List of Papers Browse by Subject Areas Author List
Abstract
Accurate coronary artery segmentation from Digital Subtraction Angiography (DSA) is critical for computer-aided cardiovascular diagnosis and intervention.
In DSA, contrast wash-in provides characteristic temporal perfusion cues, i.e., vessels progressively enhance over time, which can help distinguish true vessels from dark non-vessel tubular structures and improve recognition under low-contrast conditions by comparing neighboring frames.
However, exploiting such cues is non-trivial, as cardiac motion introduces non-rigid vessel deformation and misalignment across frames.
Clinically, radiologists typically review the entire dynamic injection process, but measure only a single key frame where the vessel is maximally opacified.
Based on this observation, we present \textbf{MACOS}, a \textbf{M}otion-\textbf{A}ware \textbf{CO}ronary \textbf{S}egmentation framework that leverages long-range temporal injection context to supervise segmentation only on a key frame with maximal opacification.
Specifically, we first propose a Motion-Aware Gated Recurrent Unit (GRU) module to explicitly estimate inter-frame displacement fields to align features before aggregation, thus enabling robust spatiotemporal modeling under cardiac motion.
Then, to leverage unlabeled intermediate frames, we further introduce a Physics-Informed Vessel Loss to enforce monotonic vessel growth driven by contrast wash-in, for providing effective regularization without requiring dense annotations.
Experiments on 971 DSA sequences demonstrate that MACOS significantly outperforms state-of-the-art methods, improving robustness on typical challenging cases.
To the best of our knowledge, MACOS is the first framework to exploit long coronary DSA sequences for weakly-supervised vessel segmentation with only key-frame annotation, and is validated on the largest coronary DSA video dataset to date.
Code is available at \url{https://github.com/Fitz-Fitz/MACOS}.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0033_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Fitz-Fitz/MACOS
Link to the Dataset(s)
N/A
BibTex
@InProceedings{WuHan_WeaklySupervised_MICCAI2026,
author = { Wu, Han AND Xiong, Xiaosong AND Song, Yanli AND Zhan, Yiqiang AND Zhou, Xiang Sean AND Wu, Dijia AND Shen, Dinggang},
title = { { Weakly-Supervised Coronary Artery Segmentation from DSA Sequence via Motion-Aware Modeling } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16880},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper presents a weakly supervised framework for coronary artery segmentation in digital subtraction angiography (DSA). The method leverages temporal information from DSA sequences to improve segmentation performance without requiring additional annotations. Specifically, the framework uses preceding frames in the sequence to guide the segmentation of the final frame. The authors introduce a Motion-Aware Gated Recurrent Unit that models inter-frame misalignment caused by cardiac motion by estimating displacement fields, which are then used to align frames during segmentation. In addition, the method incorporates a physics-informed loss that penalizes the disappearance of vessel regions after alignment, thereby promoting more consistent vessel representation across frames. The contribution is supported by component-wise ablation studies and evaluation on a relatively large dataset of 971 angiography sequences, with comparisons against state-of-the-art methods demonstrating improved segmentation performance.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
- Clear: The paper is well structured, with a clear presentation of the motivation, method, and results.
-Methodological contribution: The proposed weakly supervised framework leverages temporal information in DSA sequences, with well-motivated components to address motion and vessel consistency without requiring additional annotations. It is evaluated on a large dataset, outperforms state-of-the-art approaches, and is supported by an ablation study demonstrating the contribution of its main components.
- 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.
-
Novelty positioning: The claim that this is the first work to explore long-sequence modeling for DSA vessel segmentation appears overstated, as Zhang et al. (2025, DOI: 10.1109/TMI. 2025.3540886) also proposed a spatio-temporal approach that leverages temporal information for this task. Authors should provide a discussion of this prior work and a more precise explanation of how the present method differs from it.
-
Limited discussion: Some parts of the paper would benefit from clearer explanation, particularly how the displacement fields are derived from the feature maps. Additionally, clarify why T=6 frames was selected, and include information on ethical approval.
-
- 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?
There is no information on the ethical review board for the dataset.
- 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
This paper is clear and well-structured. It introduces a weakly supervised framework for coronary artery segmentation in DSA, with a well-motivated use of temporal information to address motion and vessel consistency without requiring additional annotations. The evaluation is well described, using a large dataset, comparisons with state-of-the-art methods, and an ablation study that supports the contribution of the main components.
The paper would be strengthened by a more precise positioning of its novelty, particularly in relation to prior spatio-temporal work such as Zhang et al. (2025). In addition, some methodological details would benefit from clearer explanation, especially how the displacement fields are derived and why T=6 frames was selected. Including information on the ethical approval of the dataset would also be important for the completeness of the manuscript.
- 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?
The paper introduces a weakly supervised framework for coronary artery segmentation in DSA. The use of temporal information is well justified, effectively addressing motion artifacts and ensuring vessel consistency without the need for additional annotations. The approach is technically sound, and the evaluation is robust, featuring a large dataset, thorough comparisons with state-of-the-art methods, and an ablation study that demonstrates the value of the main components.
While the paper would benefit from a more precise positioning of its novelty relative to prior spatio-temporal work and a clearer explanation of some methodological details, these issues do not outweigh the overall strengths of the submission.
