List of Papers Browse by Subject Areas Author List
Abstract
Accurate modeling of liver anatomy and its vascular architecture from medical imaging is critical for surgical planning and navigation in liver cancer treatment. Although liver Couinaud segments are fundamentally defined by the branching topology of portal and hepatic veins, current methods typically treat vessel and Couinaud segmentation as isolated tasks, neglecting their intrinsic anatomical dependencies. To bridge this gap, we introduce a vascular skeleton-aware learning framework that jointly addresses both tasks through three key components. First, we propose a multi-branch Tri-directional Convolution (TriDConv) module to explicitly capture anisotropic vascular morphology from CT data along orthogonal spatial axes. Second, we design a Vascular Attention and Spatial-Topology (VAST) mechanism that enforces long-range structural coherence and topological continuity by integrating sequential attention with anatomical geometric constraints. Third, we introduce a novel point-sampling strategy leveraging vascular representations as geometric priors to guide 3D Couinaud segmentation. Extensive evaluations on public datasets demonstrate that our method consistently outperforms state-of-the-art approaches by significant margins across both volumetric and surface distance metrics. Code is available at https://github.com/YueQiu0911/VSA-Liver.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1418_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/YueQiu0911/VSA-Liver
Link to the Dataset(s)
N/A
BibTex
@InProceedings{QiuYue_Toward_MICCAI2026,
author = { Qiu, Yue AND Li, Yan AND Tong, Yuqi AND Liu, Qixuan AND Yang, Qian AND Qiu, Shi AND Heng, Pheng-Ann AND Fu, Chi-Wing},
title = { { Toward Synergistic Learning for Liver Vessel and Couinaud Segmentations } },
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 manuscript introduces a synergistic learning framework that utilizes Tri-directional Convolution (TriDConv), a Vascular Attention and Spatial-Topology (VAST) module, and a Skeleton-Aware Sampling (SAS) strategy to explicitly use extracted vascular skeletons as geometric priors for Couinaud segmentation, achieving strong performance on public datasets.
- 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 framework effectively bridges the functional definition of Couinaud segments with deep learning methodologies by using predicted vascular structures to dictate liver parcellation.
2.The VAST module’s approach to topological regularization, specifically using an Earth Mover’s Distance formulation and 3D morphological skeletonization to prevent broken vessel segments, is a rigorous and well-designed solution to a common artifact in tubular structure 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.The title and introduction heavily emphasize “Synergistic Learning”. However, the implementation details state that the vessel extraction modules (TriDConv and VAST) are frozen during the Couinaud segmentation phase. This constitutes a unidirectional, sequential pipeline rather than a truly synergistic process where both tasks mutually optimize each other. The framework leverages anatomical priors but lacks bidirectional feedback.
2.There is a gap in the ablation studies. While Table 3 evaluates the VESPA and TopoLoss components for vessel segmentation , the TriDConv module—introduced as the primary local feature extractor —is completely unablated. Furthermore, there is no ablation provided for the SAS module’s internal mechanisms (such as the Gumbel-Softmax temperature settings or feature aggregation strategies) on the Couinaud segmentation task.
- 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.
- 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 manuscript heavily promotes a “synergistic learning framework” to address liver vessel and Couinaud segmentations. However, the methodology is sequential. The implementation details explicitly state that the vessel extraction modules (TriDConv and VAST) are frozen during the subsequent Couinaud segmentation training phase. Consequently, there is no bidirectional gradient flow or mutual optimization between the two tasks. The manuscript lacks experimental evidence demonstrating true synergistic benefits, rendering the core claim unsupported.
- 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 transparently acknowledged that the framework operates as a unidirectional sequential pipeline rather than utilizing bidirectional optimization. While the initial use of “Synergistic Learning” was slightly overstated, the authors provided a compelling, clinically grounded justification for this asymmetric design (the inherent dependency of Couinaud segments on vascular structures). I can raise my score to weak accept.
