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

Abstract

Advancements in 3D vision have increased the impact of blood vessel modeling on medical applications. However, accurately representing the complex geometry and topology of blood vessels remains a challenge due to their intricate branching patterns, curvatures, and irregular shapes. In this study, we propose a hierarchical part-based framework for 3D vessel generation that separates the global binary tree-like topology from local geometric details. Our approach proceeds in three stages: (1) key graph generation to model the overall hierarchical structure, (2) vessel segment generation conditioned on geometric properties, and (3) hierarchical vessel assembly by integrating the local segments according to the global key graph. We validate our framework on real-world datasets, demonstrating superior performance over existing methods in modeling complex vascular networks. This work marks the first successful application of a part-based generative approach for 3D vessel modeling, setting a new benchmark for vascular data generation. The code is available at: https://github.com/CybercatChen/PartVessel.git.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/2592_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/CybercatChen/PartVessel.git

Link to the Dataset(s)

ImageCAS: https://github.com/XiaoweiXu/ImageCAS-A-Large-Scale-Dataset-and-Benchmark-for-Coronary-Artery-Segmentation-based-on-CT Processed CoW: https://github.com/chinmay5/vessel_diffuse Vascusynth: https://vascusynth.cs.sfu.ca/Data.html

BibTex

@InProceedings{CheSiq_Hierarchical_MICCAI2025,
        author = { Chen, Siqi AND Zhang, Guoqing AND Lai, Jiahao AND Shen, Bingzhi AND Zhang, Sihong AND Dong, Caixia AND Chen, Xuejin AND Li, Yang},
        title = { { Hierarchical Part-based Generative Model for Realistic 3D Blood Vessel } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15962},
        month = {September},
        page = {257 -- 267}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    Their contributions are as follows: (1) The claimed first work to employ a part-based approach for 3D vessel generation. (2) Represent the global structure of vessels as a key graph and local segments as sequential curves, significantly enhancing vascular modeling detail. (3) Validate on real-world datasets, demonstrating superior performance over existing methods in modeling complex vascular networks.

  • 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.

    Decompose 3D vessels into global tree-like key graphs (via RVAE) and local geometric segments (Transformer-based VAE), enabling explicit modeling of branching topology and curved tubular details—a groundbreaking approach for irregular, hierarchical anatomical structures. Validated on three diverse datasets with geometric (Chamfer, JSD) and topological (Graph Wasserstein) metrics, outperforming state-of-the-art baselines in both quantitative scores and visual anatomical fidelity, ensuring robustness for complex vascular networks. Directly supports medical applications (e.g., surgical planning, segmentation) by generating anatomically consistent vessels, addressing a critical need for high-fidelity vascular modeling in precision medicine.

  • 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 paper does not conduct ablation experiments to isolate the impact of key components (e.g., RVAE vs. non-recursive encoders, Transformer architecture), making it difficult to quantify the contribution of each design choice. (2) The current visualizations of vascular reconstruction are limited to relatively straightforward cases. To further validate the framework’s versatility, it would be valuable to investigate its reconstruction performance on finely detailed smaller branches and complex topological configurations—such as ring structures—that characterize anatomically intricate vascular networks. (3) While the paper uses a Recursive Variational Autoencoder (RVAE), a method established in 3D shape modeling, the authors should clarify how their specific RVAE implementation—tailored to vascular topology and integrated into the hierarchical framework—introduces novel innovations. Key gaps include detailing unique node attribute designs, recursive decoding for vascular branching, or how RVAE is adapted to balance global structure and local geometry, beyond leveraging a mature methodology. (4) In the paper of TreeDiffusion, the Reconstruction results are much better than that presented in this paper. In additions, there are 5 datasets used in the paper of TreeDiffusion. Why do you use all the same datasets for experiments.

  • Please rate the clarity and organization of this paper

    Satisfactory

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

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

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

    N/A

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

    (3) Weak Reject — could be rejected, dependent on rebuttal

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

    (1) While the paper demonstrates methodological rigor through comprehensive comparative experiments across three distinct datasets - integrating both geometric measurements and topological analysis to achieve superior performance - critical visualization remain notably absent. (2) Although leveraging a well-established reconstruction framework that ensures technical validity, the study’s novelty requires clearer articulation. The authors should explicitly differentiate their methodological contributions from existing pipelines, particularly in addressing the identified limitations of current topological reconstruction approaches.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [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.

    Thanks for the rebuttal. However, I am not convinced by the rebuttal, which didn’t address my concerns about the study’s novelty and the comparison with TreeDiffusion.



Review #2

  • Please describe the contribution of the paper

    This work proposes a two-stage approach for vessel generation: first, modeling the global tree-like structure, including endpoints and bifurcations, and then generating the local geometry of each vessel segment as tubular curves with varying radii and lengths. A third postprocessing stage ensures seamless integration of the generated segments into a coherent structure.

