List of Papers Browse by Subject Areas Author List
Abstract
Blood vessel segmentation is a core task in medical image analysis for the care of vascular diseases and surgical planning, yet the challenges of providing expert vascular annotations pose a major obstacle for the progress of related deep learning techniques. To address this, we propose \texttt{VesselSim}, a two-stage framework for universal 3D blood vessel segmentation that eliminates the need for real annotated data during training. First, we introduce a stochastic, geometry-driven vascular simulation framework that models recursive branching, curvature-controlled growth, and collision-aware topology, followed by domain-randomized intensity synthesis to generate 16,500 anatomically plausible 3D angiographic volumes. Second, a 3D U-Net is trained solely on this synthetic data. To bridge the domain gap from synthetic to real images at inference time, we introduce a test-time adaptation strategy via a self-supervised mask reconstruction decoder, enabling adaptation to unseen clinical scans without prior domain knowledge. We evaluate \texttt{VesselSim} in a zero-shot setting on multiple real-world datasets spanning MR and CT across several anatomical regions, including the brain and kidneys. Despite being trained exclusively on synthetic data, \texttt{VesselSim} achieves performance competitive with state-of-the-art vascular segmentation foundation models. These findings suggest that learning vessel geometry from synthetic tubular structures is effective for robust cross-domain generalization, substantially reducing the reliance on acquired medical imaging data and more importantly, expert annotations.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3913_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://healthx-lab.github.io/VesselSim
Link to the Dataset(s)
Proposed synthetic dataset: https://healthx-lab.github.io/VesselSim
BibTex
@InProceedings{RaiEri_VesselSim_MICCAI2026,
author = { Rainville, Erin AND Ananian, Melissa AND Mirolla, Tristan AND Rivaz, Hassan AND Xiao, Yiming},
title = { { VesselSim: learning 3D blood vessel segmentation without expert annotations } },
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
The main contribution of the paper is the proposal of VesselSim, a 3D blood vessel segmentation framework based on synthetic-only training plus annotation-free test-time adaptation. The method first introduces a vascular simulation pipeline that combines recursive branching, curvature-controlled random walks, collision-aware topology, and physiologically motivated radius decay to generate vessel geometries with varied tortuosity and scale. These masks are then converted into angiography-like volumes through domain randomization. A 3D U-Net is trained entirely on this synthetic data, with cbDice added to encourage tubular continuity and topology preservation. At inference time, the model is further adapted to unseen real scans through a self-supervised mask reconstruction objective, updating only the encoder’s instance-normalization parameters. The key message of the paper is that strong geometry-driven priors, combined with test-time adaptation, can substantially reduce or even remove the need for real expert annotations in 3D vascular segmentation while remaining competitive with recent foundation models.
- 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 has three main strengths. First, it addresses an important practical problem: the high cost of vessel annotation. Second, the framework is coherent, with synthetic geometry, domain randomization, cbDice, and TTA fitting together well. Third, the results are notable: despite synthetic-only training, the method clearly outperforms VesselFM on TopCoW MRA and is also competitive on clDice for TopCoW CTA and HiP-CT, suggesting strong topology preservation.
- 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 novelty is somewhat integrative rather than fundamentally new. The synthetic vessel generation clearly builds on prior lines of work such as VascuSynth (Hamarneh & Jassi, 2010), stochastic L-system-based 3D vessel synthesis (Galarreta Valverde et al. , 2013), and more recent vessel generation models such as VesselVAE, VesselGPT, and VesselDiffusion. The current simulator appears to extend prior rule-based generation with curvature-controlled random walks, collision checking, and QC heuristics, but the paper does not directly quantify how much these additions matter compared with existing simulators. Likewise, the reconstruction-based TTA is close in spirit to prior self-supervised or reconstruction-driven test-time adaptation methods such as Context Encoders, Rec-TTT, and TopoTTA. Overall, the contribution feels more like a strong systems paper than a clearly novel methodological leap.
- 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.