Review #2
- Please describe the contribution of the paper
This paper proposes a synergistic learning framework for joint liver vessel segmentation and Couinaud segment delineation from CT images. Motivated by the anatomical principle that Couinaud segments are fundamentally defined by hepatic vascular topology, the authors argue that treating vessel and segment segmentation as independent tasks is suboptimal. To address this, the paper introduces three main components: (1) Tri-directional Convolution (TriDConv) to capture anisotropic vascular morphology along orthogonal spatial axes; (2) a Vascular Attention and Spatial‑Topology (VAST) module that enhances long-range connectivity and enforces topological consistency through attention mechanisms and a topology-aware loss; and (3) Skeleton‑Aware Sampling (SAS), which leverages learned vascular skeletons as geometric priors to guide Couinaud segmentation. The method is evaluated on multiple public datasets (MSD, LiVS, MSD8) and shows improved performance over state-of-the-art methods for both vessel and Couinaud 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.
Accurate vessel and Couinaud segmentation is highly relevant for liver surgery planning, resection safety, and functional liver remnant estimation.
The paper is grounded in well-established liver anatomy, correctly emphasizing that Couinaud segments are inherently defined by vascular branching topology. This makes the joint modeling strategy medically and biologically sound.
The method is evaluated on multiple public datasets with a wide range of metrics (Dice, VOE, ASSD, HD95, ASD), covering both volumetric accuracy and surface precision. The reported gains over strong baselines (e.g., LSFNet, ASG‑PVN) are consistent across datasets, particularly for surface distance and boundary accuracy, which are critical for clinical usability.
The inclusion of ablations for VESPA and TopoLoss strengthens the technical claims and demonstrates the contribution of individual components.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
While TopoLoss is conceptually appealing, its formulation (combining EMD over persistence features and skeleton overlap) is mathematically dense and may be difficult for readers to fully interpret or reimplement without additional explanation. As an integration operator, it is also unclear about the computational handling of this topological objective, e.g., differentiation, etc. Along similar lines, the proposed framework is quite involved with several modules and loss terms, including TriDConv, VAST, VESPA, TopoLoss, SAS, which corresponds to higher architectural and optimization complexity. In addition to what has been demonstrated with the ablation study on VESPA and TopoLoss, the relative necessity of each component is not always clear.
While evaluation and ablation study have been performed, it would be curious to relate the technical assessment to downstream clinical tasks, e.g., surgical margin planning and volume estimation, that really reflect clinical value.
The datasets largely reflect standard anatomy. Generalization under pathological conditions would be valuable to test.
While the paper claims identical experimental settings for all methods, details about hyperparameter tuning and architectural parity for comparison methods are not completely clear. Dataset is public. Hyperparameters setings are unclear. The implementation of Topoloss, which is also one of the mos signficant contributions, would pose the most signficant bottleneck to reproduce.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(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 presents a technically sophisticated and anatomically principled approach to joint liver vessel and Couinaud segmentation. The synergistic modeling of vascular topology and anatomical parcellation is well motivated, and the experimental results convincingly demonstrate performance benefits over existing methods. The approach is complex and would benefit from improved reproducibility, transparency, and stronger clinical validation. The contribution is substantial in anatomy-aware, clinically motivated medical image analysis.
- 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.
Despite the sequential rather than bidirectional handling of liver vessel and Couinaud, I feel this submission has sufficient merit in incorporating anatomical prior derived from a specific medical context. While the rebuttal did not fully address the technical questions about parameters and effective optimization strategy with topological quantities raised by reviewers, including myself, they have promised a public release of the implementation, which hopefully will clarify some of these limitations, in combination with the public presentation.
Review #3
- Please describe the contribution of the paper
This paper proposes a framework that explicitly leverages the anatomical dependency between liver vessel segmentation and Couinaud segment segmentation. More specifically, it introduces TriDConv for anisotropic vessel feature extraction, VAST/TopoLoss for topology-preserving refinement, and SAS for vessel skeleton-guided Couinaud segmentation, thereby structurally coupling the two tasks.
