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

Abstract

Accurate coronary artery segmentation plays a critical role in coronary artery disease diagnosis and treatment planning. However, existing deep learning approaches are predominantly formulated as point-wise classification paradigms, where predictions are produced in a single forward pass without an explicit inference-stage structured update mechanism. This limitation becomes particularly evident in regions with ambiguous image evidence or weak structural support. We propose an Uncertainty-Guided Conservative Propagation (UGCP) framework that reformulates segmentation inference as a finite-step, structure-aware state evolution process in logit space via constrained neighborhood updates guided by conservative-inspired local interaction. The update is driven by local prediction discrepancies and modulated by voxel-wise uncertainty, allowing high-confidence regions to guide structurally unstable areas in a controlled manner. A source term anchors the update to the initial logits, preventing excessive drift during structural correction. We evaluate the proposed framework on convolutional and Transformer-based segmentation backbones using public 2D invasive coronary angiography (ICA) and 3D coronary CT angiography (CCTA) datasets (616 ICA images and 1,000 CCTA volumes). Experimental results demonstrate consistent improvements in the Dice similarity coefficient (DSC), clDice, and boundary metrics across architectures. The proposed framework enhances inference-stage structural consistency and shows potential for improving the robustness of coronary artery analysis. Our code is available at: https://github.com/chenzhao2023/UGCP.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3907_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/chenzhao2023/UGCP

Link to the Dataset(s)

ImageCAS dataset: https://github.com/XiaoweiXu/ImageCAS-A-Large-Scale-Dataset-and-Benchmark-for-Coronary-Artery-Segmentation-based-on-CT ICA dataset: https://github.com/MIILab-MTU/ICA_NJ_BinarySeg

BibTex

@InProceedings{HuaHua_UncertaintyGuided_MICCAI2026,
        author = { Huang, Huan AND Xu, Zhihui AND Esposito, Michele AND Zhao, Chen},
        title = { { Uncertainty-Guided Conservative Propagation for Coronary Artery Segmentation } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16884},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This paper focus on coronary artery segmentation based on CTA images, and primarily investigates the issue of structural instability in coronary artery segmentation. Moreover, the authors claimed that they “reformulates segmentation inference as a finite-step”. Rather than producing a segmentation result in a single step, it first generates an initial prediction and then refines it iteratively. As presented in abstract ” existing deep learning approaches are predominantly formulated as pointwise classification paradigms, where predictions are produced in a single forward pass without an explicit inference-stage structured update mechanism. ”.

  • 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 main strength of this paper lies in transforming the traditional one-time segmentation process into a gradually updated inference process. By introducing uncertainty information during the prediction phase to guide result refinement, allowing high-confidence regions to assist low-confidence regions, and incorporating a constraint to prevent errors from escalating during propagation, the method enhances the continuity and reliability of fine vascular structures while ensuring overall stability. The authors treat the segmentation problem as a process of gradual propagation and refinement differs from traditional methods and demonstrates a degree of innovation. Moreover, experiments show that in fine vascular regions, the results are more continuous and exhibit fewer breaks, indicating that the method is indeed effective for addressing structural challenges.

  • 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 notion of uncertainty in this paper is not new, and its definition is not clearly presented. For example, the symbol s is ambiguous. At the first line of Page 3, the s in W_s is not defined; only later, before Equation (1), do the authors define s_{(0)} as a linear projection of h, and refer to this as “the initial logit field. ” It is difficult to understand how the linear projection W relates to a logit field. Furthermore, around Equation (5), the authors introduce another projection f with W_f; however, the meaning of f is not explained. Regarding the “local spatial operator” U, as illustrated in Figure 1, s_p and s_q appear to interact, but in Equation (4) no iterative update scheme is specified. It is unclear what the symbols above s_p in Figure 1 represent. The term “conservative propagation” is also rather abstract; it is difficult to grasp at first reading and does not lend itself to an easy comparison with existing methods. The core idea of “conservative propagation” is therefore not sufficiently clear. In Section 1, the authors state: “In classical interaction systems, quantities evolve through local exchanges between neighboring states following structured update patterns. ” In Section 2.3, they further claim that “This interaction follows an in–out exchange structure, where the incoming contribution at one location corresponds to an outgoing component at its neighbors, yielding a locally conservative update in the absence of the source term. The source term introduces a controlled non-conservative component that anchors the evolution to the initial logits. ” From these descriptions, it seems that “conservation” refers to some form of local information exchange or update, but how this leads to a clear notion of “propagation” remains difficult to understand.

