Abstract

Pulmonary artery-vein separation is critical for clinical diagnosis and treatment planning. However, existing pixel- or voxel-based methods often pro-duce fragmented predictions, significantly reducing clinical confidence. To address above problems, we propose Graph-PAVNet, a graph structure learning framework designed for PA/PV separation. First, our Light Vessel Structured Modelling (LVSM) module constructs a topology-aware vascular graph by leveraging the inherent structural and semantic relationships within the vascular network. LVSM shifts from traditional voxel-level predictions to topology-based branch-level inference, effectively resolving prediction discontinuity. However, it is challenging for a single graph to do the separation task. Due to this issue, we propose the Modal Feature Sampling (MFS) mod-ule. MFS enriches node features by constructing a hybrid Real-Virtual(RV) feature matrix that integrates multi-source information. It also employs a dynamic feature weighting mechanism to achieve cross-modal complementarity, overcoming the challenges posed by modal discrepancies. For hierarchical inference, the Hierarchical Graph Attention Network (HGAT) stratifies nodes by vascular generation order (main to peripheral branches) and employs hierarchical masking to enforce structured inter-layer propagation. At last, we introduce a novel metric: Branch Misprediction Co-efficient (BMC) to better evaluate the clinical relevance and branch inconsistency. Experimental results show that our method outperforms existing approaches in both quantitative accuracy and clinical interpretability, offering a new paradigm for pulmonary artery-vein separation.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/3757_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/Lqy1018/Graph-PAVNet

Link to the Dataset(s)

N/A

BibTex

@InProceedings{LiQin_GraphPAVNet_MICCAI2025,
        author = { Li, Qingya and Yuan, Ye and Liu, Lu and Bao, Nan and Xu, Lisheng and Tan, Wenjun},
        title = { { Graph-PAVNet: A Graph-Based Learning Framework for Pulmonary Artery and Vein Separation Using Multimodal Feature Sampling } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15971},
        month = {September},
        page = {374 -- 383}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This paper presents Graph-PAVNet, an end-to-end graph-based learning framework for pulmonary artery and vein (PA/PV) separation. Unlike conventional voxel-wise models prone to prediction discontinuities, Graph-PAVNet transforms the segmentation task into a topology-aware node classification problem over a vascular graph. Specifically, it introduces: (1) the Light Vessel Structured Modelling (LVSM) module, which encodes the vascular skeleton as a simplified graph to preserve anatomical topology while reducing redundancy; (2) the Modal Feature Sampling (MFS) module, which constructs a hybrid real–virtual feature matrix integrating geometric, topological, and image-based attributes across four modalities; and (3) the Hierarchical Graph Attention Network (HGAT), which propagates features through vascular hierarchies using dynamic masking and hop-based gating mechanisms. A novel evaluation metric, Branch Misprediction Coefficient (BMC), is also proposed to measure topological errors. Extensive experiments on the ISICDM2020 dataset show that Graph-PAVNet achieves superior accuracy and structural consistency compared to multiple state-of-the-art models, offering a new paradigm for clinically reliable artery-vein separation.

  • 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. Topology-aware graph modeling: Traditional voxel-wise segmentation models often suffer from structural discontinuities, particularly in peripheral or low-contrast vessel regions, which can severely impact clinical reliability. Graph-PAVNet reframes the task as a node classification problem over a vascular graph, preserving the anatomical connectivity and hierarchical topology of the vessel tree. This structural modeling effectively may help address one of the core limitations of voxel-based approaches.
    2. Multimodal real–virtual feature: Graph simplification, while reducing redundancy, can lead to loss of local morphological detail. To address this, the proposed Modal Feature Sampling (MFS) module constructs four types of feature matrices based on real and virtual nodes/edges. By integrating geometric, topological, and image-based cues, the model enhances node-level semantics and improves its ability to discriminate between arteries and veins.
    3. Structural evaluation via BMC: Common evaluation metrics like Dice and Recall focus on voxel-level overlap and fail to capture structural errors such as branch disconnections. To better reflect clinically relevant topology, the authors introduce the Branch Misprediction Coefficient (BMC), which measures class inconsistency between adjacent nodes. This topology-sensitive metric aligns better with clinical concerns about vessel continuity and enhances the interpretability of model outputs.
  • 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. While Graph-PAVNet is compared with several segmentation baselines (e.g., nnUNet, Mamba-UNet), the paper does not include comparisons with recent graph-based or topology-preserving methods that are highly relevant to vascular structure analysis. Without such comparisons, it is difficult to assess how much of the observed improvement stems from the graph modeling itself versus other architectural factors.
    2. The method is only evaluated on a single dataset (ISICDM2020) with a relatively small number of cases. Given the anatomical variability and clinical complexity of pulmonary vessels, this limited dataset size raises concerns about the model’s generalizability. Additional validation on larger or external datasets would strengthen the claims.
    3. While the proposed method introduces a novel graph-based framework, the performance gains over baseline models are relatively modest. In particular, the results for pulmonary artery (PA) segmentation are not consistently better than standard models such as nnUNet and UNet, which raises questions about the method’s effectiveness in distinguishing arterial structures.
  • 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.

