Abstract

Vascular computed tomography datasets are commonly annotated only once per scan, yielding the pervasive yet under-addressed problem of \emph{single-mask} annotation noise. Existing solutions either require costly multi-rater fusion or coupled with network training, preventing explicit auditing of \emph{where} and \emph{why} labels fail. We introduce a \emph{decoupled} framework for single-mask annotation noise detection that leverages \emph{cross-sectional patch self-consistency} to produce \emph{interpretable} and \emph{auditable} noise evidence. Tubular anatomy exhibits strong cross-sectional recurrence: patches extracted orthogonally along vessel centrelines recur in appearance across locations and subjects. Thus, anatomically similar patches should have consistent masks, and disagreement signals unreliable annotation. Our method samples cross-sectional patches, retrieves intensity-equivalent neighbors via scalable vector search, and computes a \emph{patch-level noise score} from statistical mask disagreement, yielding explicit image–mask evidence for every flagged region. Aggregating scores produces scan-level \emph{quality maps} for dataset quality assessment or quality-weighted training. Experiments on the coronary CT dataset validate the detected noise for improving training robustness and reveal systematic annotation biases, i.e., transverse and oblique vessels exhibit 5.1$\times$ higher error rate than axis-aligned structures, with additional correlations to cross-sectional area and intensity.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1603_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/zhuyinheng/hidden_noise

Link to the Dataset(s)

N/A

BibTex

@InProceedings{ZhuYin_Decoupled_MICCAI2026,
        author = { Zhu, Yinheng AND Xu, Xiaowei},
        title = { { Decoupled Single-Mask Annotation Noise Detection via Cross-Sectional Patch Self-consistency } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16882},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The paper addresses the problem of localising annotation noise in a Vascular Computed Tomography dataset with a single mask or single rater. It proposes a training-free cross-sectional patch self-consistency principle to identify noisy masks. According to this principle, similar cross-sectional patches will have similar masks. Therefore, patches with highly similar appearances should have consistent masks. Quality scores are assigned to voxels based on the consistency level to penalize noisy labels during training.

  • 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 work uses cross-sectional image patches and masks as a reference for comparison, which is a very intuitive and interesting idea. Provides an interpretable and training-free algorithm. Explains the results with qualitative examples, even for failed cases.

  • 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.

    Major: 1.The method was evaluated only on the ImageCAS dataset, limiting its generalizability to out-of-distribution datasets. 2.Even though the paper evaluates the method with and without quality weighting, it does not compare with other similar methods for its performance against existing noisy mask localisation methods, eg: Dong et al. , Deep Self-Cleansing for Medical Image Segmentation with Noisy Labels 3.The use of MSE to compare image patches is not intuitive here, as structural comparison is more important, why not SSIM? Minor: 1.The computation cost of patch comparisons could have been studied. 2.Figure 1 lacks some clarity. The masks for the yellow and brown patches appear at different positions compared to their corresponding image patches. 3.The first paragraph under the Methodology section describes the Multi-rater consistency equation and provides the citation [15]. However, the cited work does not describe this equation. 4.The word “curvenetwork” in the first paragraph of the introduction is missing a hyphen or space after “curve”. 5.The paper does not state the value of “K” used, preventing the reader from recreating the results.

  • 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.

    (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?

    Lack of baselines, even though there are similar methods addressing the same problem. Also some critical hyperparameters aren’t defined to ensuring full reproducability. Improvement in results seems very marginal. The reported results seems to be from only one seed, without any CIs/uncertainty regions.

  • 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.

    The feedback agrees to include the missing parameters/values & analysis (CIs) of the work. More clarity on how the work is distinct is also provided.



Review #2

  • Please describe the contribution of the paper

    This paper proposes a decoupled framework for detecting annotation noise in single-mask vascular segmentation datasets, addressing the common scenario where only one annotation per scan is available.

    The method is based on a cross-sectional patch self-consistency principle, where anatomically similar patches—extracted along vessel centrelines—are expected to have consistent masks. Disagreement between similar patches is used as evidence of annotation noise.

    The framework operates independently of model training and produces interpretable, patch-level noise scores, which are aggregated into scan-level quality maps for dataset auditing and quality-weighted training.