  • 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.

  • 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 method is relative novel, but need more clear illustration.

  • 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

    This paper presents the Uncertainty-Guided Conservative Propagation (UGCP) framework, whose core contributions can be summarized as follows:

    It reformulates the conventional single-shot point-wise classification as a finite-step discrete state evolution process in logit space, achieving neighborhood-coupled structural interaction through a local spatial operator.

    It proposes an uncertainty quantification method based on Evidential Deep Learning, designing directional gating to enable information propagation from high-confidence regions to low-confidence regions, while introducing an edge factor to suppress cross-boundary erroneous propagation.

    Furthermore, the framework is validated on two public datasets, namely 2D invasive coronary angiography and 3D coronary CT angiography, using two representative architectures: UNet and SwinUNetR.

  • 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.

    Unlike the single-shot forward propagation in conventional segmentation networks such as nnU-Net and SwinUNETR, UGCP views inference as an iterative refinement process (T=2 steps). This test-time refinement perspective complements the mainstream training-time optimization, particularly in handling semantic segmentation problems with slender topological structures like coronary arteries, where it can effectively alleviate branch discontinuities and spurious connections.

    Unlike existing methods that merely utilize uncertainty for training re-weighting or reliability estimation, UGCP embeds uncertainty estimation into the decision dynamics of the inference stage. Through Sigmoid gating, information flow control from high-certainty regions to low-certainty regions is achieved. This confidence-guided propagation mechanism possesses clear physiological significance in vascular segmentation, as the vessel centerline typically exhibits higher certainty compared to boundaries.

    The paper validates the method on 2D ICA images and 3D CCTA volumetric data, encompassing both convolutional (UNet) and Transformer (SwinUNetR) backbone networks.

  • 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.

    UGCP introduces T=2 rounds of iterative updates, each requiring computation of neighborhood interaction, uncertainty field, and edge factor. However, the paper does not report inference time or computational cost (FLOPs). For 3D CCTA data, the iterative process may significantly increase computational burden, which raises questions regarding the feasibility for clinical real-time or near real-time applications, such as interventional surgical navigation.

    The paper criticizes post-processing methods, such as morphological operations and conditional random fields (CRF), for being disconnected from the model prediction mechanism in the introduction, but the experimental section does not compare with such classical methods.

    UGCP involves multiple sensitive hyperparameters. The authors claim that these parameters are shared between 2D and 3D experiments, but the characteristics of 2D images and 3D volumetric data differ significantly, and this generalizability lacks justification.

  • 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.

  • 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 addresses the key challenge of topology preservation in coronary artery segmentation by proposing a physics-inspired, end-to-end trainable novel framework. The validation strategy across architectures and dimensions is rigorous and holds practical clinical significance. However, the absence of computational overhead analysis and insufficient comparison with classical post-processing methods limit the completeness of this work and its potential for clinical translation.

  • 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.

    Despite certain limitations acknowledged by the authors themselves, I believe this manuscript still holds potential for publication at MICCAI 2026.



Review #3

  • Please describe the contribution of the paper

    Segementation tasks are usually modeld as point wise classification task that predict the segmentation in one Infernce step. The authors propose a method that is using uncertainty guidance for to formulate segmentation inference as a finite-step discrete evolution in logit state space. The method is using uncertainty to guide information flow form high-confidence regions towards unstable areas.The method is evaluated on CNN and Transformer-based models for 2D ICA and 3D CCTA coronary artery segmentation.

