Abstract

Existing datasets for brain vessel segmentation in CTA are limited in scale and largely restricted to major vessels, leaving peripheral vasculature underrepresented. We address this data-level bottleneck by introducing DynaVessel, a richly annotated dynamic 4D-CTA dataset, alongside a novel annotation pipeline that exploits multi-phase contrast enhancement and background subtraction to reduce the effort required for expert manual annotation of vessels. This not only enlarges our dataset by 5-6 fold, but also provides a novel data-driven method for training deep learning models to detect arteries and veins despite the wide variation in vessel contrast encountered in single-phase CTA scans. DynaVessel comprises 110 training volumes from 29 patients and 165 test volumes from 14 patients. By comparing with two similarly-sized datasets for brain vessel segmentation in CTA, we demonstrate that DynaVessel-trained nnUNet models produce superior segmentations compared to the same model trained on single phase data, with an average mDC of 0.846/0.957 for arteries/veins, shorter average directed Hausdorff distance (0.304 mm/0.08 mm for arteries/veins) and improved topology sensitivity (0.877/0.974 for arteries/veins) in the TopBrain dataset. Code, model weights, and instructions to access the dataset are available online: https://github.com/alceballosa/robust-vessel-segmentation.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1486_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/alceballosa/robust-vessel-segmentation

Link to the Dataset(s)

DynaVessel (ours): github.com/alceballosa/robust-vessel-segmentation TopCoW: https://zenodo.org/records/15692630 TopBrain: https://zenodo.org/records/16623496 VesselVerse: https://i-vesseg.github.io/vesselverse/

BibTex

@InProceedings{CebAlb_Robust_MICCAI2026,
        author = { Ceballos Arroyo, Alberto Mario AND Yadav, Shrikanth M. AND Lin, Chu-Hsuan AND Kim, Jisoo AND Young, Geoffrey S. AND Qin, Lei AND Jiang, Huaizu},
        title = { { Robust automatic brain vessel segmentation in 3D CTA scans using dynamic 4D-CTA data } },
        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 paper presents a data-centric approach for intracranial vessel segmentation in CTA by leveraging dynamic 4D-CTA acquisitions. It introduces a semi-automatic pipeline that uses phase-specific contrast differences and image registration to separate arteries and veins, enabling efficient annotation with reduced manual effort. Using this pipeline, the authors construct the DynaVessel dataset, which includes multi-phase CTA volumes with artery and vein labels. They demonstrate that training a standard nnU-Net on this multi-phase data improves robustness to contrast variability and yields improved segmentation performance, particularly for vein detection, compared to models trained on conventional single-phase 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.

    Effective use of dynamic 4D-CTA for annotation efficiency: The paper proposes a practical and well-motivated pipeline that leverages temporal contrast dynamics to separate arteries and veins. By reusing annotations across multiple phases, the approach reduces manual labeling effort, which is a major bottleneck in vascular segmentation tasks.

    Data-centric contribution with multi-phase diversity: The introduction of the DynaVessel dataset, containing multi-phase CTA volumes with artery and vein labels, is valuable. In particular, the inclusion of multiple contrast phases per patient provides richer supervision and exposes models to a broader range of intensity variations, which is often missing in existing CTA datasets.

    Improved robustness to contrast variability: The experimental results demonstrate that training on multi-phase data improves robustness across different contrast phases. The analysis of performance across phases (e. g. , Fig. 4–5) provides useful insight into how contrast timing affects segmentation quality.

    Clear and interpretable preprocessing pipeline: The artery/vein separation method based on registration and intensity comparison is simple, interpretable, and grounded in physiological contrast differences, making it easy to understand and potentially reproduce.

    Use of standard and well-established backbone (nnU-Net): By building on nnU-Net, the paper isolates the contribution of the data and preprocessing pipeline rather than architectural changes, which strengthens the claim that performance gains are driven by improved data quality and diversity.

  • 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: The paper does not introduce a new segmentation model or learning framework, but instead relies on a standard nnU-Net backbone. The main contribution lies in the data curation and preprocessing pipeline.

