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

Abstract

Accurate quantitative cerebrovascular assessment and clinical intervention rely fundamentally on vessel parsing in Digital Subtraction Angiography (DSA) to segment and distinctly label the vascular tree. However, the intricate vascular topology, pronounced vessel overlap, and inconsistent contrast enhancement inherent in DSA present formidable obstacles; consequently, most existing works have been limited to binary vessel segmentation, overlooking the fine-grained anatomical categorization required for detailed clinical analysis. To address this, we propose AngioParse, a multi-expert driven framework explicitly designed for fine-grained vascular parsing. AngioParse ensures parsing precision through three key mechanisms: (1) a dual-stream expert synergy module that improves localization accuracy by fusing a high-recall global binary prior with fine-grained anatomical features from heterogeneous experts; (2) a self-consistent structural grounding mechanism that regularizes fine-grained parsing with a high-recall binary vascular prior by constraining the union of anatomical-wise predictions to lie within the vessel support, improving connectivity and topological validity; and (3) a decoupled orthogonal prediction module that enhances class discrimination by separating multi-target learning into parallel branches to minimize feature confusion. Experiments on a public DSA dataset demonstrate that AngioParse not only achieves state-of-the-art segmentation performance but also exhibits superior capability in accurately parsing and differentiating six distinct vascular categories.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3344_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

N/A

Link to the Dataset(s)

N/A

BibTex

@InProceedings{ZhuKai_FineGrained_MICCAI2026,
        author = { Zhu, Kai AND Cao, Le AND Chen, Li AND Cheng, Jun AND Mou, Lei AND Zhao, Yitian},
        title = { { Fine-Grained Cerebrovascular Parsing in DSA via Structurally-Grounded Semantic Disentanglement } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16893},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This paper proposes AngioParse, a structure-aware framework for fine-grained cerebrovascular parsing in DSA. The method combines a binary vessel prior with multi-branch class-wise predictions via two main components: self-consistent structural grounding (SCSG) and orthogonal semantic disentanglement (OSD). The framework further integrates CNN-based local features with SAM-based global representations.

    The problem of moving from binary vessel segmentation to fine-grained anatomical parsing is important and clinically relevant. The paper presents a reasonably engineered pipeline and demonstrates improved performance over several 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 transition from binary vessel segmentation to fine-grained anatomical parsing is meaningful for downstream clinical analysis. 2.The idea of combining a global vessel prior with class-wise predictions via consistency constraints is intuitive and well-motivated. 3.The method shows consistent gains over baseline models, and the ablation studies provide some support for each component.

  • 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 claims “semantic disentanglement”, and Fig. 2 suggests feature-level separation (“redundant vs diverse features”). However, the orthogonality loss is applied on output probability maps A_i rather than intermediate feature representations. 2.The paper does not clearly describe how the final segmentation is produced:

    • The framework concatenates M_{prior} and A_i, but the fusion mechanism is not specified
    • It is unclear how conflicts between branches are resolved
    • The role of P_{ens} is limited to loss computation and not used in inference 3.The paper describes the model as “MoE-driven”, but:
    • There is no gating or routing mechanism
    • The model simply concatenates features from multiple branches 4.Although the method targets fine-grained parsing, the merging of CA and VA into a single category in quantitative evaluation, along with the lack of clear visualization of VA in qualitative results, raises concerns about the model’s ability to distinguish anatomically similar but distinct structures.

    5.The role of SAM is not fully justified. It is unclear whether it contributes meaningful anatomical reasoning or simply acts as a pretrained backbone.

    1. The necessity of certain design choices (e. g. , dual decoders for SAM features, final concatenation with M_{prior}) is not sufficiently validated. 7.The clarity of the framework diagram (Fig. 2) could be further improved, as several intermediate steps are either omitted or not fully consistent with the textual description. Missing intermediate operations: The transition from intermediate features (e. g. , F_1 \in \mathbb{R}^{32\times H \times W}) to the final probability map M_{prior} is not explicitly illustrated. According to the formulation, M_{prior} is obtained via a sigmoid activation, but this step is not clearly depicted in the diagram.
      • Ambiguity in feature representation: The text distinguishes between local features (CNN) and global features (SAM-based). However, this distinction is not clearly reflected in the diagram, where multiple branches (e. g. , F_2, F_3, F_4) are presented without explicitly indicating how “local” versus “global” information is encoded or preserved. 8.The definition of the baseline in Table 2 and Table 3 is not entirely clear. It appears that the baseline is obtained by removing certain components from the full model rather than using a standard architecture, but the exact configuration is not explicitly described.
  • 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.

  • 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 evaluation does not fully support the claim of fine-grained parsing, as CA and VA are merged into a single category and VA is not clearly visualized in qualitative results. The role of the SAM branch is also unclear, with limited evidence that it contributes meaningful anatomical reasoning beyond serving as a pretrained backbone. Furthermore, several design choices (e.g., dual decoders, final concatenation) are not sufficiently validated.

  • 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.

