List of Papers Browse by Subject Areas Author List
Abstract
Anatomical trees are critical for clinical diagnosis and treatment planning, yet their complex and diverse geometry make accurate representation a significant challenge. Motivated by the latest advances in large language models, we introduce an autoregressive method for synthesizing anatomical trees. Our approach first embeds vessel structures into a learned discrete vocabulary using a VQ-VAE architecture, then models their generation autoregressively with a GPT-2 model. This method effectively captures intricate geometries and branching patterns, enabling realistic vascular tree synthesis.
Comprehensive qualitative and quantitative evaluations reveal that our technique achieves high-fidelity tree reconstruction with compact discrete representations. Moreover, our B-spline representation of vessel cross-sections preserves critical morphological details that are often overlooked in previous’ methods parameterizations. To the best of our knowledge, this work is the first to generate blood vessels in an autoregressive manner.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/0668_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/LIA-DiTella/VesselGPT-MICCAI
Link to the Dataset(s)
http://ecm2.mathcs.emory.edu/aneuriskweb/repository
BibTex
@InProceedings{FelPau_VesselGPT_MICCAI2025,
author = { Feldman, Paula and Sinnona, Martin and Delrieux, Claudio and Siless, Viviana and Iarussi, Emmanuel},
title = { { VesselGPT: Autoregressive Modeling of Vascular Geometry } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15975},
month = {September},
page = {670 -- 680}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes to use a GPT-style model to generate vascular trees autoregressively, cross section by cross section. The model is trained to generate arteries from the circle of Willis that have aneurysms, using data from the Aneurisk dataset. The authors suggest a two-stage approach: (1) cross sections are embedded as a vocabulary, and (2) the learned vocabulary is used to generate new vascular trees in an autoregressive manner. The main contributions of the paper are the following.
-
The authors introduce a new method for encoding vessels as a sequence of cross-sections. Each node is represented by its spatial coordinates and the control points of a B-spline that defines the outline of its cross section. The nodes of the vascular tree are ordered using graph preorder traversal, with a zero attribute vector to represent branch termination.
-
They propose to use a Vector-Quantitized Variational Encoder (VQ-VAE), introduced in previous work, to learn a better embedding of the nodes, incorporating information about neighbor nodes and the distribution of vascular features.
A GPT-style transformer is trained to generate a vascular tree by predicting a sequence of nodes using the embeddings provided by the VQ-VAE model. These embeddings are then decoded back to the original B-spline representation by the VQ-VAE decoder.
- The authors also present a method for reconstructing a mesh from the sequence of cross sections, where they interpolate the surface by signed distance fields and reconstruct the mesh with the marching cubes algorithm.
Finally, the author compares their method qualitatively and quantitatively with existing methods.
-
- 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 proposed autoregressive approach is novel. While autoregressive models have been investigated for graph generation, their application to generate synthetic vascular networks has not been explored yet. This paper takes advantage of recent advances in modeling sequential data, such as those used in language processing, by using a GPT-style architecture.
2) The method proposed to encode vascular trees using cross-sections represents a novel contribution. Accurately modeling the non-circularity of cross-sections is crucial for computational fluid dynamics simulations (especially in aneurysms). Previous methods either generated the mesh itself or represented the vascular tree by a set of nodes with radius. While the mesh-based approach allows a lot of freedom in the vascular tree shape, it can also introduce topological errors. Conversely, the centerline-based approach offers better control over the topology but relies on the assumption that cross sections are circular, which is not always the case. The proposed representation achieves a good trade-off between both approaches.
3) The authors made the effort to compare their method with other existing models, which is often neglected in generative methods. This occurs because it can be challenging to quantitatively assess the realism of the generated meshes. This is particularly true in emerging applications, where the novelty of the approach is frequently used to justify the absence of comparisons with previous methods.
- 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 main weakness of the paper lies in the validation of the proposed method.
1) If I understood correctly, the “Dv” and “Dec.” decoders shown in Figure 1 are actually the same. If this is correct, please use consistent naming to avoid confusion.
2) During the centerline extraction with VMTK, are aneurysms treated as branches? It seems to be the case in Figure 1. a. Do you differentiate in your data processing between saccular aneurysms (which can be defined as branches) and fusiform aneurysms (which are simply deformations of the vascular surface)?