    Limited validation of annotation quality: The ground truth is generated via a semi-automatic pipeline (iCafe + thresholding) and validated by a single neuroradiologist. There is no assessment of inter-observer variability or annotation uncertainty. Given the difficulty of artery–vein separation and distal vessel labeling, the reliability of the annotations is not sufficiently established.

    Evaluation limitations and metric choice: The paper emphasizes a modified Dice coefficient (mDC) that measures sensitivity but does not penalize false positives. This is non-standard and may bias results toward higher scores. Standard Dice, precision-recall analysis, or statistical significance testing are not sufficiently reported, making it harder to fully assess performance improvements.

    Limited evidence for generalization across scanners and sites: All data appears to be acquired from a single scanner and institution. Despite claims of robustness to real-world variability, there is no evaluation on multi-center or multi-scanner datasets, which limits confidence in clinical applicability.

  • 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

    The paper addresses an important and practical problem in CTA vessel segmentation and presents a thoughtful data-centric approach. The idea of leveraging dynamic 4D-CTA to reduce annotation burden and improve robustness across contrast phases is well-motivated and has clear potential impact. The empirical gains, particularly for vein segmentation, are promising.

    To further strengthen the work, the authors may consider clarifying and reinforcing several aspects:

    Annotation quality and validation: Since the dataset is a key contribution, additional discussion on annotation reliability (e. g. , inter-observer variability or quality control procedures) would improve confidence in the results. Evaluation metrics and reporting: Including standard Dice scores alongside the proposed mDC metric, as well as additional analysis of false positives/precision, would provide a more complete picture of segmentation performance. Generalization: The paper would benefit from a more cautious discussion of generalization, especially given the single-center data source. Clarifying this limitation explicitly would strengthen the overall presentation. Positioning of contributions: The paper would be clearer if it explicitly frames itself as a data-centric contribution (dataset + annotation pipeline) rather than a methodological advance, which would better align expectations with the presented results.

    Overall, this is a solid and practically relevant contribution, and addressing the points above would further improve its clarity and impact.

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The paper presents a well-motivated and practically relevant data-centric contribution, introducing a semi-automatic annotation pipeline that leverages dynamic 4D-CTA to efficiently generate artery and vein labels. The resulting multi-phase dataset is valuable, and the experiments demonstrate clear improvements in robustness to contrast variability, with particularly strong gains in vein segmentation. The analysis of performance across contrast phases is also insightful and highlights an important challenge in CTA-based segmentation.

    While the methodological novelty is limited and some concerns remain regarding dataset size (in terms of unique patients), the overall contribution, especially the combination of annotation strategy, dataset, and demonstrated impact on robustness, is meaningful for the MICCAI community. The work provides useful insights and a practical direction for improving vessel segmentation.

    Overall, the paper is slightly above the acceptance threshold, and I would support acceptance provided the authors clarify the evaluation setup and better position their contribution.

  • 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 #2

  • Please describe the contribution of the paper

    This work introduces a dynamic CTA dataset with whole brain paired arterial and vein annotations, including an annotation protocole for the dynamic modality. Authors show that training segmentation models dervied from such dynamic data can improve vasculature segmentation performance, even in out-of-domain 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.
    • The provided dataset is very relevant for the community and unique; to my knowledge, no other dynamic CTA datasets are available.
    • Authors show that models derived from such dataset outperform vasculature segmentation when compared to prior datasets.
  • 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.
    • Novelty: While the dataset + protocol is definitely a contribution in the field, the derivation of a vasculature segmentation model is limited and weakly evaluated. Authors make use of an nnUnet model -as were used in prior works in diverse tastes, e.g. TopCow/TopBrain- with limited baseline comparisons and ablations. Beyond the dataset, it is not clear what is the ‘methodological contribution’ of this work.

    • Labels: While authors could have obtained a timepoint-specific vessel mask for each CTA timepoint in a series; instead, a single arterial / vein mask is annotated from the Peak arterial/venous phases. Given that the dataset is the main novelty in this work, this represents a limitation, since no single timepoint enhances the whole brain vasculature so that it the labels can be considered as a ground-truth for all the timepoints in the series. Such differences are even strengthened in some pathologies (e.g. stroke), where delayed bolus arrival to the ipsilateral brain hemisphere creates larger bolus arrival delays.

