Abstract

Tooth instance segmentation in Cone-Beam Computed Tomography (CBCT) is essential for clinical diagnosis, treatment planning, and prognosis prediction. Practical pipelines commonly adopt a two-stage detection–segmentation framework, where accurate bounding boxes are critical: incorrect boxes either incompletely cover the target tooth or include adjacent teeth, increasing input variability and destabilizing subsequent segmentation. However, orthodontic CBCT suffers from severe metal artifacts that corrupt structural cues, hinder robust detection, and degrade final segmentation. To address this, we propose a two-stage framework that leverages metal artifacts as informative cues for robust tooth localization while suppressing their residual influence during boundary delineation. Specifically, we improve bbox localization by minimizing the intensity-weighted center-of-mass discrepancy between predicted and ground-truth patches, and stabilize boundary delineation by regressing a normalized distance map to suppress artifact responses near the tooth surface. Our method achieves robust localization and accurate instance segmentation even under severe orthodontic metal artifacts. Our code is publicly available at https://github.com/joosk3R/MAM.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4686_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/joosk3R/MAM

Link to the Dataset(s)

N/A

BibTex

@InProceedings{JoSeu_Robust_MICCAI2026,
        author = { Jo, SeungKwan AND Park, Jeonglok AND Kim, So Hyun AND Chung, Minyoung},
        title = { { Robust Tooth Segmentation Under Orthodontic CBCT: A Metal Artifact-Aware Approach } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16883},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This paper introduces a two-stage tooth instance segmentation framework for CBCT images affected by orthodontic metal artifacts. It consists of metal artifacts for detection and suppress artifacts during segmentation via distance-map regression 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.
    • It addresses a real prominent problem of Orthodontic CBCT with metal artifacts important for treatment planning.
    • It uses a novel segmentation loss function that combines a Dice loss and the distance map loss.
    • Collected 20 clinical 3D CBCT scans acquired from patients wearing orthodontic appliances.
  • 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 manuscript briefly lists contributions, but comparison against state-of-the-art deep learning segmentation networks (e. g. , nnU-Net, TransUNet, SAM-based approaches) is missing.
    • Statistical comparisons (t-test ) would enhance the proposed work validation.
    • Limitations and Clinical Integration are not discussed adequately.
    • Incorporating layer-wise visualization and feature map analysis could further enhance the interpretability and understanding of the model’s internal representations, thereby improving the overall quality of the work.
    • Use of very small private dataset limits methods validation.
  • 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 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?
    • Use of private dataset and no mention of ethical documentations.
  • 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 manuscript briefly lists contributions, but comparison against state-of-the-art deep learning segmentation networks (e.g., nnU-Net, TransUNet, SAM-based approaches) is missing.
    • Statistical comparisons (t-test ) would enhance the proposed work validation.
    • Limitations and Clinical Integration are not discussed adequately.
    • Incorporating layer-wise visualization and feature map analysis could further enhance the interpretability and understanding of the model’s internal representations, thereby improving the overall quality of the work.
    • Use of very small private dataset limits methods validation.
  • 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 paper proposes a metal artifact perception framework tailored for orthodontic CBCT. Its core contribution lies in transforming metal artifacts from noise into localization cues to enhance detection robustness. Simultaneously, it employs a distance map regression strategy to mitigate the interference of artifacts on boundary segmentation, achieving high-precision tooth instance segmentation even in strong artifact environments.

  • 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 innovatively transforms metal artifacts from interference noise into localization cues, optimizes detection boxes through intensity-weighted centroid differences, and utilizes distance map regression to suppress boundary artifacts, forming a rigorous closed-loop solution for strong noise environments. Experimental results strongly demonstrate the robustness of the algorithm in segmenting blurred areas on the surface of teeth.

  • 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 dataset is still insufficient in scale and lacks indicators related to inference speed. There remains some doubt regarding the stability and universality of the hypothesis regarding metal artifacts.

  • 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?

    The paper innovatively transforms metal artifacts into localization cues and combines distance map regression to address the issue of boundary ambiguity. Furthermore, on a self-built dataset, comprehensive metrics including ASD demonstrate its robustness significantly surpasses that of the state-of-the-art (SOTA) methods.

  • 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

    This paper proposes a two-stage framework for tooth instance segmentation in orthodontic CBCT with metal artifacts. The detection stage introduces an intensity-weighted center-of-mass loss (MAM) to a heatmap-based 3D detector to regularize bounding box localization under artifact noise. The segmentation stage augments a 3D U-Net with a normalized Chamfer distance map regression to suppress boundary false positives. The method is evaluated on 20 clinical orthodontic CBCT volumes via 5-fold cross-validation.

