Abstract

The SYNTAX score is a clinically established tool for assessing anatomical lesion complexity in coronary artery disease and guiding subsequent treatment. However, automated SYNTAX scoring is commonly formulated as a direct regression problem from coronary angiography videos to patient-level scores. In this work, we reformulate SYNTAX scoring as a vessel segment identity-preserving anatomical reasoning problem and propose a hierarchical modeling framework that explicitly aligns learning with the clinical workflow. Our approach maintains vessel segment identity across frames and views, estimates stenosis severity at the segment level, and aggregates evidence hierarchically according to coronary anatomy. Simultaneously, to address the scarcity of domain-specific data, we integrate and complete multiple public coronary angiography datasets, constructing a large-scale resource featuring completed vessel segmentation and derived structural annotations. Experiments demonstrate that vessel segment-level stenosis embedding enhances explanatory power and reduces prediction variability compared to baseline models, with the R^2 score improving by 0.201 and dev STD decreasing by 18.4%. These results highlight the necessity of structure-aligned modeling for reliable and stable automated SYNTAX scoring from multi-view coronary angiography videos. The GitHub link is https://github.com/VersaceSu7/SYNTAX_score_777.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5426_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: https://papers.miccai.org/miccai-2026/supp/5426_supp.zip

Link to the Code Repository

https://github.com/VersaceSu7/SYNTAX_score_777

Link to the Dataset(s)

CardioSYNTAX dataset: https://zenodo.org/records/14005818

BibTex

@InProceedings{FuSuz_Modeling_MICCAI2026,
        author = { Fu, Suzhong AND Dong, Jingqi AND Ding, Xuan AND Sun, Rui AND Yang, Yiming AND Cui, Shuguang AND Li, Zhen},
        title = { { Modeling Clinical Workflow for SYNTAX Scoring from Coronary Angiography Videos } },
        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

    This paper reformulates automated SYNTAX scoring as a vessel segment identity‑preserving anatomical reasoning task and proposes a hierarchical modeling framework aligned with clinical workflow. It integrates multiple public coronary angiography datasets to build a large‑scale annotated resource and achieves better prediction performance and stability than baseline methods

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

    1.It proposes a vessel segment identity-preserving hierarchical framework that aligns with clinical workflow for SYNTAX scoring. 2.It designs a coronary artery representation network using stenosis attention heatmaps and multi-view videos for accurate score prediction. 3.It integrates and completes multiple public coronary angiography datasets to provide large-scale structured annotations. 4.It enables more stable and interpretable SYNTAX predictions across heterogeneous multi-view data with improved 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.

    Insufficient experimental details are provided, which hinders result reproducibility

  • 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.

    (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?

    I recommend to accept this paper, as it successfully addresses the automated SYNTAX score prediction task with a clinically aligned hierarchical framework, delivers clear innovation and reliable performance, and provides a valuable dataset for the community.

  • 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.

    N/A



Review #2

  • Please describe the contribution of the paper

    The paper at hand describes a multi-stage pipeline to perform syntax scoring. First, a frame attention step selects frames of interest, then a vessel segmentation segments the coronary arteries, subsequently, the vessels are split into their subsegments, and bounding boxes are created for each vessel subsegment. Then for each subsegment the degree of stenosis is estimated and a stenosis attention heatmap is aggregated over all segments, this heatmap is then leveraged to filter the original image and fed into a vision encoder and some aggregation modules to end up with the final syntax score.

  • 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 whole pipeline is a smart way to capture the across view segment scores and is well motivated by the clinical workflow of the scoring. Moreover, the proposed workflow outperforms the previously reported SYNTAX scoring performance on the Cardiosyntax dataset.

  • 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 segmentation step does not look very robust looking at the videos provided, hence the value of the automatically generated segmentation masks is not that great. In general, with such a exhaustive multi-step pipeline it would be good to also evaluate the individual steps to capture potential butterfly effects and minimize potential failure cases. Moreover, I highly disagree with the choice of performance metric. For a regression task with clear clinically relevant thresholds it better captures the performance with respect to the clinical task at hand if the individual risk class-wise metrics are reported as well. However, I’ve seen that the original publication on the Cardiosyntax also uses this - in my opinion - not task-optimal metric.

