List of Papers Browse by Subject Areas Author List
Abstract
Large-scale collection of 3D vascular data is inherently difficult, resulting in scarce publicly available datasets. Existing methods represent vessel edges as discrete point sequences, which cannot guarantee smoothness and lose millimeter-scale size information through normalization. We propose 3-Phase Chebyshev Curve Diffusion, a framework that generates vessel curves directly in a continuous polynomial coefficient space, an approach not previously explored for vascular generation. Chebyshev coefficients yield a fixed-dimensional, length-independent representation that inherently ensures C-infinity smoothness without post-processing. We further design auxiliary modules that recover mm-scale size lost due to normalization and a mechanism that preserves overall curve trajectories. On the ImageCAS coronary artery dataset, we achieve superiority over two existing methods across key anatomical metrics, including up to a 5.2-times reduction in curvature Wasserstein distance, while confirming mm-scale accuracy.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4807_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/mozzi-bro/Chebyshev_Diffusion
Link to the Dataset(s)
ImageCAS dataset: https://github.com/XiaoweiXu/ImageCAS-A-Large-Scale-Dataset-and-Benchmark-for-Coronary-Artery-Segmentation-based-on-CT
BibTex
@InProceedings{MoJih_Anatomically_MICCAI2026,
author = { Mo, Jihwan AND Yoo, Sangbaek AND Choi, Jaesoon AND Chang, Dong Eui},
title = { { Anatomically Accurate 3D Vessel Generation via 3-Phase Chebyshev Curve Diffusion } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16889},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The core contribution lies in proposing a novel vessel curve representation (Chebyshev polynomial coefficient space) and a corresponding three-phase diffusion generation framework, which is primarily a methodological innovation.
- 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 representing vessel centerlines as Chebyshev polynomial coefficients is an elegant choice. Chebyshev polynomials achieve minimax-optimal approximation on [−1,1], providing clear theoretical advantages over B-splines (knot-placement dependency) and Fourier bases (Gibbs phenomena at open endpoints). The K=48 expansion achieves 0.003–0.020 mm reconstruction error, far below the ~0.3 mm voxel resolution, demonstrating the compact efficiency of this representation.
2.It is very interesting for polar decomposition and three-phase diffusion. Decomposing Cartesian coefficient pairs (c_n,k, c_b,k) into magnitude r_k and angle θ_k enables diffusion in domain-appropriate spaces. The ablation study convincingly validates this design (removing polar decomposition degrades curvature Wasserstein by 11×). The decision to sample θ_k≥1 from von Mises distributions is data-driven (Pearson correlation <0.04), reflecting careful analysis of the data distribution
3.It presented the comprehensive and convincing evaluation of the method. The evaluation spans tree-level (CD), edge-level (Wasserstein distances for curvature, tortuosity, radius, tapering), physiological constraints (Murray law compliance), and statistical coverage (Bounds compliance). The improvement of Murray’s law compliance to 35–37% is particularly noteworthy for physiological plausibility. The ablation study clearly isolates each component’s contribution.
- 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.For the presentation of topology, it is limited to tree structures. The method is constrained to binary tree topology by the Stage 1 RVAE, unable to handle loop structures (e. g. , circle of Willis) or more complex vascular networks. This severely limits generalizability. Discrete diffusion models [18,17] mentioned in related work already handle arbitrary network topologies, representing a notable regression. The authors should discuss specific technical challenges for extending to general graph topologies in more detail.
2.It somewhat unfairness of baseline comparisons. VesselGPT [8] was originally designed for cerebral vasculature, not coronary arteries, and supports only unconditional generation. Although the authors adapted it, cross-domain transfer may not represent VesselGPT’s true capabilities. VesselGPT supports only unconditional generation while this method and PartVessel use GT topology as conditioning, making qualitative comparison (Fig. 3) inherently unfair. The authors should provide more details about the adaptation and discuss its limitations. Moreover, with only two baselines (one being the foundation framework), the comparison lacks breadth. A comparison with diffusion in B-spline coefficient space, for instance, would more convincingly demonstrate the advantage of the Chebyshev representation.