- 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 problem formulation is well motivated. Since Couinaud segments are anatomically defined by vascular topology, the idea of modeling vessel and segment prediction jointly rather than as isolated tasks is conceptually sound and anatomically meaningful.
2.The overall method design is coherent. TriDConv targets directional vascular structures, VAST/TopoLoss aims to preserve connectivity and topology, and SAS transfers vascular information to Couinaud segmentation. The role of each component is relatively clear.
3.The experimental results are generally positive. The method reports improvements over baselines in both vessel segmentation and Couinaud segmentation, and the qualitative results suggest reductions in jagged boundaries and false positives.
- 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 the usefulness of anatomical priors has already been well established in prior work, the paper does not sufficiently explain why a skeleton-aware sampling strategy is specifically required in this setting, rather than a simpler prior-injection scheme. The justification for SAS should be clear.
2.The comparisons and ablations are not sufficiently thorough. The paper does not independently analyze the contributions of TriDConv, SAS, or the proposed vessel-segment coupling, and the Couinaud experiments mainly rely on a “baseline + Ours” setup.
3.The paper would benefit from stronger qualitative evidence for the VAST mechanism. Since VAST is claimed to improve vascular path connectivity and branching consistency, more direct visualizations, such as 3D renderings or centerline/skeleton comparisons, would be helpful. The current 2D examples in Figure 2 are not sufficient to clearly demonstrate these topological improvements.
4.The term “synergistic learning” appears somewhat overstated relative to the actual training strategy. In the experimental setup, the Couinaud segmentation stage uses pre-trained and frozen TriDConv/VAST modules, making the method closer to a sequential two-stage pipeline than a fully joint learning framework. As such, the degree of “synergy” may be somewhat overstated.
- 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?
My recommendation is mainly based on the major weaknesses noted above. In particular, the paper does not sufficiently justify why the proposed SAS design is necessary beyond simpler anatomical-prior integration, and its practical sampling procedure is not clearly explained. The evidence for VAST is also limited, as the current qualitative results do not convincingly demonstrate the claimed improvements in vascular connectivity and branching consistency. Finally, the central framing of “synergistic learning” appears somewhat stronger than what is supported by the actual two-stage training strategy. Taken together, these issues weakened the methodological claim and led me to assign a weak reject score.
- 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.
The rebuttal clarifies that SAS differs from simple prior concatenation by using learned skeleton-based scores to guide sampling. However, my main concern remains unresolved. It is still difficult to assess whether the proposed SAS mechanism provides a clear benefit beyond simply injecting anatomical vascular priors into the Couinaud segmentation branch. As a result, the effectiveness of the proposed algorithm itself is not sufficiently demonstrated, and I do not find enough evidence to change my original weak reject recommendation.
Author Feedback
Dear Reviewers:
Thank you for your time and effort and also your constructive feedback. We appreciate the recognition of our method as bridging Couinaud anatomy and deep learning (R1), as anatomically principled (R2), and as a coherent design with positive results (R3). We have carefully looked into the comments and are happy to incorporate them to improve the manuscript.
Common: We’ll add TopoLoss/SAS details (λb, λt, SAS temperature τ, feature dimensions, and training hyperparameters) and release code upon publication.
(R1 W1, R3 W4) Thank you for pointing out “synergistic”. Your general understanding of the operation flow is correct, as we do not apply bidirectional optimization. However, this design is motivated by the anatomical vessel-to-Couinaud dependency, which is inherently asymmetric: vessels provide fine-grained anatomical landmarks and vascular branching cues, that directly inform Couinaud segments, whereas Couinaud labels provide only coarse region-level context and do not explicitly encode vessel geometry. A bidirectional design is technically new and worth future work. Also, our method is clinically explainable and practically robust. Similar staged synergistic designs were explored in prior work such as SynCHMR [Zhao et al., CVPR 2024], using HMR-derived priors to guide metric SLAM, where the dense scene guides a scene-aware SMPL denoiser, with synergy realized through sequential information transfer. We understand your concerns and will carefully revise the wording and claims in the paper.
