uses quantitative coronary analysis (QCA) to extract vessel contours and MLD, allowing users to adjust control points to simulate desired %DS (e.g., ≥70% for severe cases). Synthetic images balance class distributions, enriching underrepresented severe stenosis cases. The method achieves a mAP50 of 0.717 (vs. 0.688 baseline) on the internal dataset and generalizes to ARCADE, with balanced datasets improving severe stenosis detection (Figure 4). The one-stage YOLO pipeline streamlines detection and classification, reducing computational burden. The re-annotated ARCADE dataset ensures clinical relevance. The evaluation uses a large internal dataset (7,894 images) and ARCADE (1,200 images), reporting mAP50 for detection and F1 for classification across synthetic data ratios (×1, ×2, ×4). Table 2 shows consistent gains (e.g., mAP50: 0.717 at ×4), and Figure 4 compares balanced vs. imbalanced datasets, highlighting severe stenosis improvements. External validation on ARCADE tests generalizability.

  • 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 YOLO-based one-stage detection and classification model, while efficient, leverages an established architecture (Jocher et al., 2024, ref 10). Similar single-stage pipelines exist in CAG (e.g., Faster R-CNN in Wu et al., 2020, Medical Image Analysis; YOLO in Zhang et al., 2021, IEEE TMI). The novelty lies in augmentation, not the detector. able 2 reports mAP50 gains (e.g., 0.688 to 0.717) and Figure 4 shows balanced dataset improvements, but no statistical tests (e.g., t-tests, Wilcoxon) confirm significance. Prior CAG studies (e.g., Yang et al., 2022, Radiology) use such tests for rigor. The baseline is a YOLO model without augmentation, but no comparisons with SOTA CAG methods (e.g., nnFormer, Zhou et al., 2022, MICCAI; DeepCoro, Avram et al., 2022, ref 2) or other augmentation techniques (e.g., GAN-based, Frid-Adar et al., 2018, ref 12) are provided. While %DS classification aligns with clinical thresholds (ref 16), no correlation with clinical outcomes (e.g., revascularization decisions, patient events) or QCA gold standards is provided, unlike robust CAG studies (e.g., Fischer et al., 2018, JACC). The claim of “optimizing decision support” lacks evidence of reduced diagnostic errors. ARCADE’s original annotations lack precise %DS, requiring QCA re-annotation. However, no inter-annotator agreement (e.g., Cohen’s kappa) or clinician validation details are provided, unlike rigorous CAG datasets (e.g., van Hamersvelt et al., 2019, European Radiology).

  • 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 has provided an anonymized link to the source code, dataset, or any other dependencies.

  • 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/2025/en/REVIEWER-GUIDELINES.html

    The Methods section’s diffusion model description (e.g., “zero-convolution layers”) is dense, needing simpler clinician-oriented explanations (e.g., “generates narrowed vessels”). The Experiments omit statistical significance, reducing clarity of performance claims (e.g., mAP50 gains). Baseline comparisons are limited to a no-augmentation YOLO, lacking SOTA context, which disrupts evaluation clarity. Minor repetition (e.g., class imbalance in Introduction and Experiments) could be streamlined for MICCAI’s space constraints.

  • 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 — could be accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The manuscript presents a compelling diffusion-based augmentation pipeline, significantly improving stenosis detection and classification by addressing data scarcity and class imbalance. Its clinical feasibility, robust evaluation, and open-source dataset make it a strong MICCAI candidate. Minor revisions are needed:

    Strengths: The ControlNet-based augmentation with user-guided severity control is novel, tackling CAG’s severe stenosis underrepresentation. Evaluation on large internal (7,894 images) and re-annotated ARCADE datasets shows clear gains (mAP50: 0.717), with balanced datasets enhancing critical cases. The open-source ARCADE dataset boosts research impact. Weaknesses: The YOLO-based detector is standard, lacking architectural novelty. Missing statistical tests and SOTA baselines weaken performance claims. Clinical validation (e.g., QCA correlation) and ARCADE annotation details are underdeveloped. Revisions Needed: Adding statistical tests, broader baselines (e.g., nnFormer), clinical correlations, and ARCADE validation details would address these, ensuring competitiveness. These are minor, as the core method and results are strong.

