List of Papers Browse by Subject Areas Author List
Abstract
Accurate coronary stenosis segmentation in X-ray angiography (XA) remains fundamentally limited by reliance on scarce and biased manual lesion annotations. Clinical datasets, although realistic, provide only incomplete and skewed sampling of pathological variations, constraining generalization. We address this challenge by reframing stenosis segmentation as a controllable synthetic learning problem and propose a physics-driven sim-to-real framework that eliminates dependence on real lesion labels. We generate a scalable synthetic dataset from patient-specific coronary models reconstructed from CTA, where stenosis variations with diverse severities and morphologies are deliberately designed in 3D space. A physics-based projection process produces geometry-preserving synthetic XA images with precise image-annotation alignment, enabling balanced and distribution-aware supervision beyond biased clinical sampling. A segmentation network trained on this synthetic data is subsequently adapted to real XA through uncertainty-guided self-supervised refinement. Predictive uncertainty is estimated online to select reliable pseudo-labels and drive confidence-aware consistency learning, promoting robust adaptation in ambiguous regions. On the ARCADE benchmark, our method surpasses existing self-supervised and few-shot approaches and approaches fully supervised performance without using any real stenosis annotations. These results demonstrate that physics-driven sim-to-real learning offers a principled pathway toward annotation-independent coronary lesion analysis.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2779_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/BaochangZhang/StenSim2realSeg
Link to the Dataset(s)
N/A
BibTex
@InProceedings{ZhaBao_Breaking_MICCAI2026,
author = { Zhang, Baochang AND Dewi, Fatia Kusuma AND Liu, Shuting AND Yousefi, Amir AND Schunkert, Heribert AND Ghotbi, Reza AND Navab, Nassir},
title = { { Breaking Annotation Dependency in Coronary Stenosis Segmentation via Physics-Driven Sim2Real Learning } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16893},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper addresses stenosis segmentation using a physics-driven sim-to-real framework. The overall pipeline consists of two stages. First, a physics-based simulation process generates synthetic X-ray angiography (XA) images together with corresponding ground-truth annotations. Second, a supervised segmentation network is trained on the synthetic data and further adapted to real XA images through uncertainty-guided self-supervised refinement. The results suggest the potential of simulation-based training for learning a stenosis segmentation model with reduced dependence on manual annotation.
- 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 introduces an uncertainty-guided adaptation strategy for real-domain refinement, which provides a feasible and practically relevant solution for sim-to-real learning in XA stenosis segmentation.
- Using CTA to synthesize XA images and generate guidance for stenosis analysis is an interesting idea.
- The main contribution beyond the baseline XA-Sim2Real framework is reasonably supported by the experiments and ablation studies. In particular, the results indicate that the proposed uncertainty-guided learning strategy improves adaptation performance.
- 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 goal of the sim-based method is to increase coverage of lesion types in the existing dataset. However, the lesion-generating algorithm does not account for lesion heterogeneity in the real world as its input. Purely randomized samples may not represent the actual lesions and may introduce bias.
- In Figure 2, the differences between methods are unclear. Figure 2A should indicate the true and false detections for each method.
- It is not described how the domain transfer model used in this study is trained. If the domain transfer model is also trained based on the same data,
- The signal intensity of the vessels are associated to the hemodynamic and dose of the contrast agents, which is not considered in the model, making the approach not as pathology specific as it claims.
- The DRR operator mentioned in the Methods section is not sufficiently described. Its formulation and implementation remain unclear.
- The implementation details are insufficient, especially for the synthetic data generation process. More detail is needed to understand how the patient-specific models, stenosis variations, projection settings, and annotation generation are constructed.
- Dice score is not a good measure for stenosis detection.
- The around 0.5 Ins-F1 is low for a detection task
- Lower Ins-F1 is presented in the proposed model with substantially higher Ins-Recall. This indicates that the proposed method provides much greater detection without sacrificing precision. This needs to be thoroughly discussed.
- Statistical significances are not tested between methods.
- 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
Minor points: CCTA should be spelled out when it is first mentioned.
- 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?
Overall, the paper presents an interesting sim-to-real framework for stenosis segmentation and addresses an important practical problem in medical image analysis. However, Figure 2 raises a concern that the proposed method predicts multiple stenotic regions that are not present in either the original ground truth or the authors’ re-annotated ground truth. This suggests possible false positives or over-segmentation of stenosis-like regions.
- 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
Review #2
- Please describe the contribution of the paper
1) Systematically generate synthetic XA data from CTA under controlling coronary stenosis by parametrical definition. 2) With synthetic data guidance, the gap between sim to real is corrected by self-supervised learning refinement, therefore, the annotation efforts on the real world data is mitigated.
- 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) integration of physical simulation for synthetic data generation, based on existing CTA dataset, the annotation efforts of XA datasets are reduced. 2) SSL provides a possible direction to mitigate the sim2real gap, where the real annotation is missing for detection/segmentation task 3) detailed experiment with reasonable accuracy improment
- 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 re-annotation motivation is not clear. Need to compare the differences between original annotation and re-annotation to show the reason. Otherwise, the motivation is more likely to improve the proposed method performance on this re-annotated data 2) The std of performance is huge based on 300 testing sample. No confidence or statistical difference analysis conducted
- 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 title might not be good, as the physical driven is based on simulation but not real physical driven DL models. The current one can mislead the reader
- 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?
