List of Papers Browse by Subject Areas Author List
Abstract
Accurate gingival segmentation from Cone-Beam Computed Tomography (CBCT) is critical for periodontitis management and implantology, yet remains challenging due to low soft tissue contrast and the adhesion of adjacent tissues such as the cheeks and the tongue. In standard CBCT, the gingival margin is essentially indistinguishable from the surrounding mucosa, so existing methods either rely on additional intra-oral scan (IOS) data or apply voxel-based deep-learning methods that struggle with boundary precision. To address these limitations, this paper proposes a protocol-and-method co-design. First, we introduce a chair-side scanning protocol that uses a mouth retractor and an impression tray to physically isolate the gingiva from buccal and lingual tissues during acquisition, creating a continuous high-contrast air interface that makes the gingival margin imageable in CBCT in the first place; the protocol requires no scanner modification and adds less than one minute to the standard CBCT workflow. Second, building on this controlled input, we propose a cross-surface representation (CSR) that recasts the 3D volumetric segmentation as 1D radiodensity profile regression along tooth surface vertex normals, together with a band-aware Gaussian regression (BGR) module that produces a continuous, topologically consistent confidence field on the tooth manifold. Experiments demonstrate that the method outperforms strong 3D voxel baselines, achieving expert-level accuracy without auxiliary IOS data.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2720_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{MeiLan_CrossSurface_MICCAI2026,
author = { Mei, Lanzhuju AND Li, Yuan AND Wu, Han AND Gu, Wen AND Wang, Chong AND Tonetti, Maurizio AND Cui, Zhiming AND Shen, Dinggang},
title = { { Cross-Surface Representation Learning for Accurate Gingiva Prediction in CBCT } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16889},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper addresses CBCT-based gingival margin prediction, a clinically important task for periodontal assessment and implantology that is severely limited by low soft-tissue contrast in standard CBCT acquisitions. The paper makes a dual contribution: first, a standardised physical scanning protocol using a mouth retractor and impression tray to create a continuous high-contrast air interface between the gingiva and surrounding oral tissues; second, a cross-surface representation (CSR) that transforms the inherently 3D volumetric segmentation problem into structured 1D radiodensity profile regression along tooth surface vertex normals, combined with band-aware Gaussian regression (BGR) to produce a smooth continuous confidence field for the gingival margin. Evaluated on 28 clinical cases (684 teeth), the method substantially outperforms voxel-based baselines including nnU-Net across all four tooth categories.
- 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.Physical protocol co-design is the defining innovation. 2.Cross-surface representation is anatomically motivated and principled 3.Band-aware Gaussian regression avoids discrete classification
- 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.Applicability to standard (non-protocol) CBCT scans is uncharacterised. 2.Per-fold breakdown is absent 3.Normal direction limitations for multi-rooted teeth are not discussed
- 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.
(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?
This paper makes a compelling Clinical Translation contribution by co-designing a physical scanning protocol and a surface-normal-anchored algorithmic pipeline that together enable CBCT-based gingival margin prediction at near-expert accuracy, without auxiliary IOS data. The hardware-software integration, the cross-surface representation, and the consistent multi-tooth-category results are all strengths. The main weakness, lack of characterisation on non-protocol scans, is an important limitation for clinical translation and should be addressed in discussion, but it does not diminish the contribution of demonstrating that the method works well under the proposed protocol.
- Reviewer confidence
Not confident (1)
- [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 main contribution of this paper is that it does not directly segment the ambiguous gingival boundary in CBCT using a conventional voxel-based formulation, but instead reformulates the task on a tooth-surface geometric prior through a cross-surface representation based on surface-normal intensity profiles. Combined with Gaussian soft-band regression, the method is designed to more stably predict a thin and ambiguous boundary. The standardized physical scanning protocol, which improves the separability of the input itself, is also a practical contribution.
- 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 CSR idea is anatomically well motivated and, to the best of my knowledge, is a novel and well-matched formulation for this task, making it more convincing than simply applying a larger segmentation model.
2.The performance gains are non-trivial. In particular, the improvements over nnU-Net in both MAE and IoU suggest that the proposed representation provides a meaningful practical advantage.
3.Figs. 1–3 effectively connect the clinical problem, the methodological design, and the resulting performance, making the overall message easy to follow.