  • 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.

    Focus on downstream clinical utility (stenosis detection) is well explained.

    Differentiation from generative-quality-focused works (e.g., SegGuidedDiff) is valid and clearly articulated.

    The use of YOLO models, given the reproducibility and unavailability of newer coronary detection methods, is a practical and well-justified choice.

    References to current challenges in reproducibility in this domain further support their design decisions.

    Authors admit the lack of formal metrics but justify reliance on expert QCA-based annotation.

    They show awareness of the importance of reproducibility and propose adding model-clinician agreement analysis in future work.



Author Feedback

We thank the reviewers for their valuable feedback. The comments identified important areas for improvement, including evaluation of detection and generative models, ablation studies, and clearer interpretation of results (i.e. statistical analysis). We acknowledge these points and will incorporate the comments in the revised version. -Regarding experimental scope (R1, R2, R3), we focused on validating the clinical effectiveness of our proposed method in coronary stenosis detection, prioritizing practical downstream impact over generative quality evaluation. Although SegGuidedDiff (ref12) emphasizes high mask fidelity image synthesis (R2), our goal was to evaluate how controlled augmentation improves lesion detection performance. -For architectural choice (R1), we agree that comparisons with other coronary detection models (e.g., SSASS [Lee et al., 2023, MICCAI ARCADE Challenge]) could offer additional insights. However, recent coronary detection methods do not provide publicly available code or data, making direct comparison infeasible. In contrast, we selected YOLO models, which are reproducible and show strong performances on lesion detection tasks. -Regarding limited evaluation of generative model (R2, R3), one of the generative model metrics, FID, which is used to measure realism of real images, often fails to capture anatomical features in medical images as represented in ref12. In addition, although FID is suitable for typical image generation, it is less appropriate for our model, which focuses on generating rare samples to address class imbalance. -Regarding missing ablation on text prompt and model conditions (R2, R3), we followed previous study (Oh et al., 2024, MICCAI), which discovered providing more lesion-related textual information improves generation quality. -For limited generation diversity (R2), %DS is defined solely by lumen diameter, making diameter reduction the most clinically relevant factor. Unlike other domains where broader image diversity is important, our method focuses on targeted diameter adjustment. We do not apply uniform narrowing; instead, we introduce anisotropic, severity-controlled reductions to mimic real-world variation. This design ensures anatomical plausibility without altering non-lesion regions. -For the data scarcity experiment (R2), while we collected a large-scale coronary angiography dataset, many researchers only have access to limited-scale datasets. Thus, we show that our method retains performance under data scarcity settings, highlighting higher plausibility of practical use. -For unclear description of quantitative evaluation (R2), we used YOLOv11 under diverse setting for simulating various model complexity and parameter size and show ours consistently better than the baseline on medium/large model settings. However, as this figure can show mixed information, we will exclude model size ablation and instead emphasize performance gain for balanced augmented set for severe lesion. -For missing statistical analysis of experiments (R1), we used paired t-tests with bootstrapping to evaluate the performance gains of our proposed methods following Ribli et al. (2018, Sci. Rep). mAP50 improvements with the augmented dataset were significant in both in-house (p<0.01) and the public dataset (p<0.01), and the model trained with class-balanced augmented dataset significantly outperformed the model trained with the imbalanced augmented dataset (p<0.01). -We acknowledge the absence of inter-annotator agreement metrics (R1). Our %DS labels were annotated by an experienced clinician using quantitative coronary angiography (QCA), providing a reference for consistent labeling minimizing variability. We recognize the relevance of model-clinician agreement and inter-rater consistency (e.g., Cohen’s kappa). -For clinical correlation of clinical threshold and outcomes (R1), according to ACC guidelines (ref 16), %DS severity level is used to guide revascularization and additional investigation.



Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.

    N/A

  • 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    The use of diffusion models for data augmentation for coronary imaging may be a useful idea to present at the conference. Besides all reviewers seem to be in favor of the paper.



back to top