List of Papers Browse by Subject Areas Author List
Abstract
Domain randomization methods have shown strong cross-domain robustness in neuroimaging, yet their use for 3D cerebrovascular segmentation remains limited. Existing vascular synthetic pipelines largely model vessel morphology and appearance in isolation, with limited explicit control of organ-level anatomical context during image generation. We propose an anatomy-conditioned DR framework for zero-shot 3D cerebrovascular segmentation using only synthetic training images that addresses these challenges through three task-specific components: anatomically consistent background context, morphology-aware vessel enrichment, and contrast-aware randomized label-to-image synthesis. A 3D CNN is trained exclusively on synthetically generated images and corresponding vessel masks, with no target-domain images or masks used for training or model selection. Under strict zero-shot evaluation on unseen TOF-MRA and CTA datasets, our method improves over a vessel-centric synthetic baseline and matches or exceeds a vessel-specific foundation model on CTA, while remaining competitive on TOF-MRA. These results show that anatomy-conditioned DR can provide a label-efficient alternative to large foundation-model pipelines for robust zero-shot 3D cerebrovascular segmentation. Our code is available at \url{https://github.com/erc-caravel/anatomical_vessel_dr}
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2678_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/erc-caravel/anatomical_vessel_dr
Link to the Dataset(s)
N/A
BibTex
@InProceedings{PenMat_AnatomyConditioned_MICCAI2026,
author = { Pentassuglia, Matteo AND Zhang, Xiaoming AND Billot, Benjamin AND Zuluaga, Maria A.},
title = { { Anatomy-Conditioned Domain Randomization for Zero-Shot 3D Cerebrovascular Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16881},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes an anatomy-constrained domain randomization (DR) framework for zero-shot 3D cerebrovascular segmentation. The framework does not rely on target-domain images or annotations and is trained exclusively on synthetic images. Its core components include: providing anatomically consistent background context using whole-brain label maps; performing morphology-aware vessel enhancement using murine vascular data; and applying contrast-aware randomized label-to-image synthesis.
- 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, Methodology approach: Unlike traditional vessel-centric synthesis methods, this paper constrains the synthesis process using whole-brain anatomical label maps to preserve realistic anatomical spatial organization.
2, Data utilization: It introduces graph-based murine vascular corrosion cast data to enhance vessel morphology, which may help address the scarcity of high-quality human microvascular annotations.
3, Evaluation protocol: The study performs strict zero-shot evaluations on unseen TOF-MRA and CTA datasets, it may demonstrate the model’s robustness to modality shifts and varying acquisition protocols.
4, Clinical potential: The method provides a label-efficient alternative for clinical centers lacking annotated data.
- 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, Over-simplified anatomical constraints: Although 9 classes of brain tissue labels were extracted, they were merged into two groups (CSF/ventricles and solid tissues) during the intensity sampling stage. This simplification weaken the constraining effect of fine anatomical structures on the synthetic images.
2, Differences in biological priors: Patching murine vascular data into human brain anatomical maps, despite the use of an intracranial mask constraint, still presents inherent differences in topological structure and vessel distribution density.
3, Limited performance on high-resolution modalities: On 7T TOF-MRA (the SMILE-UHURA dataset), the method’s Dice score (68.2) is slightly lower than the vesselFM-full baseline (70.8).
4, Absolute metrics on CTA need improvement: Although the method outperforms the synthetic baseline, the Dice score on TopCow24 CT is only 41.3.This indicates remaining challenges with false negatives or false positives when handling strong domain shifts.
- 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
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?
Recommendation: Weak Accept.
This paper proposes a practical anatomy-constrained DR framework to addresse the challenges of extreme class imbalance and modality generalization in cerebrovascular segmentation.
Although the Dice metrics on specific tasks like 7T MRA do not fully surpass those of heavily annotated foundation models, and there are simplifications in merging anatomical structures, its performance under the “strict zero-shot” protocol remains competitive.
Specifically, it shows significant improvement over existing baselines on the CTA modality.
- 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 explore domain randomization for cerebrovascular segmentation to enable robust and more accessible adaptation of segmentation models to new domains without the need of acquiring annotations. The proposed method infuses anatomical priors from T1-weighted imaging and morphology-aware vessel enrichment from murine corrosion-case data into the synthesis process. Training a DynUNet segmentation model solely on synthetic data yields on-par results with a foundational vessel model (vesselFM) and supervised baseline on two datasets (CereVess-MRA, SMILE-UHURA) and significantly outperform both on a third dataset (TopCoW24 CTA).