- 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 main weakness is the limited evidence for generalizability. Since all data were acquired using the authors’ specialized physical scanning protocol, it remains unclear whether the method would perform similarly on routine clinical CBCT scans. As presented, the reported gains appear to reflect a combination of algorithmic improvement and acquisition-specific advantages.
2.The evaluation scale is limited. Although 684 teeth may appear substantial, the study includes only 28 patients, and it is not clearly stated whether the data split was performed at the patient level.
3.The experimental comparisons are somewhat limited, as the baselines are mainly older voxel-based models. The evaluation would be more convincing if it also included relatively recent and stronger 3D segmentation baselines, such as 3D UXNet or SwinUNETR, to better position the proposed method against the current state of the art.
4.Reproducibility and ethical reporting are insufficiently documented. Important details such as data acquisition approval, IRB/consent, the split protocol, and the training/evaluation setup for tooth segmentation are not described clearly enough.
- 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?
This paper addresses an important clinical problem, and the CSR+BGR framework built on a tooth-surface prior is well aligned with the structure of the task. The quantitative results also show clear improvements over the baselines. However, by MICCAI standards, the evidence for generalizability and the overall evaluation design are not yet fully convincing. In particular, all data were collected under a specialized acquisition protocol, the comparisons remain mostly limited to voxel-based baselines, and the dataset scale and split protocol are only sparsely described. Overall, the idea is interesting and has clear potential, but in its current form it is difficult to strongly support acceptance, and the paper feels more like a borderline submission that would benefit from additional validation.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The rebuttal adequately addresses most of my major concerns. In particular, the authors clarify that the physical scanning protocol is a necessary prerequisite for making the gingival margin visible in CBCT, rather than merely an acquisition advantage. They also explain that all baseline methods were evaluated on the same protocol-acquired data and patient-level five-fold split, which helps isolate the algorithmic contribution of the proposed CSR+BGR framework. The additional results with stronger recent baselines such as SwinUNETR and 3D UXNet further strengthen the experimental validation.
Although some concerns remain regarding the dependence on the specialized acquisition protocol, and further validation across broader clinical settings would still be valuable, the rebuttal makes the necessity and scope of the protocol sufficiently clear. The authors also address the concerns about patient-level splitting, ethics approval, anonymization, and reproducibility details. Overall, I find that the rebuttal resolves the main issues that led to my original weak reject recommendation. Therefore, I am willing to change my recommendation to accept.
Review #3
- Please describe the contribution of the paper
This work presents a novel framework that combines specialized physical scanning protocols with geometry-prior-driven deep learning to achieve accurate gingival margin prediction in cone-beam computed tomography (CBCT). The main contributions are as follows: (1) the introduction of a standardized physical scanning protocol that uses a mouth opener and impression tray to physically isolate the gingiva from the buccal and lingual tissues during scanning, creating a high-contrast air interface; (2) the proposal of a Cross-surface Representation (CSR) that transforms the 3D voxel segmentation task into a 1D intensity distribution analysis along the surface normals of teeth; and (3) the design of a Band-aware Gaussian Regression (BGR) module that predicts a continuous confidence field on the point cloud manifold, providing an alternative to traditional discrete voxel classification.
- 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.Pragmatic clinical insight: Instead of solely relying on algorithmic solutions to address the low soft-tissue contrast in CBCT, this study innovatively introduces clinical physical aids—a mouth opener and impression tray—to reshape the imaging boundaries. This front-end approach to improving data quality demonstrates practical thinking and strong clinical awareness.
2.Novel task formulation: The unconstrained 3D voxel segmentation problem is elegantly reformulated as a scalar field regression task constrained by geometric priors derived from the tooth surface mesh. Through CSR, the method extracts features along the tooth-surface normal direction corresponding to the “bone–gingiva–air” transition, effectively reducing background anatomical noise. This formulation offers theoretical advantages in handling adherent soft-tissue edges.
3.Empirical validation: Experiments show that the proposed approach achieves a mean absolute error (MAE) of 0.38 mm without requiring auxiliary intraoral scan (IOS) data for registration. This has the potential to simplify clinical periodontal assessment workflows and reduce cumulative errors from multi-modal registration.