  • 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 addresses a practically important and underexplored problem, namely annotation noise detection in the single-mask setting, which is common in vascular CT datasets. The core idea is intuitive and well-grounded, leveraging anatomical recurrence to define a self-consistency criterion that is easy to interpret. The method is decoupled from model training, which is a strong advantage over training-based noise handling approaches and allows explicit auditing of annotations. The use of cross-sectional patch extraction with Bishop frames is technically sound and ensures stable geometric representation. The framework provides interpretable outputs, where each detected noisy region is supported by explicit patch-pair evidence. The analysis reveals systematic annotation biases (e.g., higher errors in oblique vessels), which is insightful and adds value beyond performance improvement. The method demonstrates practical utility, as quality-weighted training improves boundary-sensitive metrics, indicating that detected noise is meaningful.

  • 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 overall performance improvement is modest, particularly in standard metrics such as Dice, which raises questions about practical impact. The evaluation is limited to a single dataset, making it unclear how well the method generalizes to other anatomies or imaging modalities. The approach relies on accurate centreline extraction, but the sensitivity of the method to errors in this step is not discussed. The similarity metric (MSE) used for patch retrieval is relatively simple and fails in cases involving rotation or intensity variation, as shown in the failure analysis. The method focuses on noise detection rather than correction, which limits its direct applicability in automated pipelines. Some assumptions, such as cross-sectional recurrence across subjects, may not hold equally well in more heterogeneous datasets.

  • 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

    This is a well-motivated and thoughtfully designed approach. The interpretability and decoupled nature are strong aspects. Expanding evaluation across datasets and improving robustness of patch similarity could further strengthen the work.

  • 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 clear and practical solution to annotation noise detection in a challenging single-mask setting. The methodology is well-motivated and produces interpretable results, with additional insights into annotation bias.

    However, the empirical gains are relatively modest, and the evaluation is somewhat limited.

  • 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.

    The authors has clearly defined the raised concern in the review.



Review #3

  • Please describe the contribution of the paper

    1.This work introduces a novel decoupled framework for single-mask annotation noise detection, especially for vascular segmentation.

    2.Based on the physical characteristics of vessels, the paper points out that planes orthogonal to the vessel centerline and the corresponding segmentation masks should exhibit strong similarity, and leverages this strong similarity to identify annotation noise.

    3.Experiments on a coronary CT dataset show the effectiveness of the proposed method in improving training robustness and detecting noisy labels.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    1.The proposed framework addresses the single-mask annotation issue without requiring additional annotations or a large computational burden.

    2.The proposed method fully exploits the geometric characteristics of vessels and their self-consistency to detect segmentation noise.

    3.The proposed method is well aligned with the problem setting, and the solution is also highly interpretable.

  • 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 current work seems to lack comparisons with existing methods for addressing single-mask annotation noise. However, since the reviewer is not an expert in this specific area, it is unclear whether similar work already exists.

  • 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

    1.Some references should be preceded by a space, for example, “model training[21]” and “contrast agent propagation[24]” on Page 1. 2.Some reference formatting appears to be inconsistent, and the authors are advised to check it carefully.

  • 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?

    See above

  • Reviewer confidence

    Somewhat confident (2)

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

    Accept

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

    1.This work holds practical value for addressing and correcting label noise in cardiovascular segmentation. 2.The proposed method is well-motivated and rationally designed, effectively incorporating the physical characteristics of blood vessels. 3.The authors have satisfactorily addressed the reviewers’ concerns and queries in their rebuttal.



Author Feedback

We thank the reviewers and AC for the constructive feedback and thoughtful assessment of our work. We especially appreciate the recognition of the paper as a decoupled and interpretable framework for annotation noise auditing in the practical single-mask setting. We address the major concerns below.

## Comparative analysis We agree that comparisons with existing noisy-label methods provide useful context. However, our framework addresses a fundamentally different objective from existing single-mask noisy-label learning methods.

Existing methods are primarily training-coupled robust learning frameworks whose objective is to improve segmentation robustness on a presumed “clean” evaluation set. Potential annotation noise is inferred indirectly from optimization dynamics (e.g., sample loss, prediction disagreement, feature divergence, pseudo-label refinement). As a result, it is often ambiguous whether a flagged region corresponds to a mislabeled annotation or simply a hard sample during optimization.

In contrast, our framework is fully decoupled from network training and instead formulates annotation inconsistency as violations of anatomical cross-sectional self-consistency directly from the image-mask pair itself. The framework therefore evaluates the consistency of the annotation itself rather than model behavior during optimization.

This difference also changes how the output is consumed. Existing methods mainly identify potential noisy labels for exclusion/replacement within the training loop, whereas our framework provides explicit evidence for human QA: a suspicious patch together with multiple image-near-identical patches exhibiting consistent masks.