2) One of the most recent method (Kuipers et al. MIDL 2024) is mentioned in the introduction but not included in the validation. It might be good to justify this choice; was the code unavailable?
3) The metrics used to measure performance do not clearly demonstrate the advantages of the proposed method over existing approaches. From the results of Table 1, one might conclude that the method by Feldman et al. yielded the best results. While the author suggests that the good scores of Feldman et al.’s method may be due to memorization of the samples, this claim should be supported by more evidence.
Using the distribution of vessel morphometrics, such as tortuosity and length might demonstrate the advantage of the method better? The authors appear to have already computed length similarity and tortuosity similarity, which are reported in the text. Using the similarity between distributions as an additional metric to include in Table 1 may show the advantages of the proposed method. Additionally, incorporating other metrics, such as radius similarity or mesh surface curvature similarity, could put in light the limitation of the other methods, particularly Feldman et al., which appears “bumpy” in Figure 2. Perhaps topological metrics could also be beneficial, as some of the models from Sinha et al. seem to include cycles in Figure 2. While in the qualitative study of Figure 2, the proposed method seem to provide the most realistic results, the quantitative analysis should also reflect this.
4) In Figure 2, showing both success and failure cases for each method would make the qualitative analysis more trustworthy by making the choice of the samples less biased. Given that the quantitative evaluation does not strongly support the new model, it is important for the reader to be convinced by the qualitative results.
5) The sampling of cross sections appears quite low in Figure 1, which may lead to a smoother mesh after interpolation. Have you tried a more dense sampling too?
6) The use of free shapes instead of circular cross-sections is mentioned as a contribution of the model. However, it is unclear whether this actually improves the realism of the vascular shape. In Figure 2, the cross sections of the generated meshes looks circular, and there are no quantitative metrics to measure this improvement. What was the deviation from circular cross-sections in the original Aneurisk dataset? Did the use of splines enable a more accurate representation of this non-circularity? Including illustrations of some cross-sections or additional metrics that measure the non-circularity in both the ground truth and generated meshes would support the authors’ claims.
- 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.
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — could be rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The concept of representing vascular trees as a sequence of cross sections and the auto-regressive approach are promising. Looking at Figure 1, it appears that the proposed model indeed enhances the realism of the generated vessel. However, the validation needs improvement, as the quantitative analysis is insufficient to support and the authors’ claims regarding the performance of their model.
- 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.
My main concern was the quantitative evaluation, that did not demonstrate well the advantages of the proposed method. The authors have provided new quantitative results in the rebuttal that strongly support their method. Therefore, I think the paper should be accepted.
Review #2
- Please describe the contribution of the paper
- This paper proposes a vascular structure modeling method based on an autoregressive transformer architecture. The method parameterizes vascular as a sequence of nodes and contours, and succeeded to model a branching tree structures.
- 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 modeling algorithm is technically sophisticated . This will be an early work that uses auto-regressive approach to generate vascular structure. -In terms of applications, successful modeling of branching tree structures is considered as novel.
- 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 generation approach is certainly new, but the degree of improvement shown are relatively small. -The experimental part to show the method’s efficacy relies only on existing quantitative measures. Their clinical impacts or applications have not been discussed.
- 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.
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — could be rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
-The paper focuses only vascular modeling accuracy, but the theoretical progress and the degree of performance improvement are limited.
- 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.
My evaluation of this paper is neutral. Considering the novelty of modeling branching structures, I could not justify rejection.
Review #3
- Please describe the contribution of the paper
The authors propose the first autoregressive method to automatically generate vessel tree meshes. They re-parameterize the mesh and encode node representations with a VQ-VAE, which are used to generate vessels autoregressively using a decoder only GPT 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.
- Re-parametrization using the full cross section instead of a single radii parameter is a fantastic addition, which enables more realistic vessel morphologies.
- One explicit motivation for the work is that [9] can only generate relatively small vascular segments. This is backed up with good qualitative evidence (height 20)
- Qualitative analysis shows exceptional results
- 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.
- its not entirely clear what the technical contributions of the paper as these are not explicitly stated.