    • The work is not easy to read. Please check below my comments about it.

  • 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 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
    • Are images defaced/skull-struppied for full patient anonymization purposes?
    • The abstract is not easy to read, unless someone is expert in this field. There is no background nor introduction: why this topic is important to be addressed? what are the problems with current models/datasets?
    • Also abstract: “ (…) models trained on our dataset produce superior segmentations compared to the same models (…)” –> but then, only results of the proposed model are mentioned. Authors could summarize here how “much better” these results are.
    • Can you briefly include what is icafe? It is mentioned along the text; but note that not everyone reading this work will know about it.
    • Introduction: “(…) because of the abscence of interphase motion (…)” –> What is interphase motion in this context? Does it refer to patient motion? Then, this is not 100% guaranteed.
    • Did you compensate for patient head motion? How?
    • Algorithm: is very hard to follow without any context. What is Sa, Sv, Xa, Xv? Please define.
  • 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 novelty of this work lies exclusively in the labelled dataset, as models derived from it have been widely explored in previous works, with more extensive validations and analysis (which are mostly missing here). Being the dataset the most relevant contribution here, the annotation protocol still derives labels from single timepoints rather than separate masks for each of the.

  • 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

    The main contribution of this paper is the collection and annotation of a large-scale 4D-CTA brain vessel dataset with arterial and venous labels. By leveraging dynamic 4D-CTA acquisitions, the authors efficiently generate multi-phase CTA volumes for each patient and reuse annotations across phases, effectively increasing the dataset size by a factor of five to six. The paper demonstrates through experiments that training deep learning models on this large, multi-phase dataset improves the robustness and accuracy of brain artery–vein segmentation, particularly under varying contrast conditions.

  • 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.

    -Novel and valuable dataset: The most significant contribution lies in the construction of a relatively large 4D-CTA dataset for brain vessel segmentation, which is rare in this domain. The use of dynamic CTA allows efficient annotation reuse across phases and provides rich temporal contrast information.

    -Clear empirical benefit of data scale: The experiments convincingly show that increasing the amount and diversity of training data leads to improved artery and vein segmentation performance, including better topology-related metrics.

  • 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.

    -Restricted impact on the broader community: Although code and models are promised to be released, the dataset itself does not appear to be publicly distributable. This limits the potential impact of the work on the broader research community and its reuse in follow-up studies.

    -Limited methodological novelty: While the dataset itself is novel, the technical approach is not. The segmentation models and training strategy rely entirely on established deep learning frameworks (e.g., nnU-Net) and standard loss functions. No new network architecture, learning strategy, or theory is introduced.

  • 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 paper presents a valuable and well-curated 4D-CTA dataset and clearly demonstrates the benefits of large-scale, multi-phase data for brain vessel segmentation. However, the lack of methodological novelty and the limited accessibility of the dataset reduce its impact. Thus, this is considered to a borderline submission for MICCAI.

  • Reviewer confidence

    Very confident (4)

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

    Accept

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

    The construction of a large-scale 4D dataset is a valuable and novel contribution. This effort goes beyond the typical scope seen in prior work. Overall, I believe the paper meets the acceptance threshold.



Author Feedback

We thank the reviewers for their insightful feedback, which will significantly improve our manuscript:

Generalization (R1): To clarify, DynaVessel (DV) comes from one hospital using a Canon-Toshiba scanner, and we evaluate the model on O.O.D. datasets from 2 countries in another continent (Europe): 1) a subset of ISLES, 2) TopCoW (+derived datasets), which was acquired with a Siemens scanner. Covering more scanners/hospitals would be ideal, but the fact a model trained on N=29 patients with multi-phase CTA can achieve SOTA performance on O.O.D. data suggests strong generalization

Data availability/Anonymization (R2,R3): Due to changing federal government and institutional mandates, our institutional IRB, compliance and data security administrators have instituted stricter oversight procedures for data sharing, which we have been working through for the last 3 months. We preclude reidentification by removing patient information from DICOM headers and removing facial bones/soft tissue from all scans (no skull stripping). We’re optimistic this will satisfy our new compliance responsibilities and allow us to release DV before the paper is published