3.With Stage training, the authors acknowledge a significant gap in mean bifurcation count for LCA (3.2 vs. GT 4.8), which directly affects tree-level CD and overall generation quality. However, the Stage 1 modifications are limited to auxiliary heads; the topology generation itself is not improved. Since Stage 1 topology constrains Stage 2 inputs, this upstream bottleneck may systematically limit overall quality and deserves deeper analysis. We would like to see the detail discussion about it in the revised version
4.The motivation emphasizes the value of generated vessels for segmentation training or endovascular robot simulation. While the experimental section validates no downstream applications. There is no demonstration that generated vessels improve segmentation via data augmentation, nor any evaluation of generated sample suitability in simulation environments. This leaves the practical clinical value without direct evidence.
5.Experiments are conducted only on ImageCAS coronary arteries (LCA and RCA). Given the relatively regular structure of coronary arteries (mainly binary trees), performance on more complex vascular structures such as cerebral and pulmonary vessels remains unknown. Cross-anatomical generalization is not validated.
- 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
1.Recommend reporting GT Murray’s law compliance: The ±20% threshold is used, but the GT compliance rate itself is not reported. If GT compliance is much higher than 35–37%, substantial improvement room exists; if GT is also low, the metric of clinical significance should be reconsidered.
2.We would like to see the sufficient discussion of generation diversity. The qualitative results mention “PartVessel” two samples are nearly identical while ours are visually distinct. A quantitative diversity metric such as average pairwise distance among generated samples would more rigorously support this claim.
- 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 core contribution, performing vessel curve diffusion generation in Chebyshev polynomial coefficient space, is a genuinely novel and technically deep methodological innovation. This representation elegantly addresses fundamental problems of discrete point-sequence representations (no smoothness guarantee, length-dependent dimensionality) while providing minimax optimal approximation guarantees. The polar decomposition and three-phase diffusion design have clear physical motivation and statistical justification, supported by thorough ablation experiments. The mm-scale recovery via ScaleHead is unique among existing vessel generation methods, with clear clinical translation value.
- Reviewer confidence
Very confident (4)
- [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 #2
- Please describe the contribution of the paper
The authors propose a vessel generation method based on the PartVessel framework. A 3-phase Chebyshev curve diffusion scheme is used to generate vessel curves while ensuring C^∞ smoothness. The method is evaluated on a public dataset and shows improved performance over two existing methods while preserving anatomical features.
- 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 3-phase Chebyshev curve diffusion and temperature-calibration mechanisms improve vessel generation, enhancing curve smoothness and mm-scale size consistency.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
The use of public datasets demonstrates the performance of the proposed method in specific cases. The validation could be strengthened by further testing on synthetic data that represents extreme or challenging scenarios.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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 3-phase Chebyshev curve diffusion and temperature-calibration mechanisms improve vessel generation, showing good performance on public datasets, but validation could be strengthened with synthetic tests under extreme or challenging scenarios.
- 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
This paper builds on PartVessel, which generates synthetic vessels in three steps: generating a key graph representing overall vessel topology, generating each vessel edge, and assembling the full vessel tree. Unlike PartVessel, which represents each edge as a discrete point sequence, this paper represents each edge as a continuous polynomial function using Chebyshev polynomials. This gives C∞ smoothness by construction and a fixed 102-dimensional vector per edge regardless of edge length or complexity. These 102 coefficients are generated using three separate diffusion models, handling bending magnitude, bending direction, and radius profile independently. Auxiliary modules ScaleHead and EdgeRadiusHead recover mm-scale dimensions lost through normalization, which is applied during training to stabilize 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.
Strengths:
1.Diffusing in Chebyshev coefficient space is genuinely new for vascular generation. Chebyshev polynomials achieve minimax-optimal approximation, giving 0.003–0.020mm reconstruction error against a 0.3mm voxel resolution using just 48 coefficients per edge.
2.Smoothness is guaranteed by the representation itself. This directly shows in results: curvature variance of 21.5 vs PartVessel’s 107.2, which nearly matches ground truth (23.9).