- 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.
Heterogeneous testing: diverse angiography settings included such as both TOF, CTA as well as a dataset 7 Tesla TOF. Even though slightly limited in number of subjects, this heterogeneity allows reliable assessment in realistic domains.
- 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.
Applicability: while the proposed anatomical conditioning seems to aid performance gains, it states slight limitation in application. Namely, to train the synthetic-only segmentation model in a given domain, paired anatomical and vessel imaging (e.g. T1w + TOF/CTA) needs to be available subject-wise. The method has shown promising generalization in segmentation performance, it might allow for training in a close domain where paired is available, then translation to target domain; thus is not a crucial limitation.
- 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
Statistical significance: Authors could show statistical significance in results table (e.g. with *) besides the text, and revise the text for consistent reference to significance (for example “(paired Wilcoxon p<10−6)” vs. “(p=0.0676)”)
Inclusion of time-constraints of on-the-fly synthesis could improve transparency and expectations about application of the approach in future research
- While clDice is a great choice, inclusion of a distance-based metric (e.g. Average Hausdorff distance) could further improve robustness of evaluating segmentation performance of thin vessels (Dice is overpowered by bigger structures)
In future work, I would recommend
- Include TopCoW-MRA in evaluation -> especially interesting to see ablation results of “anatomy-constrained vessel enrichment” which was deemed essential for TopCoW-CTA
- Benchmark against a supervised nnUnet -> might outperform DynUNet, thus could be a stronger baseline
- 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 an important contribution for neurovascular research, with a small but intuitive novelty upon existing frameworks. The methodology can be easily adapted and thus could enable wide-spread impact in accelerating methodological and clinical research concerning cerebral vessels. Experiments are robust and results are well presented as well as properly interpreted and discussed.
- 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
The authors present a carefully designed domain randomization framework for cerebral vessel segmentation that uses only synthetic data. Main technical contributions include the use of anatomical constraints and a vessel specific synthesis strategy (donor vessel enrichment). The authors evaluate their method in zero shot fashion on external 7T TOF MRA and CTA datasets and compare it to a supervised baseline and a vessel foundation model. The proposed method outperforms both baselines in the setting with the strongest domain shift.
- 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.
- Builds successfully on and improves upon ideas from recent literature (e.g. vesselFM). The method outperforms vesselFM in the CTA domain where the domain shift is strongest, supporting the main message of the paper that the proposed method helps generalization.
- The donor-vessel enrichment idea of using high resolution non-human vessel data to improve small vessel performance in human data is novel and authors show in ablations it is helpful.
- The authors use anatomical constraints like intracranial masking and show the modality dependent effect (improves CTA performance more).
- The experimentation and ablations are quite extensive, including statistical testing for superiority claims, and show the benefit of domain randomization on unseen domains 7T MRA and CTA.
- 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.
-One limitation of the proposed domain randomization approach is that it has lower performance on the original training modality (TOF-MRA), with a lower Dice score than standard supervised training on the CereVess dataset. -In addition, the method requires paired T1-weighted and TOF-MRA data, which may limit its applicability in settings where such paired data are not available.
- 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
-The authors should define abbreviations at first use, for example GMM on page 5 and CSF on page 4. -The authors could, for example in future work, provide supervised in-domain baselines on the target datasets to improve the interpretation of the results. For example, training on CTA and evaluating on TopCoW CTA would help readers understand the achievable performance with labeled target-domain data and the gap between the proposed zero-shot method and expected performance under domain-specific training.
-The clarity of the paper could be improved. The framework is quite complex, and concepts such as vessel overlay and “donor” patches are currently difficult for the reader to understand.
-Could the authors discuss alternative ways to create the seed maps in cases where paired T1 and angiography data are unavailable? How does the method translate to settings where CTA data are available but the TOF-MRA data are out of domain? More generally, which aspects of the method are transferable or reusable in other training settings?
-Could the authors clarify the intended application of this method and discuss the tradeoff between in-domain performance and cross-domain generalization? In particular, under what conditions would the loss in source-domain performance be acceptable?
-The authors use exclusively synthetic data. Is there a reason for omitting the real TOF-MRA data from the IXI dataset? I could be misunderstanding this but including real data alongside synthetic data could help recover some of the lost performance on the CereVess evaluation set.