  • 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?

    The paper presents a novel approach with the currently best reported performance on SYNTAX scoring. There are some major weaknesses, which however do not outweigh the positive sides for a conference paper.

  • 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.

    Still an accept from my side



Review #3

  • Please describe the contribution of the paper

    1.The article presents a hierarchical modeling approach that preserves the identity of vessel segments. This framework mimics the thought process of cardiologists when interpreting images, first examining local vessel segments and then summarizing them hierarchically into the left and right coronary artery systems. 2.Methodologically, this approach combines discrete vascular segments with continuous spatial features by generating narrow attention heatmaps. This ensures the consistency of the identity of the same anatomical vascular segment across different viewpoints and motion frames, reducing ambiguity during multi-view feature fusion. 3.Currently available multi-view coronary angiography datasets suffer from fragmented annotation. The authors integrated multiple public datasets and supplemented the vessel segmentation labels for approximately 1.2 million images, providing a valuable resource for subsequent research.

  • 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.This paper proposes a hierarchical modeling framework that preserves vessel segment identity. By extracting the vessel centerline and constructing a connected graph, this method decomposes the vessel tree topology into structurally stable linear vessel segments, allowing vessel identity to be anchored to the topological structure rather than to the pixel appearance, which is highly susceptible to changes in multiple viewpoints, cardiac motion, or contrast agents. 2.The study introduces a “segment-aware decision field” and a hierarchical feature aggregation mechanism, which aligns with the actual clinical workflow of cardiologists. Simultaneously, the model strictly groups and hierarchically summarizes multi-view video features according to the physical anatomical structure of the left and right coronary arteries, effectively preventing erroneous entanglement and confusion of features between different anatomical systems. 3.Regarding data utilization and domain resource development: Addressing the long-standing pain points in the field of coronary angiography multi-view video, such as the scarcity and fragmentation of high-quality labeled data, the authors innovatively integrated multiple publicly available datasets. They completed complete vessel segmentation labels for up to 1.2 million images, constructing and committing to open-sourcing this imaging dataset with extremely high clinical translational potential.

  • 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 experimental section of this paper only compares the proposed framework with the single baseline model CarS and its variants. If this study claims that its hierarchical aggregation and spatiotemporal processing are superior to traditional “black box” models, it should introduce more representative spatiotemporal video analysis baselines for comparison. For example, models such as CathAI’s automated coronary interpretation could be introduced, with slight modifications to their feature extraction backend to adapt them to the SYNTAX regression task, thus more convincingly demonstrating the superiority of this strategy. 2.When transforming discrete segment-level narrowing assessments into continuous spatial representations, the paper states that the Gaussian form is not optimal. Therefore, comparative experiments with different spatial kernel functions or continuous distribution strategies based on accurate lumen profiles should be added to the experimental section. Otherwise, the design of this core module will lack solid theoretical support and sufficient experimental verification. 3.When non-intersecting blood vessels overlap in a specific two-dimensional viewpoint, the algorithm inevitably generates “pseudo-bidding nodes.” This pseudo-topological structure incorrectly truncates what should be complete blood vessel segments, fundamentally altering the connected graph structure from that viewpoint. Furthermore, the segmentation diagrams shown in Figure 2 from different viewpoints are positioned relatively close together, failing to adequately illustrate the effect. Therefore, it is necessary to clearly explain how the current method identifies and eliminates “pseudo-bidding points,” and supplement this with relevant robustness analysis or ablation experiments, specifically demonstrating how it effectively avoids entanglement of erroneous anatomical features. 4.The mechanism for obtaining the key parameter $s_k$ appears to be a black box, as the paper doesn’t explain in detail how this crucial parameter is derived. If it’s automatically estimated by an algorithm, then specific implementation details and the accuracy of this step must be provided, as it directly determines the quality of the subsequent heatmap generation. 5.The keyframe extraction method is opaque. In Figure 2, the feature information for the entire framework comes from selecting keyframes from the case sequence. In real coronary angiography videos, many frames are invalid due to factors such as incomplete or dissipated contrast agent and angiography deviation caused by heartbeat. The paper does not explain how these frames were selected. It is recommended to supplement the specific algorithmic logic for keyframe selection. 6.The current evaluation is based solely on cross-validation using a mixture of these publicly available datasets. Given that the annotation and completion strategies for these datasets are designed by the researchers themselves, this introduces certain risks. The model must be rigorously validated on an independent external clinical cohort completely uninvolved in the training and label completion pipeline. Otherwise, the model may simply overfit to the self-designed pseudo-label completion logic and may not truly generalize to the complex and ever-changing real-world clinical environment.