  • 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.

    (2) Reject — should be rejected, independent of rebuttal

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

    This paper presents a novel graph-based framework for pulmonary artery and vein separation and introduces several interesting modules, including hierarchical graph attention and multimodal feature sampling. However, the overall contribution is weakened by several important issues. The method is evaluated on a single, small-scale dataset, limiting confidence in its generalizability. Comparisons with more recent or topology-aware baselines are missing, which makes it difficult to contextualize the performance gains. Additionally, the quantitative improvements are limited, and in the case of pulmonary artery segmentation, the results are even worse than those of simpler models like nnUNet and UNet.

  • 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.
    1. Although the author emphasized in the reply that the focus of the work is the effectiveness of PA-V separation rather than the comparison of methods, the comparison with the existing related methods in the field can reflect the current level and performance baseline of the field, and only then can the advantages of the author’s work be evaluated.
    2. Regarding the data set and effectiveness, the author emphasized in the reply that a lightweight experimental design was adopted, but the method used by the author is a deep learning method, which is a data-driven method. Therefore, sufficient data volume is a prerequisite for ensuring the effectiveness of the method and the credibility of the experimental results. The author has very little data for each training and testing. The advantages shown in the final results may come from the similarity of the data and the statistical deviation of a small amount of data, rather than the method itself.



Review #2

  • Please describe the contribution of the paper

    The study presents a dedicated solution for pulmonary artery-vein separation by constructing graph-based vascular structures and employing graph convolutional networks (GCNs) to integrate topological and imaging information for node feature extraction and learning.

  • 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 and Novel Methodology: The authors propose an innovative graph classification approach combining geometric and spatial features through GCNs:

    Comprehensive consideration of:

    Inter-node relationships (bifurcation points)

    Adjacent point interactions within tracheal lumina demonstrating effective utilization of both global and local clinical information.

    Specific attention to the clinically significant, yet previously overlooked challenge of arterial-venous confusion in semantic segmentation.

  • 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. Dataset Limitations (Chapter 3): The evaluation using only 6 cases raises concerns about generalizability and robustness

    2. Unclear data partitioning in Section 3.1: 24 cases divided into 4 subgroups, and 4 subgroups used for training. This would lead to overlapping among the training, validation and test sets. It is recommended to ensure that there is no intersection between any two of the training, validation and test sets. It is recommended to ensure that there is no intersection between any two of the training, validation and test sets.

  • 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 does not provide sufficient information for reproducibility.

  • 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

    Implementation Details: For feature fusion in Section 2.2: The weighted addition approach requires comparative analysis with concatenation alternatives. Missing ablation study on fusion strategy selection

    Validation Requirements: Evaluation on larger datasets needed to ensure fairness. Must address dataset split issues to establish result credibility

  • 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?

    The methodology of this paper is clear and it has paid good attention to the clinical issues of vessel segmentation. However, there are problems in validation. The dataset used for validation by the authors is relatively small, which affects the generalization and robustness. At the same time, an extremely important point is that the authors must explain the issue of data partitioning, which affects the credibility of the current results.

  • 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 #3

  • Please describe the contribution of the paper

    The paper proposes a Graph-based learning framework capable of organically incorporating the topological structure of the vessel trees into the learning mechanism in a hierarchical fashion.

  • 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 author’s propose a promising alternative to current methods in order to incorporate topological information directly into the feature processing of the network. This is a very strong prior to incorporate into the architecture, and honestly a refreshing take on the otherwise concurrent blackbox DL approaches.

    The method is concise and clear. The results section is thorough.

  • 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 question that I don’t feel the paper touches upon is where to obtain a good initial volume to compute a the necessary skeleton for the architecture. I feel this is a chicken-egg problem of ‘I need a good vessel estimate to derive a good skeleton, but I need to solve the segmentation task before I can acquire a good vessel volume’.
    How does the model perform if the skeleton contains inaccuracies due to obtaining the skeleton from a fragmented/incomplete segmentation? Do we require a human annotator to annotate a skeleton before the proposed architecture can be employed?

    While I appreciate the extensiveness of Table 2. I would have liked to see the number of trainable parameters each model has in order to judge the parameter efficiency of each architecture. How does the proposed architecture compare in terms of number of parameters? I would assume the proposed approach poses a very strong inductive bias, allowing the architecture to perform better and with less data. Furthermore, graph operations are notoriously memory hungry in contrast to standard grid-based architecture. How does the proposed method compare in terms of memory consumption?

  • 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 does not mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • 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

    Minor formatting and grammar issues:

    • Throughout the paper there are multiple instances of missing spaces. Eg: Title of section 2.2 before ‘(MFS)’, Title of section 2.1 before ‘(LVSM)’, contributions section (2) before ‘(MFS)’, section 3.1 before ‘(Positive Predictive Value)’, etc., as well as before citations [] brackets.
    • Section 2.3 has a sigma equation with missing closing bracket.
    • Section 4, last sentence of first paragraph: “…Our next research direction is trying to hence the robustness of…”, I believe the authors meant to write something else than ‘hence’, perhaps ‘enhance’?
  • 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 proposed method opens up a new paradigm for designing topology-aware architectures. This is a novel and interesting idea that differs itself from the concurrent ‘black-box’ deep learning scaling trends.