Good method based on physical simulation, reduced the training efforts of synthetic data generation
- 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.
Accept, I like the idea and feel it is right direction to go
Review #3
- Please describe the contribution of the paper
The paper proposes to create a patient specific synthetic dataset for coronary stenosis segmentation. They used a specific coronary model reconstructed from CTA.
- 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.
Physics-based generation of data. Performance comparable to supervised methods but with synthetic data. Code and synthetic data will be shared.
- 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.
I would have liked a dedicated state-of-the art/related work section.
- 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
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?
This paper is well-structured and proposes a novel method that outperforms current approaches. However, I am not much familiar with this field.
- Reviewer confidence
Somewhat confident (2)
- [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.
N/A
Author Feedback
We thank the reviewers for the constructive feedback and positive assessment of the proposed sim-to-real framework for stenosis segmentation.
1.Re-annotation Motivation(R2). As described in Sec.2.4 and Sec.3.1, the proposed framework simulates stenoses spanning 30%-100% severity, while the original ARCADE annotations only include stenoses with ≥50% luminal narrowing in vessels of diameter ≥1.5 mm. The re-annotation was therefore introduced to provide a more comprehensive evaluation for mild stenotic regions rather than to favor the proposed framework. Importantly, Tab.1 shows consistent improvements under both the original and re-annotated settings, indicating that the re-annotation does not alter the relative performance trends among methods.
2.Figure 2 Clarification(R1). We agree that Fig.2 can be clearer. In the revised figure, we will use the re-annotated mask as the reference and explicitly mark true-positive and false-positive regions for each method. The additional predictions of our method are primarily confined to the vessel lumen, which is further supported by the uncertainty maps in Fig.2B showing localized responses around narrowing-like regions rather than non-vessel areas. These responses may highlight vessel regions requiring further clinical inspection rather than arbitrary over-segmentation.
3.Synthetic Generation Clarification(R1). The proposed simulation framework is designed to provide controllable stenosis diversity for supervision generation. Although randomized sampling is used to increase diversity, stenosis generation remains constrained by CTA-derived patient-specific coronary anatomy with controlled lesion severity and morphology. We agree that full physiological modeling of hemodynamics and contrast dynamics is beyond the current scope. Here, “pathology-aware” refers to anatomically meaningful stenosis appearance and narrowing patterns, while XA appearance is further approximated through DRR rendering and domain transfer. The DRR operator follows the physics-based implementation adopted in DeepDRR [14], including attenuation-based projection for vessels and contrast agents. The CUT-based domain transfer model is trained in an unpaired manner using generated DRR images and unlabeled XA images from the public XCAD dataset, which is independent from the stenosis segmentation dataset used in this work. Additional implementation details will be clarified in the revised manuscript and released code.
4.Evaluation Metrics and Statistical Analysis(R1,2). We agree that Dice alone is insufficient for stenosis segmentation evaluation; therefore, we additionally reported instance-level metrics including Ins-F1 and Ins-Recall. Ins-F1 is computed under a strict lesion-wise mask matching criterion, where small overlap differences can strongly affect instance matching in challenging XA stenosis segmentation. Notably, all methods in Tab.1 achieve relatively moderate Ins-F1 values, reflecting the intrinsic difficulty of stenosis segmentation in XA images. Since XA-Sim2Real is the most comparable annotation-free baseline trained using the same synthetic dataset, paired Wilcoxon signed-rank analysis based on the reported test results further shows significant improvements in Dice (0.515 vs. 0.486, p=0.0129), Recall (0.635 vs. 0.457, p<1e-49), and Ins-Recall (0.843 vs. 0.645, p<1e-20), while no significant difference is observed for Ins-F1 (0.570 vs. 0.578, p=0.195) under the re-annotated setting. The improved Recall and Ins-Recall indicate enhanced sensitivity to suspicious stenotic regions, while the relatively large standard deviations further reflect the challenging nature of stenosis segmentation in low-contrast XA images.
5.Minor Revisions(R1,2,3). We will further revise the manuscript to improve abbreviation consistency, title clarity, and the discussion and positioning of related 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 manuscript was reviewed by three expert reviewers. While acknowledging merits in the proposed work, the reviewers also raised some concerns about methodological novelty and validation. Such concerns should be addressed to strengthen the value of the work.
- 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 reviewers’ concerns have been successfully addressed in the rebuttal, making the work a solid contribution.
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 authors mostly addressed the reviewers’ concerns. The reviewers found the work interesting.
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.
This paper presents a physics-driven sim-to-real framework for coronary stenosis segmentation that substantially reduces dependence on manual XA annotations through synthetic data generation and uncertainty-guided adaptation. Reviewers highlighted the novelty of combining patient-specific CTA-based simulation with self-supervised refinement, as well as the potential clinical value of reducing annotation burden in a challenging task.
The main concerns focused on synthetic data realism, re-annotation rationale, implementation details, and statistical validation. The rebuttal provided convincing clarification regarding re-annotation, simulation pipeline, independent domain adaptation strategy, and statistical significance analyses. While limitations remain regarding physiological realism and broader validation, the overall methodology is well-motivated, supported by comprehensive experiments, and demonstrates meaningful improvements over the most relevant annotation-free baseline. Overall, the paper offers a promising and practically relevant direction for annotation-efficient coronary stenosis analysis.