Methodological novelty (R1,R2,R3): while we don’t propose a new architecture/loss function, DV is the largest CTA dataset for brain vessel segmentation and the first to use Dynamic 4D-CTA to simplify & scale up the annotation process. We appreciate R1’s encouragement to position our contribution as data-centric and will increase the emphasis on this in the revised version

Limited ablations (R2): we did dataset-level ablations: Figs. 4 & 5 show ablations of the impact of contrast at inference (tested on all phases) and the impact of training on a single image per patient (at peak vein/peak arterial contrast). We hope these results will address R2’s concerns about ablations

Metrics (R1): R1 is right: mDC can be “gamed” by over-segmenting a scan. However, the adHD metric will severely penalize cases where the boundaries of the predictions are far from the targets, as it sums minimum distances from each GT-surface point to the nearest predicted surface point. Our annotations cover more vessels than existing datasets, so traditional metrics wouldn’t convey the impact of our approach on the O.O.D. data. Still, we present extra metrics for binary vessel segmentation with the DV model on 3 datasets:

Metric DVtest VVerse TopBrain Dice 0.64 0.20 0.30 Precision 0.55 0.12 0.18 Recall 0.79 0.65 0.90

Qualitative analysis reveals that FNs are infrequent while FPs mostly comprise very distal/scalp vessels (Fig 7c) and bright background objects that can be removed with a brain mask. In addition, we will add statistical significance tests comparing the DV model to all baselines in the revised version

Data quality assurance (R1,R2):The 16cm detector used for the Dynamic-4D-CTA protocol means the patient experiences little motion during the 60 second data acquisition, minimizing inter-phase patient motion. We also use intra-patient rigid registration (validated by a radiologist) between the peak arterial phase and other phases to ensure all images are properly aligned (Sec 2.2). We’ll detail the radiologist’s quality verification procedure for vessel annotations and discuss the limitations due to the lack of inter-rater agreement metrics in the camera ready

Single-phase annotations (R2): Dynamic-4D-CTA acquires data for all vascular phases, allowing retrospective selection of optimal peak phases per patient. We ensure all relevant vasculature parts are highlighted in the peak phases by including only healthy patients with symmetric vascular trees and no occlusions. Evaluation on cases from TopCoW (e.g., ct_001) that have partly occluded or non-opacified arteries confirmed that DV nnUNet segmentation accuracy is robust to such artifacts and pathologies

Clarity (R3):We will improve clarity and accuracy in the abstract (incl. task relevance), the description of iCafe, and variable definitions




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 presents a data-centric approach for brain vessel segmentation in CTA by leveraging dynamic 4D-CTA to construct a multi-phase annotated dataset and improve robustness to contrast variability.

    Reviewers agree that the dataset and annotation pipeline are valuable and practically relevant, but raise major concerns about limited methodological novelty, unclear validation of annotation quality, and limitations in evaluation metrics and generalization claims. I suggest that the authors address these concerns in the rebuttal.

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    The paper presents limited methodological innovation, but reviewers agree that the data contribution could be valuable for the community. I could recommend provisional acceptance, but strictly conditioned on the authors making the dataset publicly available. If the dataset and annotations cannot ultimately be made publicly available, the paper should not be accepted.



Meta-review #2

  • 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.

    The rebuttal did not address all the reviewers’ concerns, being the main concern the lack of methodological novelty of the work presented. Certainly the preparation of the dataset and the dataset per se are valuable, but without being able to access the whole dataset, and without accompanying it with relevant disease-related clinical information, its impact and usefulness may be limited. Suggest this paper being more appropriate for one of the workshops of the conference rather than for the mainstream track. In addition, in the rebuttal it is stated that statistical significance could be added, but it does not specify which tests of statistical significance authors are referring to or whether they already have done these analyses.



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.

    The paper receives mixed reviews (2 positives and 1 negative). Although this work does not provide a new method, however, it provides a valuable and well-curated 4D-CTA dataset and demonstrates the benefits of large-scale, multi-phase data for brain vessel segmentation. Authors promise to open-source this dataset, which will be valuable for the MICCAI community. I recommend acceptance.



back to top