(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 addresses an important and practical problem, 3D blood vessel segmentation without expert annotations, and presents a technically coherent solution. The proposed framework combines a geometry-driven synthetic vessel generator with annotation-free test-time adaptation, which is a meaningful direction for reducing dependence on costly real vascular labels. What mainly pushed the paper above the acceptance threshold for me is the empirical evidence: as shown in Table 1 (page 6), VesselSim, despite being trained only on synthetic data, achieves the best clDice on all three benchmarks and is especially strong on TopCoW MRA (71.8 Dice / 76.2 clDice). The comparison is particularly compelling because the paper states that VesselFM was pretrained on TopCoW and HiP-CT, whereas VesselSim uses no real annotated training data. I also found the ablation study reasonably supportive of the paper’s claims, showing that cbDice improves structural/topological consistency while the reconstruction branch and TTA improve cross-domain robustness. I do have reservations—mainly that the individual components are somewhat incremental and that the TTA protocol could be described more clearly, but overall I believe the practical significance, strong synthetic-only results, and solid experimental validation make this paper marginally above the acceptance threshold.
- 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 multi-stage framework, VesselSim to enable robust, zero-shot blood vessel segmentation on 3D medical image data. Particularly, VesselSim builds on a Unet architecture trained purely on multi-scale, synthetically generated 3D angiographic data and augmented with test-time adaptation of segmentations. Synthetic generation is based on an extended stochastic braching framework aimed at generating high-tortuosity samples. The full synthetic training dataset comprises from synthetic images from previously published framworks, augmented with the authors’ high-tortuosity samples. A classical composite training objective is extended with an auxiliary mask reconstruction loss through in-painting random vessel segments. The reconstruction decoder is also utilized to update instance normalization layers of the encoder in a single test-time adaptation epoch. Compared to vessel-foundation models (vesselFM, SAM-Med3D, UniverSeg), VesselSim shows superior Dice and superior center-line Dice performance on two and three out of three real vascular datasets (HiP-CT, TopCow-CTA/MRA, respectively.
- 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.
Accessibility: synthetic-only training enables valuable zero-shot segmentations that can highly boost efficiency of clinical research in neurovascular disease. Even if used as initial annotation to be refined, this can save significant time in the model prototyping process, enable teaching and involving of junior clinical raters and ultimately lead to more precise annotations and models.
Focus on real clinical challenge: Vessel tortuosity is a highly relevant clinical challenge in blood vessel assessment. The natural variation of human vasuclature and specifically the complex shapes and curvature of vessels make not only interpretation but intervention (e.g. intra-arterial thrombectomy) challenging and in many cases potentially harmful. Thus, extending research, datasets and models that better depict tortuosity of vessels is of high clinical relevance.
- 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.
Limited methodological novelty: While the authors present a unique combination of approaches, synthesis methodology is largely based on published literature [13]. Most importantly in my opinion, only half of the used synthetic vessel data is generated with the proposed method (i.e. high-tortousity vessels), thus results might show general benefit of the stochastic branching framework (i.e. solely application of it and no extension), regardless of the proposed extension. On the other hand, test-time adaptation does count as a novel contribution.
Limited demonstration of performance: While the included datasets are important benchmarks, they are limited in anatomical representation. TopCoW-CTA and TopCoW-MRA essentially comprise of paired imaging of the same subject/anatomy, thus from a structural vessel heterogeneity perspective, does not necessarily count as two distinct datasets. On the other hand, kidney blood vessels are naturally less tortuous (as the authors’ figures also show), thus further blood vessel segmentation datasets (e.g. ) should be included to reliably support the presented conclusions.
- 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
Could the authors clarify if the ablation presents cummulative methodology?
- Does “Base Unet + recon” also include “cbDice”? Does “+recon” also imply TTA?
- This is an important detail to justify the interpretation of performance improvement yielding from these methods alone or via their combination
Did the authors test training a low-tortuosity-only baseline within your framework?
- Since high-tortuosity synthetic vessels only take up half of the training data (apart from skull and background samples) the authors cannot surely claim that the improvement on top of vessel-foundation models originates exclusively from their proposed approach of extending the referenced and used stochastic branching framework.