- 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
This paper presents a weakly supervised coronary artery segmentation framework for DSA sequences, where only the maximally opacified key frame is annotated while the full sequence is used for learning. The main contribution is a motion-aware long-sequence modeling framework (MACOS) that explicitly estimates inter-frame displacement fields for feature alignment before temporal aggregation, together with a physics-informed vessel loss that exploits monotonic vessel enhancement during contrast wash-in to regularize unlabeled intermediate frames. The method is validated on a relatively large DSA dataset and shows consistent improvements over strong single-frame and sequence-based baselines.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The paper addresses a clinically meaningful and practically relevant setting: in DSA, clinicians inspect the whole injection sequence but annotations are typically available only on a key frame. 2.The method is well motivated. Explicit motion compensation is appropriate for coronary DSA, where cardiac motion causes non-rigid deformation and simple temporal aggregation can blur vessel structures. 3.The physics-informed vessel loss is intuitive and tailored to the imaging process, leveraging contrast wash-in dynamics without requiring dense frame annotations. 4.The experimental study is reasonably comprehensive, including comparisons with CNN, Transformer, Mamba, and sequence-based baselines, as well as component ablations and temporal configuration analysis. 5.The reported results are strong and consistent across multiple metrics, and the qualitative examples suggest improved robustness on challenging low-contrast and motion-affected cases.
- 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 methodological novelty is moderate rather than truly substantial. The core design combines motion estimation/warping with recurrent temporal aggregation and a monotonicity-based regularization term; while effective, the individual components are conceptually incremental. 2.The physics-informed monotonic vessel growth assumption may not always strictly hold in real clinical DSA due to varying injection dynamics, local contrast propagation patterns, catheter/view changes, or severe motion/artifacts. The paper would benefit from a more careful discussion of failure cases and assumption violations. 3.Although the paper reports strong segmentation metrics, it provides limited analysis of the motion module itself. For example, there is no quantitative evaluation of alignment quality, no visualization of estimated motion fields in difficult cases, and no comparison with stronger motion-aware/video segmentation alternatives. 4.The experimental validation could be strengthened by reporting cross-center generalization more explicitly, clarifying whether data from the two centers differ substantially in acquisition protocol, and showing stratified performance or leave-one-center-out results. 5.Reproducibility is currently limited in the manuscript. Important implementation details such as exact sampling strategy, sequence normalization/preprocessing, model complexity/inference cost, and code availability are not yet sufficiently detailed for full 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 submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper addresses a clinically relevant and practically meaningful weakly supervised setting for coronary DSA segmentation. The proposed framework is technically sound, the motivation for motion-aware long-sequence modeling is clear, and the empirical results are consistently better than strong baselines. In particular, the gains over both single-frame and prior temporal models suggest that explicit motion compensation is useful in this setting.
My main reservations are about the level of novelty and the depth of validation. The technical contribution is somewhat incremental, and the paper would be stronger with more detailed analysis of failure cases, motion estimation quality, and cross-center generalization. Nevertheless, I find the work solid overall, and the combination of clinical relevance, method-task alignment, and strong experimental performance makes it worthy of acceptance.
- 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 paper proposes MACOS, which leverages long-range temporal information from complete DSA injection sequences for weakly supervised coronary artery segmentation. The core contributions include MA-GRU and a Physics-Informed Vessel Loss. The problem setting is clinically motivated and the experimental scale is relatively large.
- 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 temporal information during the contrast agent wash-in phase in DSA has indeed been largely overlooked by existing methods. Meanwhile, cardiac motion alignment represents a real and practical challenge, making the motivation well justified. 2.The dataset is relatively large in scale, which enhances the credibility of the results. 3.The experimental section employs a clear incremental ablation design, where components are progressively added, making the contributions of each module easy to interpret.
- 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 dice improvement of MACOS over nnUNet is only 1.40%, with also overlapping standard deviations. But there’s no statistical test (e.g. Wilcoxon test) provided in the paper, so it is hard to determine whether the improvement is statistically significant or simply due to randomness. 2.The current monotonicity assumption only holds for the wash-in phase. How should violations of this assumption be handled during the wash-out phase (e.g., at the end of the sequence)? 3.MA-GRU is essentially a combination of ConvGRU, optical flow estimation and feature warping. There is already substantial similar work in video object segmentation (e.g., DEVA, Cutie). The paper lacks review or discussion of related work and does not clarify the differences from general motion-compensation approaches in video segmentation. The monotonicity constraint in the Physics-Informed Vessel Loss is conceptually simple and lacks comparison with temporal consistency losses from other domains. Also, the motion estimation network consists of only three 3x3 convolutional layers, which may be insufficient for handling non-rigid cardiac motion. The work provides no quantitative evaluation or visualization of motion estimation quality. 4.There are writing issues and inconsistent terminology: Contribution 2 refers to Global Motion Loss + Physics-Informed Vessel Loss, and Fig. 1 caption labels (b) Global Motion Loss. However, Section 2.5 defines the total loss L_total using only L_mot and L_phy, and Global Motion Loss is never formally defined. It is unclear whether these terms are equivalent. 5.In Table 1, “SVS-Net [15]” actually corresponds to reference [5] (Hao et al., 2020), not [15] (Yang et al., 2019), which is a single-frame 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 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 proposes the MACOS framework for coronary artery segmentation from DSA, combining long-range temporal modeling with explicit motion compensation. The clinical motivation is well grounded, the dataset is among the largest for this task, and the ablation design is clear. These strengths support that the work has potential publication value. However, the main conclusions lack statistical testing, making it unclear whether the reported improvements are significant. Some technical details are insufficiently specified, although the authors state that the code will be released.
- Reviewer confidence
Somewhat confident (2)
- [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
N/A
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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 presents a weakly supervised framework for coronary artery segmentation in DSA that leverages temporal information and a physics-informed loss.
Reviewers agree the problem is clinically relevant and the proposed method is well-motivated and developed on a relatively large dataset, but note moderate novelty and missing validation details.
I suggest that the authors address the raised concerns, including 1) better position novelty against prior work, 2) provide statistical significance and stronger analysis, 3) clarify methods and fix inconsistencies, and 4) release code and models for reproducibility. Please also provide the necessary information on the ethical review board for the dataset.