3.For actual medical use (surgical simulation, segmentation training), having real physical dimensions is important. Compared methods can’t be evaluated in mm at all. This method achieves within 5% error on length and 3% on radius.
- Using Murray’s law compliance as a metric goes beyond geometric evaluation. The improvement is substantial: 0% (PartVessel), 17% (VesselGPT), 37% (theirs), suggesting the continuous radius representation produces genuinely more physiologically plausible vessels.
5.Every component of the pipeline is individually tested, including hint nodes, polar decomposition, phase decomposition, and temperature calibration.
6.Observing that higher-order angles have near-zero correlation with conditioning (Pearson < 0.04) and sampling them from von Mises distributions rather than a diffusion model is an efficient, empirically grounded architectural choice.
- 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.
Weaknesses:
1.Stage 1 (key graph generation) is unchanged and is a clear bottleneck. The paper acknowledges that Stage 1 under-generates bifurcations for LCA (mean 3.2 vs ground truth 4.8). This directly hurts CD on LCA, which the paper attributes to Stage 1 rather than curve quality. Since Stage 1 determines the overall topology of the vessel tree, leaving it unimproved limits the clinical utility of the full pipeline regardless of how good Stage 2 is.
2.Murray’s law compliance is low in absolute terms. 35-37% means roughly two out of three bifurcations violate a known physiological law. The paper frames this as a win over baselines, which it is, but the absolute number is weak for any downstream clinical application. The paper does not discuss why compliance is still so low or how it could be improved.
3.The ScaleHead is a simple and expected engineering addition. Predicting a single scalar scale factor from the root hidden state is a straightforward MLP regression task. The more interesting question is whether the scale factor could be made a direct part of the latent space or sampled from a learned distribution rather than point-estimated, which would give control over tree size at inference. As implemented, ScaleHead always predicts the same deterministic scale for a given z, limiting size diversity in generated trees. This is a limitation worth discussing that the paper does not address.
4.The VesselGPT comparison is questionable. VesselGPT was designed for cerebral vasculature and is evaluated here on the coronary arteries out of domain. Despite this disadvantage, VesselGPT outperforms the proposed method on tapering (LCA) and radius (RCA). This raises questions about how competitive the proposed method would be against VesselGPT on its intended domain and weakens the validity of the comparison overall.
5.Single dataset and single anatomy. The method is evaluated exclusively on the ImageCAS coronary artery dataset with 855/861 samples. There is no evaluation on other vessel types such as cerebral, hepatic, or pulmonary vasculature, which have different geometric and topological properties. It is unclear whether the Chebyshev representation and the 3-phase decomposition generalise beyond coronary arteries.
6.K=48 coefficients is not justified beyond reconstruction error. The paper sets K=48 based on achieving low reconstruction error but provides no analysis of the tradeoff between K and generation quality. A smaller K would mean a lower-dimensional diffusion space, potentially easier to learn. It is possible that a much smaller K would achieve similar generation quality while being simpler, but this is never investigated.
7.The paper builds heavily on PartVessel but provides only a brief and difficult to follow description of it, making the paper hard to understand without referring to the original PartVessel work. A more complete description of PartVessel would make the paper more self-contained.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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
The paper would benefit from a brief introduction to Chebyshev polynomials, as readers unfamiliar with approximation theory may find it difficult to follow the core technical contribution without this background.
- 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 primary factor in favour of acceptance is the representational shift from discrete point sequences to Chebyshev coefficient space, which is principled, novel, and demonstrably effective. The primary factors limiting the score are the unchanged Stage 1, which the paper itself acknowledges as a bottleneck, and the evaluation on a single dataset and single anatomy, leaving generalisability unclear. These open questions place the paper at Weak Accept.
- 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
Author Feedback
We thank the reviewers and AC. To our knowledge, vessel-centerline diffusion in Chebyshev coefficient space (Section 1) is novel; polar decomposition is validated by ablation (Table 2: 11× curvature degradation when removed). All quantities derive from existing data (no new training/generation).