- 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 is addressing the well known problem of generalization in cerebrovascular segmentation in a targeted and careful way with novel ideas, a nice evaluation setup and shows an outperforming solution in the most crucial setting where the domain shift is strongest. The lack of within-domain supervised baselines for 7T TOF MRA and TopCow CTA is understandable (from the point of view that this is a paper investigating generalization) but prevents the reader from assessing to what extent the proposed method “solves” the generalization problem. Authors could also better clarify the intended use, advantages and tradeoffs of the proposed method especially since binary labels in the CTA domain are readily available.
- 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 meta-reviewer for their constructive comments and for recognizing the relevance, technical soundness, and zero-shot generalization potential of our anatomy-conditioned domain-randomization framework. In the camera-ready version, we clarified the method, improved statistical reporting, and better positioned the intended use of the approach while preserving the experimental scope.
Methodological clarifications. We clarified the role of anatomical priors and tissue grouping during synthesis. Dense whole-brain label maps preserve organ-level spatial context, while grouping the nine tissue classes is a contrast-aware design choice for angiographic imaging, where vessel visibility is driven mainly by flow or contrast enhancement rather than detailed T1/T2-like tissue boundaries. Thus, anatomical labels constrain spatial organization, whereas intensity sampling is simplified to avoid overfitting to tissue contrasts that are not consistently informative across TOF-MRA and CTA.
Applicability and seed-map construction. We clarified that paired T1w scans are used to maximize seed-map quality, but the framework only requires vessel-aligned anatomical labels. For MRI without paired structural scans, approximate labels from contrast-robust or modality-appropriate parcellation tools can still condition synthesis, potentially with reduced parcellation quality.
Synthetic-only training and intended use. We clarified that the goal is not to maximize performance on source-like domains, where supervised or hybrid real-synthetic training may be preferable, but to study strict zero-shot transfer to domains where annotations are unavailable, costly, or inconsistent. Keeping training synthetic-only isolates the contribution of anatomy-conditioned DR without using real target-domain images or annotations.
- Statistical reporting. We revised statistical reporting for consistency. The ablation section now uses uniform p-value notation and states the paired Wilcoxon test once. Table 1 now includes significance markers: a superscript * indicates that the best-performing method significantly outperforms all others in the same dataset/metric column using paired two-sided Wilcoxon tests with Bonferroni-corrected p_corr < 10^-2.
Runtime transparency. We added runtime information to the Implementation Details section: full sample generation takes 0.33 s per 128^3 patch on average, with observed times ranging from 0.24 to 0.65 s.
Presentation and terminology. We corrected abbreviation definitions at first use, including CSF and GMM, and refined the descriptions of seed maps, donor-vessel patches, and anatomy-constrained vessel enrichment.
Additional evaluation suggestions. We agree that distance-based metrics, TopCoW-MRA evaluation, supervised in-domain target baselines, and supervised nnU-Net comparisons would further contextualize performance. These additions require experiments beyond the scope of the camera-ready revision and are left for future work.
- Finally, we acknowledge that murine donor vessels are an imperfect biological prior and that absolute CTA performance likely remains below fully supervised in-domain training. Nevertheless, in the present strict zero-shot setting, anatomy-constrained donor patches provide useful morphology enrichment, and CTA performance substantially improves over available zero-shot baselines under the strongest modality shift. We view reducing this biology gap and improving absolute CTA performance as important future directions.
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 assigned a score of 4 (Weak Accept), with no negative recommendations. The paper addresses a clinically important problem — zero-shot 3D cerebrovascular segmentation across modalities — and proposes a novel anatomy-conditioned domain randomization framework. The key innovations, including whole-brain anatomical constraints, morphology-aware vessel enrichment using murine vascular data, and contrast-aware randomized label-to-image synthesis, are well motivated and technically sound. The zero-shot evaluation on unseen TOF-MRA and CTA datasets demonstrates strong cross-modal generalization, particularly on CTA where domain shift is most severe. The concerns raised by reviewers — anatomical simplification, reliance on paired data, the trade-off between source-domain performance and generalization — are either design choices or inherent limitations of the problem setting, and do not undermine the core contribution. Given the consensus across reviewers and the absence of any fatal flaws, the paper is accepted without requiring a rebuttal.