  • 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. Novel two-stage approach: The paper introduces a novel methodology that separates vessel generation into two stages: first, modeling the global tree-like structure, and then generating the local geometry of each vessel segment. This decomposition is intuitive and effectively simplifies the problem, allowing for better control over both large-scale organization and fine-grained details.

    2. Clear and well-structured presentation: The writing and overall structure of the paper are of good quality, making the methodology and results easy to follow. However, there are some concerns that will be addressed in the corresponding section.

  • 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.

    Summary:

    1. Clarity and Explanation Issues
      • The role of Stage 3 (rotations and translations) is unclear, given that Stage 2 is already conditioned on the key graph.
      • Some notations and variables are not clearly defined or could cause confusion.
    2. Evaluation Concerns
      • The evaluation lacks standard generative metrics (such as COV, MMD, or 1-NNA) and does not measure novelty.
      • It is unclear how the metrics are computed—whether they compare generated samples to training data, use point cloud sampling, or extract skeletons from meshes.
    3. Training Details Missing
      • The paper does not specify training times for each stage.
      • It is unclear whether Stage 1 and Stage 2 are trained sequentially or jointly.
    4. Unsupported Claims
      • The statement “Our model achieves superior performance across all tasks.” is not justified based on the results discussion.

    Detailed comments:

    1. Regarding the approach, Stage 2 generates local geometry conditioned on the key graph produced in Stage 1. If this conditioning is already in place, why are rotations and translations necessary in Stage 3?

    2. On the generative aspect, Stage 1 generates a key graph, and Stage 2 generates local geometries. Since both are VAE-based, can Stage 2 generate multiple different local geometries for the same key graph?

    3. The descriptor C is first introduced in Stage 2, but it was already mentioned in Stage 1 without explanation.

    4. I have trouble with several symbols:
      • In the last part of the Encoding Phase section, shouldn’t it be h_z instead of z? Since h represents the encoded information, and z was already used for coordinates, this might cause confusion.
      • What does y represent in Equation 4?
      • What are x, l and q in eq 5? C is the descriptor? because a different C symbol was used
    5. Some parts are not clearly explained:
      • Stage 1 describes how to reconstruct a key graph, mentions the KL term, and appears as a VAE in the figure, but there is no discussion about sampling it seems as it is an autoencoder and not a VAE.
      • If i´m understanding correctly, In the description of Stage 2, the phrase “For each vessel segment identified in Stage 1” is misleading because Stage 1 does not explicitly identify segments. It would be more precise to say “For each vessel segment identified in the key graph,” as Stage 1 only outputs the original keygraph.
      • In Stage 2, for each node, coordinates (and other descriptors) are computed. What are the “above variables” that are used for conditioning?

    6. It would be helpful to clarify which data is used to compute the evaluation metrics. Are the generated samples compared against the training set? Are point clouds sampled from the generated meshes? Are skeletons extracted from the generated meshes?

    7. What is the basis for this evaluation? No other related work uses these specific metrics. The standard generative evaluation metrics (COV, MMD, 1-NNA) are not used, nor is novelty measured for the generated samples.

    8. I would suggest reorganizing the metrics table to distinguish between reconstruction and generation metrics. Additionally, please provide citations or a brief description of what each metric measures.

    9. How long does training take for each stage? Is Stage 1 trained first and then Stage 2, or is there an alternative training strategy?

    10. The statement “Our model achieves superior performance across all tasks.” is not accurate based on the discussion in the results section.
  • Please rate the clarity and organization of this paper

    Satisfactory

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

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

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

    1- Figure 1 is never mentioned in the text 2- In Figure 1, there is a histogram showing vessel length and the number of bifurcations. In the Implementation Details section, the authors mention that sequences are truncated to a maximum length of 200 (tokens?). How do the histograms look after this truncation? 3- Typo: I believe the word “captures” should be “capture” in the second paragraph of page 2. 4- The description of Stage 1 does not mention that RVAE has already been used for vascular modeling and generation.

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

    (4) Weak Accept — could be accepted, dependent on rebuttal

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

    The approach is novel and the results seem promissing, however major things need to be clarified

  • 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.

    I appreciate the clarifications



Review #3

  • Please describe the contribution of the paper

    This paper is the first proposal employing the part-based approach for 3D vessel generation. Unlike existing methods, authors consider both local and global geometric structures. Preserving global vascular structure, authors employ Recursive Variational Autoencoder (RVAE) which is extending Recursive Autoencoder (RAE) due to synthesizing key graph that representing proximal, bifurcation, distal point. Capturing local geometric structure, authors introduce transformer-based autoencoder with conditional variables that consider length, distance, curvature and depth.

