List of Papers Browse by Subject Areas Author List
Abstract
Computed tomography angiography is widely used to visualize vascular anatomy. Accurately distinguishing the inner lumen from the outer vessel wall is essential for many clinical and research tasks, yet in routine practice this delineation is typically performed manually by radiologists. Recent deep learning approaches aim to automate this process, but they generally rely on large, and often private, manually annotated datasets, incurring substantial labeling cost and observer-dependent variability. To address these limitations, we propose a pipeline that is pre-trained on a large set of synthetically generated contour annotations (requiring no manual labeling) and then fine-tuned using a small set of manually corrected data. We evaluate the method with a multi-annotator study to benchmark performance against human variability. After fine-tuning with n=400 cases, our method achieves a mean MAE of 0.493 mm (vs. Reader 1) and 0.643 mm (vs. Reader 2), closely matching inter-annotator variability (0.500 mm). We also report labeling cost, where n=400 corresponds to approximately 6 hours of correction time in our setting. We also analyze performance as a function of variable n manual supervision. In a blinded preference test, an independent radiologist selected the model prediction over the human annotation in 56.8% (vs. Reader 1) and 67.6% (vs. Reader 2) of comparisons. Overall, these results indicate that the proposed approach substantially reduces annotation effort while approaching observed inter-reader variability, enabling scalable and reproducible lumen and vessel-wall delineation for downstream analysis with limited manual supervision.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5035_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{BraJac_Reducing_MICCAI2026,
author = { Bracci, Jacopo AND Katzmann, Alexander AND Nguyen, Tri-Thien AND Sühling, Michael AND Maier, Andreas},
title = { { Reducing Expert Annotation Effort for CTA Vessel Delineation with Synthetic Pretraining } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16880},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
his paper addresses the labeling cost in CTA vessel delineation using a two-stage pipeline based on large-scale pretraining on synthetic annotations followed by fine-tuning with a small amount of manually corrected data. The study also provides a relatively systematic evaluation of annotation cost and inter-reader consistency.
- 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.
This problem setting of this paper has practical value. However, the technical novelty is limited and the main components of the proposed framework are established in the existing literatures.
- 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 technical novelty of this paper is limited. This work is more like a clinical application and engineering validation. Its main contribution is reducing annotation cost and supporting practical vascular analysis workflows, rather than introducing a substantially new method. For this reason, it may be better suited to a clinically oriented medical imaging journal. (2) Although the authors show that a small amount of manual correction can improve performance, it remains unclear whether such limited fine-tuning is sufficient to address the systematic bias introduced during synthetic label generation. Since synthetic labels may contain structural errors rather than simple random noise, large-scale pretraining on them could cause the model to inherit the bias of the original automatic method. Given the limited fine-tuning data and the freezing strategy used, the current results seem to reflect partial calibration rather than a thorough correction of this bias. (3) There is already lots of work on addressing the challenges of insufficient labels, such as pseudo-labeling, semi-supervised learning, and self-supervised learning. I suggest that the authors expand the related work in the Introduction and include comparisons with existing methods in the experimental section. (4) The proposed method is validated on only a single dataset. It would strengthen the paper to include experiments on at least one public dataset to further demonstrate the generalizability and robustness of the proposed approach. (5) Although the manuscript includes a blinded preference study involving radiologists, the contour annotations used for training and quantitative evaluation still appear to be primarily derived from two medical students rather than expert annotations or expert consensus. Therefore, the current results more convincingly show that the method approaches the level of non-expert corrected annotations, but are still insufficient to support a claim of expert-level performance or clear clinical applicability. If the authors wish to emphasize “human-level accuracy” or clinical usability, they should evaluate on expert annotation (gold standard).
- 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 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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper addresses a clinically relevant problem and has potential practical value. However, its methodological novelty is limited, as the pretraining–fine-tuning pipeline is already well established and has been widely applied across many fields. In addition, the current experiments are insufficient to demonstrate that limited fine-tuning can fully correct the systematic bias introduced by synthetic labels. Moreover, the lack of expert gold-standard annotations makes the claims of “human-level accuracy” and clinical applicability insufficiently supported.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
The rebuttal clarifies the clinical motivation and acknowledges several limitations. However, my main concerns remain insufficiently resolved. First, the evidence does not fully support the claim of human-level accuracy, since training and quantitative evaluation are mainly based on contours corrected by two medical students rather than expert-consensus radiologist annotations. The blinded radiologist preference study is useful, but preference does not provide a gold-standard quantitative reference. Second, the concern about synthetic-label bias remains. The reported improvement after fine-tuning shows practical calibration over synthetic-only training, but it does not demonstrate that systematic structural bias inherited from automatic contour generation has been thoroughly corrected, especially given the limited manual data and the freezing strategy. Finally, the lack of broader comparisons and external validation remains a significant limitation. Although I appreciate the clinical relevance of the work, the current evidence is still insufficient to support the main claims. Therefore, I maintain my reject recommendation.