- [9] outperforms the method on the given metrics, which the authors explain is due to misaligned metrics. This was a missed opportunity to present a new metric that better captures performance and show that the work improves upon [9,27].
- Qualitative analysis is strong, but hard to reach general conclusions with such a small sample of predictions (8)
- 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.
- 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
- I recommend specifically stating your contributions to avoid confusion.
- Table 1 formatting is messy
- 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?
Novelty of autoregressive blood vessel generation, and strong qualitative results. The quantitative results do not show that their method offers any improvement. Missed opportunity to analyse deficiencies in existing metrics and propose a new improved one.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
I am glad that the authors agree that better evaluation metrics are needed in this field. Despite similar scores to competing methods on the typical metrics, they improve on efficiency and other alternate performance metrics which are well evidenced. More importantly, I am convinced that others in the community will find their approach technically interesting with potential generalisation to other sequential generation tasks.
Author Feedback
We thank the reviewers for their thoughtful feedback. We now provide a point-by-point response to the raised concerns. (R1/R3)Quantitative evaluation does not reflect qualitative results Existing evaluation metrics have known limitations and are not always well suited to capture the anatomical complexity of vascular structures and no baseline has performed an evaluation against SOTA. Being the first ones to perform the comparison, our method performs comparably to the state of the art: it surpasses one of the baselines across all reported metrics, and when compared to Feldman et al. [9], it performs similarly, with only marginal differences. As noted by R3, there is a clear need for new evaluation strategies tailored to this domain. Despite metric limitations, our approach offers distinct advantages: it is more efficient (3 hours vs over 24 hours training time), scales to larger and deeper vascular trees (height 20 vs height 15 in Table 1), and enables the representation of more complex anatomical features by modeling non-circular cross-sections (allowing for aneurysm modelling), an aspect not reflected by standard metrics but vital for downstream applications. Comparing our method to [9] in terms of feature distributions (tortuosity and length) shows that our approach aligns closely with the dataset samples. At height 15, our method achieved a tortuosity cosine similarity of .96 and for length .87, outperforming VesselVAE, which scored .46 in tortuosity and .83 in length. At height 20, our method reached .97 in tortuosity and .88 in length, while [9] achieved .77 and .38. (R1)Advantage of using splines for cross-section representation While developing the framework, our data exploration revealed that vessel cross-sections are far from being circular. Previous methods have overlooked this detail, often assuming circular shapes. As R3 highlighted and the statistics confirm, this re-parameterization allows for more realistic vessel morphologies. The Hausdorff distance between the dataset cross-section samples, our spline representations, and various circular fits shows that splines consistently achieve lower errors (Splines [Avg: 0.86; Std: 4.18], Average circle [Avg: 15.11; Std: 16.88], Minimum enclosing circle [Avg: 6.02; Std: 7.12], Maximum inscribed circle [Avg: 11.38; Std: 12.9]), confirming splines’ superior ability to capture the shape of vessel cross-sections in datasets such as Aneurisk. Accurate modeling of aneurysms particularly benefits from this representation, as their cross-sections tend to be highly irregular. Splines provide the flexibility needed to capture their morphology faithfully. (R3)Technical Contribution Our main contributions are: (1) a novel autoregressive formulation using a Transformer-based architecture to model vascular trees as sequences, and (2) a B-spline cross-sectional representation that captures more accurately the vessel geometries allowing aneurysm modelling, something baseline methods have not explored. (R1)Aneurysms as branches Our datasets contain only saccular aneurysms. They were treated as branches during centerline extraction with VMTK, as shown in Fig. 1. (R1)Low sampling in Figure 1 The figure is intended for illustrative purposes. The number of cross-sections per vessel in the dataset is significantly higher, typically around 200, depending on the sample, calculated by the VMTK network extraction algorithm. (R1)No evaluation against Kuipers et al. We prioritized comparing against methods that are more closely aligned with our approach. The method by Kuipers et al. (2024) focuses on simulating stroke treatment in vessels with occlusions. (R2)Clinical applications not discussed Our work has potential impact in many clinical applications, such as preoperative planning, flow simulations, and synthetic data generation for training segmentation or diagnostic models. Exploring these directions is part of our planned future work. We will incorporate all clarifications in the manuscript
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.
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 #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A