(R1 W2) For TriDConv, replacing it with a standard 3D convolution under the same setting reduces LiVS Dice (83.48 to 83.05), and raises ASSD (0.67 to 0.73) and HD95 (3.54 to 3.85), supporting axis-decomposed anisotropic filtering for elongated vessels.
(R1 W2, R3 W1&2) SAS is not a simple vessel-feature concatenation. Direct concatenation adds vessel features under fixed sampling, whereas SAS uses predicted skeleton scores to modulate sampling toward vascular branching and boundary. Component checks isolate each ingredient: (i) feature concatenation only (no skeleton-aware sampling); (ii) skeleton-aware sampling only (vessel features not fed in); and (iii) the same sampling structure but with uniform neighbour weights. Their inferior performance confirms the respective contributions of vessel-feature injection, skeleton-aware sampling, and learned scoring. Table 2 further shows that SAS consistently improves Dice and ASD across different backbones, including PointNet2Plus and ASG-PVN. We use τ=1.0 to avoid overly random or overly peaked sampling; Gumbel weighted-sum preserves non-peak skeletal context compared with max aggregation and avoids uniform averaging compared with mean aggregation.
(R2 W1) For TopoLoss, Φ(·) extracts descriptors such as connected-component count and branch-length statistics for the EMD term, whereas S(·) uses the differentiable skeletonization operator.
(R2 W2&3) Fully agree. Our study focuses on anatomy-aware vessel/Couinaud segmentation on public datasets; pathological robustness will be our future work, as noted in Sec. 4. (R3 W1) For SAS, each point’s vessel-skeleton feature is projected into h_i and scored by Eq.(4); Eq.(5) then applies Gumbel Softmax to obtain differentiable selection probabilities, with τ controlling sharpness. Gradients thus guide the sampling toward vascular branching points and regions near vascular-territory boundaries, where Couinaud boundaries are ambiguous.
(R3 W3) We agree that 3D vascular-tree/centerline visualizations would better show VAST’s topological effect. Table 3 already provides evidence: on MSD, adding VESPA and TopoLoss reduces both ASSD and HD95.We test the centerline-level metrics on the MSD dataset: clDice rises from 0.739 to 0.756 when VESPA is added and to 0.757 when TopoLoss is added, while skeleton recall improves from 0.773 to 0.823 once VESPA is included, supporting better centerline coverage and fewer missed thin vessels.
Authors
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 receives mixed reviews (2 negative and 1 positive). Authors are invited to provide a rebuttal.
- 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.
This paper received mixed reviews before rebuttal and still gets mixed reviews after rebuttal. After carefully reading the paper, all reviewers comments and authors response, I found that several fundamental concerns of R3 are still valid: 1) the necessary of proposed SAS design (point-sampling method) over the much simpler direct anatomical-prior integration. There is no experimental results validating this point. 2) The evidence for VAST (topology consistency) is also limited, as there is no results regarding the topology metrics, such as clDice, Betti number, etc, and the current qualitative results do not convincingly demonstrate the claimed improvements in vascular connectivity and branching consistency. 3) It also lacks of comparison with nnUNet, which is a strong baseline for liver and vascular structure segmentation. Hence, I lean to rejection.
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.
After considering the rebuttal and all reviewers’ comments, the overall assessment remains positive. Although the term “Synergistic Learning” may be somewhat overstated, the paper’s strengths still outweigh the remaining concerns.
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.
Based on the rebuttal results, two of the three reviewers ultimately assigned an “Accept” rating, while one assigned a “Reject.” Notably, one reviewer revised their initial assessment from “Weak Reject” to “Accept”. Consequently, the overall evaluation was positive, leading to a final decision of “Accept.”