Review #2
- Please describe the contribution of the paper
The paper presents a novel method for reducing annotation burden for CTA vessel delineation using synthetic pre-training. The main contributions of the paper can be summarised as follows: 1.A two-stage, label-efficient training strategy for CTA lumen and vessel-wall delineation, combining synthetic pre-training with limited manual fine-tuning. 2.Benchmark performance against human variability via a multi-annotator study. 3.Quantify performance as a function of the size of fine-tuning data points.
The authors report plateauing performance with only a few fine-tuning cases, indicating the efficacy of the synthetic pre-training step. Frequent preference for the automatically generated delineations in a blinded reader study supports the effectiveness of the proposed 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.
Strengths: 1.The paper addresses a clinically relevant problem of the high cost associated with lumen and vessel wall annotation in CTA. 2.The proposed geometric formulation is elegant and innovative. 3.The experimental design is strong and thorough. Performance is evaluated as a function of annotation budget, and repeated random subset sampling improves robustness. 4.The evaluation is multi-faceted, spanning quantitative assessment using clinically meaningful metrics (MAE, RMAE, Dice, Hausdorff distance) and qualitative analysis through inter-reader variability and a blinded preference study.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
Weaknesses: 1.A real-only baseline is missing from the provided results, limiting the ability to fully contextualise the impact of synthetic pre-training. 2.Variability across a small number of annotators may inflate inter-reader disagreement and affect model–human comparisons. 3.The method depends on an accurate centreline patch, the impact of errors during patch extraction is not discussed. 4.The method depends on an automated method for lumen and vessel-wars contours. The impact of the quality of segmentation is not discussed. 5.Table 1 is presented but not referenced in the text.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
The authors are encouraged to take the above weaknesses as suggestions to imp
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper presents a well-motivated and clinically relevant approach to reducing annotation burden in CTA vessel delineation. The proposed two-stage training strategy, combined with an innovative geometric formulation, demonstrates strong technical insight on the part of the authors. The central hypothesis is supported by thorough and carefully designed experiments, with evaluation spanning both quantitative and qualitative assessments.
While certain limitations remain, including the dependence on accurate patch extraction and the quality of contour generation, these should be more explicitly addressed in the paper but do not detract from the overall contribution of the work.
- 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 appreciate the authors’ rebuttal and their acknowledgement of the weaknesses raised during review. While these weaknesses cannot be fully resolved within the rebuttal period, the authors have appropriately clarified the scope of the work and recognised the relevant limitations.
Review #3
- Please describe the contribution of the paper
Segmenting the inner and outer vessel walls is important for vessel assessment, but remains challenging due to reliance on manual annotations, which are often private and subject to observer variability. This study proposes a two-step pipeline: first, pre-training on a large-scale synthetic dataset without human intervention, followed by fine-tuning on a small set of manually corrected data. The method is evaluated both quantitatively and qualitatively against human annotations.
- 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 manuscript is well written and easy to follow, with a clear presentation of both the motivation and the proposed pipeline. 2.The evaluation is particularly strong and comprehensive. Beyond standard metrics such as Dice, the authors compare against human annotations using multiple complementary metrics (e.g., MAE, RAME, HD), and further include a blinded evaluation comparing human and model outputs. This multi-faceted evaluation design substantially strengthens the credibility of the results and provides a more realistic assessment of performance. 3.The paper provides detailed reporting of both synthetic and human-annotated data, including dataset sizes and annotation time. It offers valuable insight into the practical cost and scalability of the approach.
- 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.
Although the method is clearly described, reproducibility is limited since no code or trained model is provided
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
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.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
1.The topic is relevant and potentially useful to the community. 2.The work addresses an important challenge in vessel image analysis.