  • 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 motivation and the idea of propagating information from certain to uncertain regions are clear.
    • Reformulating inference as an uncertainty-guided, finite-step discrete evolution in logit state space is an interesting and novel idea.
  • 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 performance improvements on both datasets are relatively modest, making it difficult to determine the practical significance of the proposed method.
    • At present, it remains unclear what the method adds beyond the existing literature on segmentation of tubular structures, for example in CCTA or ICA.
    • A comparison with other state-of-the-art methods is lacking.
    • The propagation depth is limited to T=2, which suggests a rather local effective receptive field. Therefore, the constrained neighborhood updates should primarily involve nearby voxels. This appears inconsistent with the examples in Figure 1, where the distance between certain and uncertain regions seems substantially larger.
    • It is not evident that uncertainty must be modeled specifically via a Dirichlet distribution. Experiments with alternative uncertainty measures, such as Monte Carlo dropout or BALD, would strengthen the methodological justification.
    • The sensitivity of the method to the hyperparameters u_0, τ, and step size θ is not investigated.
    • More detailed voxel-level analyses are needed to clarify how information propagates throughout the image. In particular, visualizing the iterative evolution of both the segmentation and the uncertainty map would substantially improve interpretability.
    • The quantitative evaluation should be extended to include topology-aware metrics, such as the number of disconnected components repaired, as well as topological false-positive and false-negative measures.
  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

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

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (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?
    • The performance improvements on both datasets are relatively modest, making it difficult to assess the practical significance of the proposed method.
    • A more detailed voxel-level analysis is needed to clarify how information propagates throughout the image. In particular, visualizing the iterative evolution of both the segmentation and the uncertainty map would substantially improve interpretability.
    • The influence of the propagation depth Tshould be analyzed more systematically to better understand its effect on performance.
  • Reviewer confidence

    Very confident (4)

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

    Reject

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

    The authors addressed all comments. However, a comparison with other approaches, such as CRF-based methods or segmentation methods for tubular structures, is still missing and remains critical. In addition, using only T=2, rather than T>2, does not allow information exchange beyond nearby neighboring voxels and therefore cannot close larger segmentation gaps, which is a critical limitation. The current setting mainly enables small local refinements. Given these remaining limitations, I cannot change my previous decision.