    Thank you for the authors’ response. After checking the manuscript, I appreciate several clarifications, especially the willingness to revise the “MoE-driven” terminology and to provide disaggregated CA/VA results. However, several major concerns remain unresolved. 1.The manuscript substantially overstates the disentanglement claim. The contribution section states that the method enforces feature orthogonality in the latent space, while the actual orthogonality loss in Eq. (7) is applied to the flattened output probability maps A_j and A_k, not to intermediate feature representations. Therefore, the method demonstrates prediction-level diversity at most, rather than feature-level or semantic disentanglement. 2.The final inference mechanism remains unclear. The manuscript defines M_{prior}, the anatomy-wise sigmoid outputs A_i, and the noisy-OR ensemble P_{ens}, but it does not provide a precise formula for the final multi-class segmentation. 3.The evidence for the SAM branch remains insufficient. The manuscript motivates SAM2 as a source of long-range structural dependencies, but the current experiments do not isolate whether SAM contributes task-specific anatomical reasoning or simply acts as a strong pretrained encoder. 4.The definitions of the baselines in Table 2 and Table 3 should be made explicit. In particular, Table 3 reports a non-trivial result even when all three listed decoders are marked as removed, making the actual baseline architecture ambiguous. Overall, the response addresses some presentation issues, but the main concerns about overstated disentanglement, unclear inference/fusion, insufficient SAM-specific validation, and incomplete fine-grained evaluation remain only partially resolved.



Review #2

  • Please describe the contribution of the paper

    This paper presents a novel framework called AngioParse for fine-grained anatomical parsing of the cerebrovascular tree in DSA. The proposed SCSG and OSD, effectively address the challenges of projective overlap and vessel fragmentation. Extensive experiments comprehensively demonstrate the effectiveness of the proposed method.

  • 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 manuscript is well-written, and the figures and tables are clear and easy to interpret. The research topic is highly valuable and clinically meaningful. The experimental evaluation is thorough and comprehensive.

  • 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.As a study focused on the vessel parsing task, the proposed method relies heavily on high-quality, pixel-level annotations of DSA images. The practical application value and clinical impact of this work would be significantly strengthened if the authors could conduct cross-dataset testing to explicitly demonstrate the model’s generalization ability across different data sources. 2.The design of the Orthogonal Semantic Disentanglement (OSD) module relies on the assumption that the representations from different segmentation heads are orthogonal and uncorrelated. However, in reality, distinct anatomical vascular structures maintain inherent correlations and physical connectivity (e.g., CA connects directly with the PCA and MCA). Therefore, forcing these anatomical structures to be strictly orthogonal in the feature space seems somewhat counterintuitive and potentially unreasonable. The authors should justify this design choice or discuss its potential limitations regarding physically connected vessels.

  • 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 mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure 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?

    This paper proposes a promising solution for the fine-grained anatomical parsing of the cerebrovascular tree in DSA, which holds significant practical application value. Although there are some debatable aspects in the methodological design rationale, the manuscript is clearly written, and the experimental evaluation is comprehensive.

  • 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.

    The authors have thoroughly addressed all my concerns in the rebuttal. Given that the research topic is highly valuable and clinically meaningful for vessel segmentation, I recommend accepting this manuscript for publication.



Review #3

  • Please describe the contribution of the paper

    The paper proposes AngioParse, a framework that bridges the gap between binary vessel extraction and fine-grained anatomical parsing by combining global morphological cues with local anatomical signatures to categorize the cerebrovascular tree into clinically distinct segments. It further introduces a self-consistent structural grounding strategy to preserve topological continuity and an orthogonal semantic disentanglement module to resolve projective ambiguity in 2D DSA, improving the accuracy and reliability of anatomical labeling.

  • 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 paper proposes a dual-stream expert synergy module and a self-consistent structural grounding mechanism to address the challenges of complex vascular topology and vessel overlap. The experimental evaluation is relatively comprehensive, and the comparisons and analyses are clearly presented.

  • 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 motivation for using SAM as one branch is unclear. The authors state only that it is for global encoding, but this capability could also be achieved with ViT. 2.The authors design Lcons to enforce anatomical correctness. However, it is unclear why Mprior is not directly used as a multiplicative mask to suppress irrelevant regions. 3.The claim that DgCNN serves as a global encoder lacks sufficient justification. The experiments in Figure 5 do not clearly demonstrate its global modeling capability. 4.The hyperparameter analysis in Figure 4 appears to be conducted on the test set, which violates standard experimental protocols. 5.The performance of Swin-UNet in the comparison is unexpectedly low, and the authors do not provide a corresponding explanation.

  • 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 mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure 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 proposes AngioParse, a multi-expert driven framework for fine-grained vascular parsing, but the design lacks sufficient justification, and several experimental details remain unresolved.

  • Reviewer confidence

    Confident but not absolutely certain (3)

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

    Accept

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

    The authors’ rebuttal has addressed part of my previous concerns, including the motivation of SAM and the design of experimental schemes. Although the paper still has minor deficiencies in details and experiments, such shortcomings are acceptable given the page length limit. Overall, its merits outweigh the flaws, and it meets the acceptance criteria of MICCAI.