  • Reviewer confidence

    Somewhat confident (2)

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

    Accept

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

    While in my review I asked for specific memory and parameter count information, these were not explicitly provided in the rebuttal. This however, does not change my opinion on the overall paper being well over the standard of acceptance.

    Furthermore, I feel the criticisms of the other reviewers to be at best irrelevant and at worse incorrect: -If the task at hand happened to be something akin to a classification task, I would agree a single dataset to be a limiting factor. However, this is a segmentation paper, and as such every piece of the image can be individually used to evaluate the method. -Reviewer 2’s second criticism involved the concern that train/val/test datasets may leak. I fail to see how this is a valid concern from the content provided on the paper. While is known to happen in research (and a competent research lab would never allow this to happen), I don’t see how any author could provide evidence to disprove this during a peer-review process. As such this appears a very unreasonable concern to me. -Reviewer 1 asks for topology preserving methods for comparison. I feel this concern is more of an ablation concern as opposed to a lack in results, as this work could also implement many of the literature’s (ex. topological losses) on top of their existing architecture.




Author Feedback

We sincerely thank the reviewers for their valuable feedback. Our work introduces Graph-PAVNet, the first end-to-end trainable graph neural network for pulmonary artery-vein (PA-V) separation, motivated by the clinical need to preserve vascular topology. By integrating topological priors into the architecture, we propose a clinically grounded, “structure-first” approach. Below, we address the major concerns:

Q1: Comparisons with recent graph-based or topology-preserving methods(For Reviewer 1) A1: We emphasize PA-V separation effectiveness over methodology comparisons for two reasons: (1) Pulmonary vessels are anatomically distinct from retinal/cerebrovascular systems, making cross-domain comparisons invalid due to poor generalization. (2) Due to the ‘single variable rule’, our vascular-specific graph framework (hierarchical attention/dynamic gating) makes direct comparisons with generic networks (GAT/GCN) invalid under unequal preprocessing conditions.

Q2: Limited Dataset—Concerns about generalization (For Reviewer1 & Reviewer 2) A2: In this work, our primary contribution lies in proposing a novel paradigm that bridges Euclidean features/data structures with non-Euclidean graph representations, thereby establishing a new learning framework for PA-V separation. The key point is not to build a fully optimized model, but to demonstrate the feasibility and potential of this approach in addressing prediction discontinuities in deep learning. Thus, we adopted a lightweight experimental design to validate our core hypothesis, prioritizing conceptual innovation over extensive generalization analysis.

Q3: Concerns about the effectiveness(For Reviewer 1) A3: Compared with the previous DL methods, this work is a methodological innovation from the actual clinical needs. The topological continuity of the predicted results is a priority condition that needs to be guaranteed more, as suggested by the doctors we collaborated with. Therefore, in this work, we propose a new evaluation metric BMC to evaluate the fracture problem. It can be found that on the premise of the classical index can be equal to the model, our proposed algorithm reduces the fracture cases by 83~98%. As an experimental test of a new technology path, we think this can show that our method has a relatively good effectiveness. And we believe there’s still space for improvement.

Q4: Unclear description of dataset division(For Reviewer 2) A4: We appreciate the opportunity to clarify our dataset splitting strategy. The description used in the manuscripts is “These CT scans are randomly split into 4 subgroups (4 sets for Training, 1 set for validation, 1 set for Testing).” It means 24 CT scans are randomly but equally divided into 4 subgroups. For each subgroup(containing 6 cases), 4 for training, 1 for validation, and 1 for testing. While different subgroup is independent of each other, there is no data leakage between phases. Thank you for prompting this important clarification.

Q5: Concerns about vascular skeleton extraction(For Reviewer 3) A5: This is quite a good question! For the work, we believe current pulmonary vessel segmentation achieves stable performance (~90% Dice), and when combined with conventional post-processing, tends to produce over-segmentation rather than under-segmentation. This is why we use skeleton as raw data structure: over-segmentation has minimal impact on skeleton continuity.

Q6: Concerns about parameters efficiency(For Reviewer 3) A6: As an experimental paradigm innovation, our design intentionally prioritizes lightweight architecture. Compared with the image, the amount of data in graphs can be reduced by more than 2 orders of magnitude. And we use LVSM Module to further reduce the size, which helps reduce the parameter amount of the overall algorithm. Based on this, we believe that the number of parameters in this work is still small, and both image task learning and general graph neural network are relatively lightweight.




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’

    I agree with the reviewers’ negative feedback. The paper lacks rigorous evaluation and omits essential comparisons with existing topology-aware or graph-based methods. I strongly disagree with R3’s overly positive assessment, which significantly overstates the paper’s contribution. Its practical value remains unclear. I recommend rejection.



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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



back to top