  • 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 is the first work of part-based 3D vessel generation.
    2. Authors employ RVAE during existing methods adopt RAE and it is more reasonable for generation.
    3. To preserve the local geometric structure of realistic coronary artery, authors propose embedding conditional variables and it has novelty.
    4. The visualizations and figures are clear.
    5. The source code is made publicly available through an anonymized link.
  • 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.

    One major limitation of the paper is the absence of an ablation study.

  • 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.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

    It would be beneficial to include an experiment demonstrating how the generated vessels can be used to improve the performance of downstream tasks. This would further validate the practical utility of the proposed generative model beyond reconstruction metrics.

  • 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?

    This paper introduces the first part-based generative framework for 3D vessel modeling, which is a novel and impactful contribution to the field. The use of a Recursive Variational Autoencoder (RVAE) for modeling global vascular topology is a more appropriate generative choice compared to previous RAE-based methods. To capture local geometric details, the authors condition segment generation on vessel-specific descriptors, which adds to the model’s realism and novelty. While the visualizations are clear and the code is publicly available. Even the absence of an ablation study limits a deeper understanding of each module’s contribution, the main contributions are clear.

  • 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.

    Even the lack of ablation, the proposed approach and architecture are clear.



Author Feedback

We sincerely thank all reviewers for their constructive feedback. We appreciate the positive remarks on the novelty of our first part-based generative model for vessel generation(R1,R2,R3), the clarity of our presentation(R1,R2) and visualizations(R3), promissing result(R2,R3) and the code availability(R3).

[R1.1/R3: Ablation Study] We emphasize that the main contribution of our work lies in the overall part-based generative framework, rather than each single module. Nevertheless, we conducted ablations replacing RVAE with GraphVAE in Stage 1, and Transformer-VAE with a vanilla VAE in Stage 2. Both led to clear drops in performance. Detailed results will be included in follow-up work.

[R1.2: Result Visualization] First, our current work focuses on tree-like vascular structures, ring-like topology modeling is planned for future work. Second, it is important to note that the three datasets we used have already pose significant challenges due to their structural complexity and diversity. We will add visualizations with fine-grained branching details.

[R1.3: Clarify Novelty] Our RVAE is tailored to the anatomical structure of vessels. First, the key graph is inherently hierarchical tree-like structure, making it naturally compatible with recursive modeling. Second, we introduce vessel-specific node attributes, including segment orientation, curvature, length, and depth—that guide the latent representation to capture geometric properties. Furthermore, to our knowledge, we are the first work to apply a Transformer-based VAE for sequential vessel segment modeling, conditioned on the structural descriptors from the key graph. These contributions collectively underscore the novelty of our framework, which has also been positively recognized by R2 and R3.

[R1.4: Comparison with TreeDiffusion] We share two 3D datasets with TreeDiffusion, the remaining datasets are 2D and not applicable to our 3D framework. We additionally use ImageCAS for a more challenging evaluation. We follow its official code for evaluation and report real generated results. While they report stronger reconstruction, our method focuses on generation quality. Recently, we also compared with 3DShape2VecSet (SIGGRAPH 2023) and observed similar disconnections and isolated structures, highlighting the difficulty of vessel generation.

[R2.1: Clarity and Explanation Issues] We revised the manuscript to address all issues (detail comments 3/4/5/8/10 and additional 1/3/4) We deleted the sentence mentioned in 10, and added description for additional comments 4. In details, (1) Our method is similar to the assembly method. Stage 2 generates segments in a local canonical space, independent of the key graph, so it requires to apply rigid transformations. (2) Yes. (4) We clarify y=node degree; x=3D coordinates; l=sequence length; q=posterior; C=descriptor. Confused notation has been corrected. (5) Missing descriptions have been added.

[R2.2: Evaluation Concerns] (6) We evaluate all metrics on the test set, which comprises 10% of each dataset, with the remaining 90% used for training. For point-based metrics, we sample point clouds from the meshes, while for graph-based metrics, we evaluate on the assembled skeleton graphs. (7) The basis for our evaluation follows [24], which have been widely used for evaluating 3D shape generation. While we found they poorly reflect anatomical realism. Specially, we observed that compared methods with visibly unrealistic or fragmented outputs can still achieve deceptively high scores under such metrics. We will focus instead on domain-relevant metrics (vessel length, radius, bifurcation angles) in the future work.

[R2.3: Training Details] (9) Stage 1 and 2 are trained sequentially: ~12h and 3h respectively; assembly takes <3 min. (Additional 2) The 200-token limit does not distort the data distribution, is a validated hyperparameter.

[R3: Downstream Task] We agree on the value of downstream validation and plan to explore this in future.



Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

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

    N/A

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

    Accept

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

    N/A



Meta-review #2

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

    Reject

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

    N/A



Meta-review #3

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

    Accept

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

    N/A



back to top