Author Feedback

We thank the reviewers for their constructive feedback. Our work’s clinical value (R1/R2), SCSG/OSD effectiveness (R2/R4), and thorough experiments (R1/R2/R4) were recognized. Our responses are as follows: 1.Disentanglement & Orthogonality (R1, R2) Prediction-Level Disentanglement: L_ortho is applied to output probability maps A_i, not intermediate features, to supervise each head’s response uniqueness. Penalizing redundant co-activation forces preceding layers to learn discriminative features that resolve projective ambiguities. The Fig. 2 label will be corrected to “redundant vs. diverse predictions.”

Physical Connectivity: OSD’s orthogonality constraint operates on the prediction probability space, not the anatomical feature space, so it does not deny physical connectivity between vessels. It penalizes “all heads producing identical patterns across the entire image,” not co-activation at junction points, compatible with connected vessels. L_cons ensures the union of vessel categories covers the full binary mask, preventing the vascular tree from being torn apart.

2.Architectural & Inference Clarity (R1, R4) Inference Mechanism: The final prediction is produced by concatenating the outputs from N parallel segmentation heads with M_prior along the channel dimension, yielding the final multi-channel output. The max probability class is selected per pixel. P_ens are training-only topological anchors that enforce global consistency via L_cons.

Soft Grounding vs. Multiplicative Masking: Hard masking with M_prior risks error propagation; if the prior misses a low-contrast vessel, it is permanently deleted. L_cons provides a soft structural constraint that preserves gradient flow, allowing anatomical branches to recover uncertain regions.

3.Multi-Expert Synergy & SAM (R1, R4) SAM provides segmentation-aware priors (boundaries, connectivity) from its SA-1B mask pretraining, unlike ViT which lacks such structural priors. The CNN branch captures local textures; LoRA-tuned SAM captures long-range dependencies for fragmented vessels. Table 3 validates this: D^{CNN}_l gives 77.35% Dice, adding D^{Tr}_g reaches 77.62%, and D^{CNN}_g achieves 78.51% (clDice: 51.62% → 53.70%). D^{CNN}_g is “global” as it decodes SAM-derived global features.

Decoders for SAM features: As shown in Table 3, adding D^{Tr}_g focuses on long-range topological continuity, while adding D^{CNN}_g refines spatial alignment. Removing either leads to a drop in clDice or Dice.

4.Terminology of MoE (R1) We acknowledge the lack of a routing mechanism. We will refine “MoE-driven” to “Multi-Expert Synergy,” integrating heterogeneous features (CNN for local, SAM for global) rather than sparse conditional computation.

5.CA/VA Categorization (R1) CA and VA are distinct categories in our architecture. Their merged reporting in Table 1 was a space-saving measure; both achieve individual Dice of 0.9324 (CA) and 0.9579 (VA). We will report disaggregated metrics and VA-specific visualizations in the revision.

6.Baseline and Experimental Protocol (R1 & R4) Table 2 baseline, without SCSG or OSD, i.e., all auxiliary losses (orthogonality and consistency) are removed, uses a single segmentation head. Table 3 baseline removes all three decoders, relying only on the topmost local CNN decoder.

Swin-UNet Performance: Low performance stems from overfitting on the small dataset (175 images) and insufficient long-tail modeling. Pure Transformers lack CNN inductive biases.

Protocol: Fig. 4 reflects grid-search on the training set; the test set was only used for testing and reporting.

7.Performance on other datasets (R2) Cross-dataset validation is valuable, but our task requires pixel-level annotations for multiple anatomical vessel categories in DSA. No public DSA dataset currently provides a compatible label space. Evaluating on binary vessel datasets would not assess the proposed anatomical parsing task. We will release our code and dataset to benefit the community.



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.

    The paper addresses an important and clinically relevant problem moving DSA analysis from binary vessel segmentation to fine-grained cerebrovascular parsing. However, the current evidence leaves several issues unresolved: the final inference/fusion procedure is unclear, the “MoE” and SAM/global-encoder claims need stronger justification, the OSD design should be reconciled with anatomical connectivity and with the fact that orthogonality is applied to probability maps rather than features, and several experimental details require clarification.

  • 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 rebuttal addresses a meaningful portion of these concerns. Overall, I recommend acceptance. The topic is clinically important, the proposed framework appears to improve over the evaluated baselines, and two reviewers found the rebuttal sufficient to support acceptance.



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.

    Based on the outcome of the rebuttal process, two reviewers assigned an “Accept” rating, while one reviewer assigned a “Reject” rating. Although the paper still suffers from issues such as exaggerated claims and insufficient supporting evidence (R1), the study itself possesses “significant practical application value” (R2). Furthermore, the rebuttal successfully clarified certain doubts regarding the design rationale (R3), and the manuscript is clearly written (R2). Consequently, the overall consensus is positive, and the inclination is to “Accept” this paper.



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.

    Authors only partially addressed the reviewers concerns about overstated disentanglement, unclear inference/fusion, insufficient SAM-specific validation, and incomplete fine-grained evaluation. Please, refer to Reviewers’ post-rebuttal comments for more details.



back to top