Author Feedback

We thank the reviewers for the careful evaluation. We agree that the gains are modest. We do not claim large SOTA-level gains; UGCP is proposed as a plug-in inference-stage update mechanism for thin-vessel segmentation. To avoid dataset/backbone-specific tuning, we used the same UGCP setting across 2D ICA, 3D CCTA, UNet, and SwinUNetR. Table 1 shows consistent DSC, clDice, and HD95 improvements under this fixed setting, supporting stability rather than over-tuning. clDice/HD95 provide structure-related evidence, while explicit topology FP/FN or component analysis would be useful complementary evidence. We agree that comparisons with CRF/morphology and other tubular methods would improve empirical completeness. UGCP is not intended to replace them, but to act at a different stage: existing methods often improve continuity through architecture, loss, prior, or post-processing, whereas UGCP contributes an uncertainty-guided logit-update mechanism during inference. Thus, it can serve as a complementary mechanism. We agree that full sensitivity analysis would better characterize the design space. We used a shared setting to reduce over-tuning, not to imply parameter insensitivity. T=2 was chosen to balance local correction, over-smoothing/error diffusion risk, and computational cost. Table 2 tests component contributions rather than all design choices; progressive gains from the UQ gate, edge factor, and source anchor suggest the improvement is not arbitrary smoothing. For uncertainty modeling, we do not claim Dirichlet is the only possible choice. We use it because it provides a single-pass, logit-coupled estimate: its concentration parameters encode both class proportion and evidence magnitude, allowing uncertainty to be recomputed after each update. MC dropout or BALD-style estimates usually require stochastic sampling or Bayesian approximation, less aligned with our finite-step update. We agree that the notation and “conservative propagation” were unclear. Here, p is a spatial location and q is its 4/6-neighbor; h_p is the backbone feature; W_s maps h_p to initial logit s_p^0; s_p^t is the logit after t updates; W_f maps h_p to feature f_p for edge estimation; and U is the local operator with uncertainty gates, edge modulation, and a source anchor. “Conservative” means an exchange-structured, conservative-inspired update, not a strict conservation law: without the source term, each location exchanges incoming/outgoing information with neighbors. “Propagation” means controlled local transfer of foreground-logit information, while the source term anchors states to initial logits to reduce drift. We agree that iterative masks and uncertainty maps would make the process more interpretable. Still, the submitted update rule explicitly specifies the voxel-level behavior: at each step, each voxel exchanges foreground-logit information only with its direct 4/6-neighbors; uncertainty controls exchange direction/strength, and the edge factor suppresses cross-boundary transfer. With T=2, this yields finite local refinement rather than long-range communication, and Fig. 1 illustrates flow direction rather than global propagation. We agree that 3D iterative overhead is important. Our submitted experiments focus on structural consistency and do not claim real-time deployment. UGCP adds no extra backbone and requires no repeated backbone inference; its extra computation is limited to local logit/feature-field operations. With fixed 4/6-neighborhoods and two steps, the overhead mainly scales linearly with voxels. Thus, the added cost comes from a bounded local update rather than another model-level inference. We thank reviewers for highlighting these limitations. We view them mainly as completeness and clarity issues, not as invalidating the mechanism. The submitted results support our restrained claim: UGCP is a backbone-independent inference-stage logit update improving structural consistency across datasets, dimensions, and backbones.



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 proposes an Uncertainty-Guided Conservative Propagation (UGCP) framework for coronary artery segmentation in CTA and angiography. However, several recurring concerns prevent stronger enthusiasm:

    Limited empirical gains: All reviewers note that performance improvements are modest, making it difficult to assess practical impact. Insufficient validation depth: Lack of strong comparisons with classical post-processing methods (e.g., CRF, morphology) and missing topology-aware or interpretability analyses. Weak ablation and sensitivity analysis: Hyperparameters (including propagation depth T=2) and uncertainty modeling choices are not sufficiently justified or explored. Computational concerns: Reviewer #2 raises concerns about efficiency and clinical feasibility due to iterative inference overhead, especially in 3D. Clarity and conceptual ambiguity: Reviewer #1 highlights notation and formulation ambiguities around “conservative propagation,” uncertainty definitions, and update mechanics. Methodological positioning: Reviewer #3 questions what the method adds beyond existing tubular structure segmentation literature, suggesting incremental rather than transformative novelty.

    On the positive side, reviewers consistently recognize:

    A clear conceptual motivation (uncertainty-guided propagation for topology preservation) A reasonable experimental design across 2D and 3D settings and multiple backbones A potentially useful direction for improving structural continuity in vascular segmentation

  • 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 author’s response addresses the reviewers’ concerns.



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 reviewers generally agree that the paper addresses an important problem in coronary artery segmentation, and the proposed method is interesting. The rebuttal sufficiently clarified most concerns. Some limitations remain, particularly the lack of comparison with CRF-based methods or segmentation methods for tubular structures, and the limited T=2 for larger segmentation gaps. I encourage the authors to discuss these limitations clearly in the camera-ready version.

    Overall, I believe this paper is valuable for acceptance at MICCAI.



Meta-review #3

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

    Reject

  • Please justify your recommendation.

    I agree with R3 that the limits of the method for small local refinements and lack of comparisons to classical (e.g. CRF) methods are still significant limitations of this paper.



back to top