  • 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.

    (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?

    While this research demonstrates significant value in modeling clinical workflows and integrating large-scale open-source datasets, its core methodologies and clinical validation still suffer from significant shortcomings. First, the model fails to account for the unavoidable vessel overlap issue in real-world 2D angiography. Second, the experimental section only compares the model with a single CarS baseline and its variants, failing to adequately demonstrate its superiority over other advanced spatiotemporal video analysis frameworks; furthermore, the core “segment-aware decision field” directly employs a Gaussian distribution without sufficient ablation experimental support. Finally, regarding clinical validation, it lacks completely independent external clinical cohort testing without label completion. Overall, the paper’s current status is slightly below the direct acceptance standard, but if the authors can provide compelling robust experiments, rich baseline comparisons, and clarify the aforementioned issues during the defense phase, this work still has the potential for acceptance.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    The authors’ rebuttal has addressed the remaining concerns raised during the initial review. Specifically, the authors clarified the frame-selection mechanism, explaining that while full sequences are shown in the supplementary video, adnd high-quality clips were used for training via a ResNet-34 classifier. They also justified the choice of evaluation metrics based on clinical guidelines and provided a clear distinction between their work and CathAI (which relies on 2D images rather than temporal video data).

    The promising performance on CardioSyntax, along with their commitment to open-sourcing the dataset (120 patients, 800+ videos), code, and trained weights upon acceptance, makes this a complete paper. I recommend accepting this paper.



Author Feedback

We sincerely thank the meta-reviewer and reviewers for their constructive comments. We are encouraged that R1 recognized our segment identity-preserving hierarchical framework, structured annotations, and improved performance, while R2 noted that our pipeline is clinically well motivated and achieves the best reported CardioSyntax performance. We address the main concerns below.

R2Q1: The videos provided in the supplementary material show the segmentation results of the complete videos. We did not remove low-quality frames, especially the first few frames where no vessels are visible, because we wanted to present the full video sequence. During training, however, we only used high-quality video clips.

R1Q1, R3Q5: We annotated high-quality video segments from all video-frame data of 120 patients. We then trained a binary classifier to distinguish high-quality frames from other frames. The training set contains more than 800 videos from 120 patients. Specifically, we trained a ResNet-34 model for this task. If the paper is accepted, we will release this dataset, the code, and the trained weights.

R2Q2: We strongly agree with this concern. However, public datasets currently lack full-pipeline annotations, which is also one of the main challenges in this field. Nevertheless, our method already includes several strategies to reduce error accumulation. These include iterative training of the segmentation model for better mask generation, a more stable frame-level segment tracking algorithm, and avoiding overlap when cropping vessel videos.

R2Q3: According to clinical guidelines, the SYNTAX score can indeed be divided into several risk levels. One major reason we used the metrics in the paper is that they reflect prediction reliability while also allowing a fair comparison with CardioSyntax. In future work, we will further promote and explore more suitable evaluation metrics.