The following result mentioned in the Discussion cannot be found in the current Results, to the best of my knowledge: “Notably, with the additional skull simulation, the mean Dice score was improved from 0.61 to 0.72 before TTA for the TopCow MRA dataset” If the authors have conducted experiments with and without skull simulation, it would be valuable to include in the presentation of Results for the relevant datasets (TopCow).
- 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 authors present an important direction of research, however the work as is has multiple limitations that might prevent justification of the authors’ conclusions. Part of the stated novel contribution (i.e. high-tortuosity vessel generation) is applied in parallel to existing generation methodology and not tested in isolation, thus reliability of the seen performance improvements might be questionable.
- 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 #3
- Please describe the contribution of the paper
VesselSim proposes a two-stage annotation-free methodological framework for 3D blood vessel segmentation. Firstly it used a process of stochastic branching simulation that generates 16,500 anatomically similar vascular patches spanning both low- and high-tortuosity morphologies, incorporating physiologically informed parameters (Murray’s law for radius scaling, breadth-first growth, collision avoidance). Secondly, it trains a 3D U-Net exclusively on this synthetic data using a composite loss that includes a centerline boundary Dice term, augmented by a self-supervised mask reconstruction auxiliary task designed to enable test-time adaptation (TTA) to unseen real clinical scans.
The framework is then evaluated zero-shot across three real-world datasets covering brain (CTA, MRA) and kidney (HiP-CT) vasculature and compared against vascular and general foundation models.
- 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.Zero-shot generalization without any real annotations Variability of geometry driven synthetic pretraining is supported with competitive performance across foundational models(FM) . VesselSim matches or is similar to FM despite training only on synthetic data and across three datasets (TopCoW CTA and HiP-CT) 2.Vascular simulation framework: Paper extends prior rule-based methods by incorporating curvature-controlled biased random walks, Murray’s law radius scaling, collision avoidance, and a quality control pipeline. 3.Lightweight and targeted test-time adaptation: Paper presents restricting TTA updates to instance normalization parameters with empirical validation. 4.Well-structured ablation study and evaluation across diverse imaging modalities and anatomical regions.
- 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 acknowledges that both TopCoW and HiP-CT are included in VesselFM’s training corpus, which confers a potentially large advantage to that baseline. A more detailed discussion of this asymmetry in the results section would be benefited. 2.Given the high variance (e.g., HiP-CT VesselSim: 31.6 ± 21.5 Dice), differences between methods may not be statistically significant. 3.Major vessels (carotid arteries, vertebral arteries, main renal arteries) are clinically important and their omission may limit applicability in certain downstream tasks 4.The abstract and introduction consistently spell the method name as “VessselSim” (three s’s) in the Conclusion section, while the rest of the paper uses “VesselSim.” This should be corrected. 5.Unreadably small parameter legend in Figure 1 and the absence of SAM-Med3D from the qualitative comparison. 6.Include at least one figure showing example synthetic vascular patches alongside their corresponding real-domain counterparts to visually ground the domain gap discussion.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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
Improvement of images with increase font size and separate images for more resolution and clarity.
- 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?
VesselSim designed synthetic data generation framework, combined with a lightweight TTA strategy, can match specialized vascular foundation models trained on large real-data which makes it well- above thresholds of acceptance but also presents with modifiable areas namely asymmetric comparison with overlapping training data, lack of clarity with TTA setup and high variance which prevents clinical adoption.
- 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
N/A
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.
All three reviewers independently scored your paper a “Weak Accept,” recognizing the practical significance and clinical relevance of your work. The consensus is that VesselSim provides a highly compelling framework. Achieving zero-shot performance that matches or even exceeds foundation models (like VesselFM), especially considering that VesselFM had the asymmetrical advantage of being pretrained on the evaluation datasets, is a notable accomplishment for the community. The reviewers also praised your integration of physiologically informed synthetic generation (e.g., Murray’s law, collision avoidance) to model challenging, high-tortuosity vessels. However, the reviewers have raised several important points regarding clarity, methodological details, and presentation. You are suggested to address the raised issues in your final version.