1.VesselGPT comparison [R1, R3]. VesselGPT (MICCAI 2025), originally evaluated on cerebral vasculature, is autoregressive; we adapt it to ImageCAS coronary arteries and, since VesselGPT supports only unconditional generation, evaluate 1,000 samples (Section 4). Across 14 cells (Table 1), ours obtains the best in 11—curvature and tortuosity Wasserstein best on both arteries; Murray’s compliance (35–37%) surpasses PartVessel (0%) and VesselGPT (15–17%). VesselGPT leads on tapering (LCA) and radius (RCA), aligned with its B-spline cross-section parameterization (Section 2); our 3-phase design trades local radius fidelity for global geometry. CD: RCA best (Ours 59.0 / PartVessel 72.8 / VesselGPT 119.9), LCA close to PartVessel (Ours 59.1 / 57.0 / 99.0); the LCA gap is from Stage-1 bifurcation undercount, not curve quality.
2.Stage-1 bifurcation undercount [R1, R3]. Stage-1 extends PartVessel’s RVAE to additionally predict degree-2 hint nodes (intermediate trajectory waypoints, Section 3.3), used to condition Stage-2’s curve generation. Removing hint nodes degrades curvature Wasserstein by 11× (Table 2). Per-tree node count rises 11.45→23.87 (LCA, +12.42) and 7.21→16.79 (RCA, +9.58); hint nodes contribute +12.42 (LCA, all) and +9.33 (RCA, ~97% of the increase). Under the same RVAE capacity, this larger prediction target likely reduces bifurcation accuracy; a diffusion-based Stage-1 is a future direction.
3.Murray’s law and GT reference [R1, R3]. Following R1, we computed GT compliance on ImageCAS: strict (±20%) = 16.6%/15.9% (LCA/RCA); loose (±50%) = 53.1%/62.4%. Held-out 163 is consistent (16.6%/17.1%). GT mean of (Σr_child³)/r_parent³ is 1.46/1.40 vs ideal 1.0—reflecting that GT radii are skeleton-derived from standard centerline-extraction preprocessing (not direct mm measurements), inherited equally by all methods. Under this preprocessing-bounded reference, our 35–37% (2.2× GT, uniformly best) represents substantial movement toward physiological plausibility; closure to ideal Murray would require advances in centerline-radius estimation itself.
4.Diversity quantification [R1]. Manifold coverage on LCA (fraction of GT 6D-feature k-means clusters reached): Ours 0.70 vs PartVessel 0.50.Raw diversity (coefficient of variation of per-edge curvature/radius/length): +14% (LCA), +59% (RCA) over PartVessel.
5.ScaleHead determinism [R3]. ScaleHead is deterministic for fixed z. By design, Stage-2 operates in normalized coordinate space (Section 3.1); mm-scale recovery is architecturally confined to Stage-1’s root latent z, not an arbitrary choice. Tree-level scale diversity arises from different z; length/radius match GT within 5%/3% (Section 4.1). Decoupling scale from z would require two-stage redesign (future direction).
6.K=48 and B-spline baseline [R3, R1]. K=48 reconstruction error (0.003–0.020 mm) is well below voxel resolution (~0.3 mm, Section 3.3). Smaller K (e.g., K=24) also yields sub-voxel reconstruction; we chose K=48 since the model handles it stably with extra precision margin (K tunable to dataset/compute). We considered B-spline coefficient diffusion: per-curve adaptive knots would add a Stage-1 prediction target—and given the Stage-1 trade-off from hint nodes (Topic 2), would likely worsen bifurcation accuracy; fixed shared knots compound a hyperparameter with K. Chebyshev avoids both via an orthogonal global basis (Section 2).
Limitations. Loop-topology via graph diffusion (Section 5), downstream applications, cross-anatomy validation, and synthetic stress testing are future directions. A short Chebyshev intro and PartVessel-summary clarification fit within camera-ready limits.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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.
Reviews are positive and note significant empirical improvements and a relatively novel vessel representation/learning method.