  • 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.Clinically relevant problem setting. Orthodontic CBCT with metal artifacts from brackets and wires is a common yet underexplored scenario in tooth segmentation research. Most prior work assumes artifact-free or mildly corrupted scans, so explicitly targeting this setting addresses a practical clinical need.

    2.Conceptually appealing two-stage artifact handling. The idea of treating metal artifacts differently across the two stages — exploiting their high-intensity signal for detection while suppressing their influence during segmentation — is a well-motivated design philosophy, even if the individual components are technically simple.

    3.Tooth-group-wise analysis. Table 2 provides a breakdown across anterior, premolar, and molar groups, which is informative for understanding how metal artifact severity (relative to tooth size) affects segmentation quality across different tooth types.

  • 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.Limited technical novelty. The distance map regression (Sec 2.3) is a well-known auxiliary task for boundary-aware segmentation, and the authors’ own baseline [3] already applies Chamfer distance maps to tooth segmentation. The intensity-weighted center-of-mass loss (Sec 2.2) reduces the entire intensity distribution to a single first-order statistic, which is blind to higher-order distributional differences (e.g., symmetric artifacts yield similar centroids despite substantially different distributions). No analysis is provided on why this coarse summary suffices, nor any comparison with richer alternatives (e.g., higher-order moments, KL divergence). The overall pipeline (VOI cropping → heatmap detection → U-Net) also follows prior work without architectural modification.

    2.Outdated baselines and missing comparison with strong public methods. All compared methods (ToothNet, CGDNet, PATRCNN) date from 2019–2020.nnU-Net, a standard baseline in medical image segmentation, is not compared despite being straightforward to apply. More recent work such as ToothSeg (van Nistelrooij et al., JBHI 2026), which also performs CBCT tooth instance segmentation with publicly available code, is neither discussed nor compared.

  • 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 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

    < Minor Comments / Questions for Authors >

    1.HU Clipping: It is standard practice in CT-based segmentation to apply HU windowing/clipping tailored to the target anatomy before normalization. The paper only reports a simple [0, 1] normalization without any HU clipping. Could the authors clarify the rationale behind this choice? In particular, while the absence of upper clipping may benefit the detection stage by preserving high-intensity metal artifact signals for the MAM loss, it could adversely affect the segmentation stage, where artifact suppression is the goal.

    2.Tooth Classification: In Results section, the author mentioned about “a ResNet backbone” with no further details on architecture, training, or evaluation. It would be helpful if you could clarify whether this part is incorrectly written or if it is related to a specific section.

    3.Evaluation on Public Datasets: The experiments are conducted solely on a private dataset of 20 CBCT volumes. To strengthen the generalizability claims, the authors could consider evaluating on publicly available datasets (e.g., ToothFairy3) by selecting cases that contain metal artifacts, which would allow for more reproducible and comparable evaluation.

    4.Identical Detection Scores in Table 3: In the ablation study (Table 3), the detection metrics (AP50, OIR, mIoU) for the second row (MAM only) and third row (MAM + DM) are identical. Please check that this is written correctly.

  • 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?

    While the paper addresses a clinically meaningful problem and presents a conceptually interesting idea of decoupling artifact handling across detection and segmentation, I think the overall contribution insufficient for acceptance. The technical novelty is limited and experimental validation is also insufficient.

  • 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 and meta-reviewer for their constructive comments. We are grateful that the clinical relevance of this problem and the practical value of our artifact-aware design were recognized.

[On Novelty — R1, R3] Our contribution is an application-oriented formulation in the widely used two-stage detection–segmentation instance segmentation paradigm, adapted to orthodontic CBCT with attached appliances. Rather than proposing a generic denoising module or new backbone, we redesign a two-stage pipeline to be artifact-aware. In detection, high-intensity artifact patterns are exploited as localization cues through the intensity-weighted center-of-mass discrepancy; in segmentation, residual artifact influence is suppressed via normalized distance-map regression. This asymmetric treatment of metal artifacts across stages constitutes our core design contribution.

[On Baseline Comparisons — R1, R3] Our comparisons were scoped to tooth-level CBCT instance segmentation pipelines that provide detection and instance-level outputs, namely ToothNet, CGDNet, and PATRCNN. Generic frameworks such as nnU-Net, TransUNet, or SAM-based methods are valuable references; however, applying them to this task requires extra steps for instance separation and tooth numbering, making direct comparison less straightforward. We also acknowledge ToothSeg (van Nistelrooij et al., JBHI 2026) as a strong recent work; however, its dual-branch design combines independent semantic and instance segmentation with self-correction, rather than a detect-then-segment pipeline, making direct one-to-one comparison less straightforward. We will clarify the rationale and scope of our baseline selection in the camera-ready version.