- 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
Author Feedback
We thank the reviewers for their constructive feedback and for recognizing the clinical relevance (R1-R3), rigor of the evaluation (R2, R3), and value of reducing annotation cost (R1-R3). In response to the meta-reviewer, we clarify that our contribution is methodological with a clinical focus. While the pre-training/fine-tuning paradigm is well established (R1), the methodological contribution lies in combining synthetic contour pretraining with an “elegant and innovative” (R2) geometric formulation, namely contour-based polar radius regression, and low-data manual correction for CTA lumen and vessel-wall delineation. The evaluation reflects the clinical focus by assessing annotation efficiency, model-reader versus inter-reader disagreement, and blinded radiologist preference, rather than segmentation accuracy alone. R1 raises a concern about systematic bias in synthetic labels, and R2 similarly asks how initial automated contour quality affects the method. We do not claim that limited fine-tuning eliminates all such bias. Rather, the submitted experiments quantify its practical reduction: compared with synthetic-only training, fine-tuning improves all reported metrics for both readers and reduces large-error cases. The blinded preference study adds an independent assessment, with fine-tuned outputs often preferred by the board-certified radiologist. Thus, while residual bias is a valid limitation, the results support practical correction toward human delineation behavior rather than reproduction of synthetic labels. We agree with R1 that the broader label-efficient learning literature should be discussed more explicitly. We will expand the related-work discussion to acknowledge that direct comparisons with alternative label-efficient approaches are missing and represent valuable future work, especially given the lack of public paired CTA lumen/wall contour benchmarks. External validation would further strengthen generalizability, as noted by R1.However, to the best of our knowledge, no public dataset readily matches our setup: CTA cross-sections with paired lumen and outer-wall contours, comparable centerline sampling, and reader/annotation-effort analysis. Furthermore, although our dataset is single-site, it spans multiple vascular territories and multiple volumes and is therefore not limited to one narrow anatomy. Regarding reader variability (R2) and reference annotations (R1), we acknowledge that two annotators cannot capture the full range of variability, and that expert-consensus radiologist contours would provide a stronger reference. However, the contour task is performed on 2D cross-sectional slices, enabling high-quality data collection from trained students under a rigorous protocol. While not a substitute for expert-consensus contours, the blinded radiologist study adds clinical validation beyond purely metric-based contour comparison. We appreciate R2’s constructive suggestions. A “real-only” train-from-scratch baseline would further isolate the contribution of synthetic pretraining, and we will state this as a limitation. Regarding centerline patch extraction, we agree that its impact is relevant for fully automatic deployment, and it will be discussed as a limitation. Nevertheless, because the submitted study samples cross-sections from manually annotated centerlines, it intentionally isolates contour delineation, leaving centerline-detection errors outside the current evaluation. We will also add the missing reference to Table 1.Regarding reproducibility, we share R3’s commitment to open research. We are seeking internal clearance to release a simplified model version and a representative subset of the data to ensure the community can reproduce and build upon our findings. Overall we believe the submitted evidence supports a task-specific methodological pipeline that addresses a clinically relevant annotation bottleneck and benchmarks performance against reader variability and blinded radiologist preference
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.
There is a strong disagreement between one reviewer and the other two. Thus, a rebuttal from the authors is needed. Please clarify if your contributions are more on the methodological or application side.
- 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 novelty of the method is relatively minor, but the setting for reducing the annotation time is relevant to the community. The value of the paper lies in its application, not in the method, as the authors claim. As noted in Rev. 1, the method using synthetic labels can pose an issue that must be acknowledged in the limitations. Something that would have strengthened the paper is to release code and pretrained weights for reproducibility.
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.
While the initial reviews showed a stark divergence among the panel, the authors’ comprehensive rebuttal successfully resolved major methodological concerns. Crucially, the authors clarified that their evaluation was executed strictly at the patient level via cross-validation, refuting any threat of slice-correlation data leakage. They also clarified that their pipeline avoids expensive, unscalable biomechanical meshing by using segmentation solely as a soft prior. Although the single-center, 95-patient cohort remains a minor experimental constraint, the model’s performance stability under simulated missingness (up to 30%) provides compelling validation of its architectural robustness.
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 reviewers generally recognize the clinical relevance and practical value of the work. The rebuttal reasonably clarifies the main contribution. I lean slightly toward acceptance, and encourage the authors to address concerns regarding human-level accuracy and supportive validations.