R3Q1: To the best of our knowledge, CathAI is not a full-pipeline automatic coronary angiography interpretation model. It uses 2D images rather than temporal video data, and its final goal is stenosis detection and severity estimation. In addition, its code, weights, and dataset have not been publicly released, making it difficult to reproduce. Among current works on temporal coronary angiography analysis, a representative example is STQD-Det published in TPAMI, but it also does not release any files that would allow reproduction.

R3Q2: What we intended to express is that we do not claim the Gaussian distribution used in our method is the optimal choice. We used it mainly because it provides a continuous representation. This is a good research direction and problem. In future work, we will continue to explore more suitable representation methods.

R3Q3: In fact, the purpose of using bifurcation points for vessel segmentation and tracking is to generate stable video crops for each vessel segment. We remove video crops with high overlap. DeepCoro then outputs the stenosis severity for each input video crop. We map the stenosis severity of all cropped videos into heatmaps using Gaussian distributions. The heatmap represents all suspicious stenosis regions in the image. In our view, incorrect bifurcation points do not significantly affect the actual meaning of the heatmap. This is because, regardless of whether a vessel is mistakenly divided into multiple segments, if stenosis exists, it will still be reflected in the heatmap.

R3Q4: The extraction process of $s_k$ is not a black box. We follow Algorithm 1 in VesSAM, which performs vessel segments splitting based on the skeleton graph and bifurcation points.

R3Q6: We fully agree that external validation is very important. However, to the best of our knowledge, there is currently no other publicly available large-scale multi-view video coronary angiography dataset. This limitation is also one of the key motivations for our research in this field.




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.

    This paper addresses the important problem of automatic SYNTAX scoring for coronary arteries. The approach taken is interpretable breaking down the problem into artery segmentation and stenosis calculations. All reviewers have seen the merit but there are some concerns raised. Authors are invited to submit a strong rebuttal addressing these concerns.

  • 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.

    All three reviewers are on board for acceptance for this paper addressing an important problem of SYNTAX scoring. Many aspects of a good paper are present in this one.



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.

    There is consensus among the reviewers that the paper addresses an important clinical problem and proposes a clinically aligned framework. The rebuttal clarified the main concerns. I therefore recommend acceptance.



Meta-review #3

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    The initial reviews were generally positive but raised several substantive concerns. R1 supported acceptance and emphasized the clinical alignment, structured annotations, and performance gains, while noting that some experimental details were insufficient for reproducibility. R2 also recommended acceptance but raised concerns about segmentation robustness, error propagation in a multi-stage pipeline, and the lack of clinically risk-stratified metrics. R3 initially recommended weak rejection, mainly due to limited comparison with other spatiotemporal or angiography models, insufficient analysis of the Gaussian decision-field design, possible failure modes caused by vessel overlap or pseudo-bifurcation points, unclear frame selection, and the lack of an independent external clinical cohort. The rebuttal was useful in clarifying several points. It explained that the supplementary videos showed complete sequences, whereas training used selected high-quality clips identified with a ResNet-34 classifier; it also clarified the frame-selection procedure, the origin of the segment parameter through the VesSAM skeleton-graph splitting procedure, and the distinction between the proposed video-based SYNTAX scoring pipeline and CathAI-like stenosis-estimation systems. The rebuttal also explained the rationale for using the reported regression metrics for comparability with CardioSyntax and described measures intended to reduce error accumulation, including iterative segmentation refinement, segment tracking, and overlap removal during vessel cropping. These clarifications were sufficient for R3 to move from weak reject to accept, and R2 maintained an accept recommendation. Some limitations remain. The submitted paper still does not provide an independent external clinical validation cohort, and the absence of risk-class-wise clinical metrics limits assessment of clinical decision impact. The Gaussian form used for the decision field is acknowledged as a practical continuous representation rather than an optimized or fully validated design, and there is limited ablation of alternative kernels or lumen-profile-based spatial representations. The comparison set is also relatively narrow, relying mainly on CarS and variants rather than a broader set of adapted video or angiography baselines. These issues should be considered when interpreting the clinical readiness of the method. Overall, the strengths outweigh the remaining limitations for a MICCAI conference paper.



back to top