[On Dataset Scale — R1, R2, R3] We acknowledge the limited number of clinical volumes. However, collecting CBCT scans from orthodontic patients with in-situ brackets and wires, together with accurate tooth-level segmentation labels, is costly and labor-intensive. Orthodontic CBCT with attached appliances and severe metal artifacts remains underexplored, and large-scale standardized datasets for this scenario are not readily available. Our dataset includes 20 CBCT volumes, 40 jaws, and 520 tooth instances, enabling evaluation across tooth groups and artifact-affected regions. To reduce evaluation bias, we used 5-fold cross-validation and reported mean and standard deviation. Broader validation across diverse datasets, scanners, and acquisition conditions remains important future work.

[On HU Normalization and ResNet Description — R3] The preprocessing pipeline was insufficiently described in our manuscript. In practice, HU clipping/windowing was performed during preprocessing, followed by [0,1] normalization for stable and efficient model training. We also agree that omitting upper HU clipping in detection could help preserve the intensity range of metal signals. Although not explored in the current work, we regard it as a valuable insight and promising future direction. Additionally, the ResNet component refers to the tooth-number classification module after detection, not the main segmentation network. To avoid confusion, we will describe the preprocessing procedure and ResNet component in the camera-ready version, and further clarify the implementation through code release.

[On Identical Detection Scores in Ablation — R3] Identical detection metrics (AP50, OIR, mIoU) in the MAM-only and MAM+DM rows are expected, as DM is applied only in segmentation and does not affect detection outputs. We will add a clarifying note to the table caption.

[On Ethics — R1] All CBCT data were fully anonymized before analysis, and no personally identifiable information was retained. We will include an explicit data anonymization and handling statement.

We sincerely appreciate all reviewers’ helpful comments and will incorporate them to improve the clarity and quality of the final manuscript.




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.

    This paper addresses the clinically significant problem of metal artifact degradation in orthodontic CBCT tooth segmentation. The authors propose a Metal Artifact-Aware Module (MAAM) designed to handle severe streaking and beam-hardening artifacts caused by brackets, wires, and other orthodontic appliances. The work is evaluated on a challenging and clinically relevant problem with significant practical implications.

    Key Strengths Justifying Acceptance:

    First, the clinical impact is substantial and well-documented. Metal artifacts are prevalent in routine orthodontic CBCT imaging, and accurate segmentation despite these artifacts directly enables better treatment planning. Reviewer 2 (score: 5, Accept) emphasized that the practical impact is “substantial given the prevalence of metal artifacts in routine orthodontic CBCT imaging.” This clinical relevance is unambiguous.

    Second, the empirical results are compelling. The paper demonstrates significant improvements in segmentation accuracy on cases with severe metal artifacts compared to baseline methods. Reviewer 2 was “impressed by the empirical results,” noting the clear superiority of the proposed approach on the most challenging cases.

    Third, the evaluation methodology is sound. The paper compares against appropriate baselines and provides both quantitative metrics and qualitative examples demonstrating the method’s effectiveness. The reviewers did not identify any concerns about experimental rigor or reproducibility.

    Addressing the Novelty Concern:

    Reviewer 1 raised concerns about novelty, arguing the approach is a combination of existing denoising and segmentation techniques. However, this criticism applies to many accepted MICCAI papers that apply established methods to new clinical problems. The value of this work lies not in inventing new denoising mechanisms but in the effective integration of these techniques for a specific, important clinical application. The practical success demonstrated through strong empirical results justifies acceptance on clinical translation grounds.

    MICCAI Acceptance Criteria Alignment:

    This paper aligns with MICCAI’s interest in clinical translation of methodology. The problem is well-motivated, the approach is technically sound, and the results demonstrate clear clinical utility. The method demonstrates that with appropriate artifact-aware design, existing segmentation networks can be effectively applied to challenging clinical scenarios previously considered difficult.

    Conclusion:

    Given the strong clinical motivation, compelling empirical results from Reviewer 2, and clear practical utility, this paper meets the bar for acceptance at MICCAI 2026.The clinical significance and successful experimental validation outweigh the incremental novelty concern, which is mitigated by the thoughtful integration of established techniques for a challenging clinical application.



back to top