Accordingly, our framework provides two key relative advantages in the practical single-mask setting: (i) interpretability/auditability: every flagged region is directly associated with explicit image-near-identical patch evidence that can be inspected by a human QA reviewer; (ii) annotation-centric quality control: the framework directly evaluates consistency of the annotation itself, enabling dataset quality auditing rather than only improving downstream training robustness.

Our goal is therefore not primarily to develop another training-time denoising strategy, but to identify a criterion that correct annotations themselves should satisfy. We will clarify this distinction and add comparative discussion with existing SMND methods in the revised manuscript. Per MICCAI rebuttal policy, we cannot include new experimental comparisons during rebuttal.

## MSE vs. SSIM and centerline sensitivity SSIM is difficult to scale to our setting (~10^6 patches and ~10^12 pairwise comparisons). MSE was chosen primarily for computational feasibility and compatibility with scalable vector retrieval. Its limitations (e.g., rotation/window-level sensitivity) are already explicitly analyzed in the failure cases (Fig. 6).

Similarly, centerline/Bishop-frame errors mainly reduce retrieval recall rather than precision. Our design prioritizes precision (ensuring detected noise is truly suspicious and auditable), while recall can potentially be improved by denser sampling over rotation/shift spaces at increased computational cost. We will further clarify this discussion in the revised manuscript.

## Reproducibility clarifications We will clarify the implementation details requested in the meta-review, including: K=100 (chosen as a trade-off between variance estimation stability and bin locality), clarification of the image/mask manifold illustration in Fig.1, and attribution/clarification of Eq.1 (“segmentations obtained repeatedly from the same image” in [15]). Minor typo/formatting issues will also be corrected in the revised manuscript.

## Statistical uncertainty reporting We will revise the presentation of the existing quantitative results to include statistical uncertainty measures (e.g., confidence intervals / mean±std) for clearer interpretation of the reported gains.




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.

    The paper presents an elegant and interpretable “decoupled” framework for detecting annotation noise in medical imaging when only a single rater’s mask is available. By leveraging the anatomical recurrence of cross-sectional vascular patches to identify inconsistencies, the work addresses a practical challenge with an approach that requires no training overhead. While reviewers generally lean toward acceptance, the primary concerns center on a lack of comparative rigor, specifically the absence of established baselines for noisy label detection. Furthermore, the submission requires additional technical clarity regarding hyperparameter selection and the sensitivity of the method to preprocessing steps like centerline extraction.

    Rebuttal Instructions: The authors must provide a comparative analysis against at least one established baseline for noisy label detection in a single-mask setting to demonstrate a clear relative advantage. To ensure reproducibility, the authors are required to provide missing hyperparameter values (specifically “K”), clarify discrepancies in Figure 1, and address the citation for the multi-rater consistency equation. Additionally, the rebuttal should justify the choice of MSE over structural metrics like SSIM for patch similarity and include a discussion on the framework’s sensitivity to errors in the centerline extraction process. Finally, authors should provide statistical uncertainty measures, such as confidence intervals, to confirm that the reported performance gains are meaningful.

  • 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.

    Decoupled Single-Mask Annotation Noise Detection presents a highly practical, training-free framework that effectively addresses single-rater label inconsistencies by exploiting the geometric and cross-sectional recurrence of tubular vascular structures. The authors successfully addressed initial concerns regarding comparative baseline rigor and hyperparameter sensitivity by clarifying that the framework acts as a decoupled, interpretable auditing tool for human quality assurance rather than a standard training-coupled optimization pipeline. A unanimous post-rebuttal consensus to accept (Scores: 3 $\rightarrow$ 4, 5, 5) and concrete commitments to incorporate critical hyperparameter definitions ($K=100$) and statistical uncertainty measures, this work offers a valuable and reproducible contribution to dataset curation in medical imaging.



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.

    The paper proposes a simple yet effective method for identifying noisy regions within CT vessel annotations by measuring the consistency between appearance changes and mask shape changes across adjacent frames. The approach is well-motivated, and the authors further demonstrate that explicitly accounting for the identified noisy annotations during training yields measurable improvements in downstream segmentation performance. The reviewers converged on a positive recommendation, and I concur with this consensus in supporting acceptance.



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 presents a practical and interpretable framework for annotation noise detection in the single-mask setting. Reviewers highlighted the decoupled design and anatomical motivation as strengths, while concerns focused on comparisons and reproducibility details. The rebuttal addressed most of these points. Based on the reviewer feedback and rebuttal, the paper is above the acceptance threshold.



back to top