- 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 method’s strong reliance on upstream geometric priors renders it potentially sensitive to error propagation. Both the CSR and BGR modules are fully dependent on the accuracy of the tooth surface meshes extracted by nnU-Net. However, in real-world CBCT clinical data, metal artifacts—such as those from crowns or implants—can lead to failed tooth segmentation or severely distorted surface meshes. The manuscript lacks an analysis of how such upstream mesh extraction errors may impact the robustness of downstream gingival predictions.
2.The manuscript compares the BGR network, which leverages point cloud and surface normal features, with conventional 3D CNNs (e.g., 3D U-Net, DeepLabV3+) that operate directly on raw CBCT voxels. While this comparison demonstrates that “surface-constrained” approaches outperform “unconstrained voxel classification,” it does not sufficiently establish the inherent advantage of the Gaussian regression (BGR) itself. To make the evaluation more rigorous, I suggest the authors to include a baseline model that also operates on the tooth surface manifold—e.g., a PointNet++ or graph neural network performing direct binary classification on the point cloud.
3.Although the physical scanning protocol effectively enhances image contrast, the manuscript does not discuss its applicability boundaries in extreme clinical scenarios. For instance, in patients with severe gag reflex, limited mouth opening, or partial tooth loss, the placement of the physical tools may be constrained, which could limit the generalizability of the proposed method.
4.The introduction quickly jumps to the limitations of voxel-based CNNs but overlooks a deeper discussion of the methodological context. Given that the core of this work is the transformation of the 3D volumetric gingiva segmentation problem into a tooth-surface-based manifold problem, I recommend that the authors supplement the Introduction or Related Work sections with a discussion of existing methods for tooth, gingiva, and alveolar bone segmentation in medical image analysis, as well as recent literature on surface-, mesh-, or graph-based learning (e.g., [1][2]). Currently, the manuscript cites only 16 references; adding relevant studies would provide a more comprehensive background.
5.I recommend that the authors supplement the manuscript with figures showing the actual 1D intensity profile curves extracted from the three regions. Additionally, visualizations of the continuous Gaussian confidence fields predicted along the tooth surfaces could be included. Providing these visualizations would greatly enhance the interpretability of the method and increase readers’ confidence in its effectiveness.
[1] Wang C, He Z, Chen L, et al. ITMatch: Arch-Guided Semi-supervised Tooth Arrangement via Iterative Confidence Evaluation[C]//International Conference on Medical Image Computing and Computer-Assisted Intervention. Cham: Springer Nature Switzerland, 2025: 268-277.[2] Zhou, Z., Chen, Y., He, A., Que, X., Wang, K., Yao, R., & Li, T. (2024). NKUT: dataset and benchmark for pediatric mandibular wisdom teeth segmentation. IEEE Journal of Biomedical and Health Informatics, 28(6), 3523-3533.
- 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
Please see weakness above.
- 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?
Addressing key challenges in medical image analysis does not always require stacking increasingly complex network architectures. The authors of this work precisely identify the core difficulty in CBCT gingival segmentation and propose an elegant solution that combines physical pre-scanning isolation with algorithmic post-processing constraints. This geometry-driven representation learning aligns naturally with periodontal anatomical features. Overall, the approach is methodologically sound and represents a compelling and meaningful exploration.
- 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 author’s response to my comments, which to some extent resolved my confusion. Although this paper still has some shortcomings, I believe it is a meaningful work. The author should complete all the promised revisions in the final manuscript, and I recommend accepting this paper.
Author Feedback
We thank the Meta-AE and reviewers for recognizing the clinical motivation, the protocol and method co-design, and the consistent gains over voxe-based methods. We address the major concerns below. (1) Protocol scope and isolation of algorithmic gain [Meta reviewer, R1.1, R2.1]. The protocol is a prerequisite that makes CBCT gingival margin imaging tractable. In routine CBCT, buccal and lingual soft tissues adhere to the arch with nearly identical intensity to gingiva, leaving the margin invisible even to experienced clinicians, and no algorithm can recover information that is not encoded. Our lip retractor and impression tray protocol introduces this air interface as a simple chair side add on to the existing CBCT workflow, using routine clinical aids with no scanner modification. All comparison methods in Table 1 use the same CBCT data acquired under this protocol and the same patient level five fold split. The gains over nnU-Net (0.38 vs 0.50 mm MAE, 56.2% vs 34.6% IoU@0.3mm) therefore reflect the algorithmic contribution rather than any acquisition advantage. (2) Baseline comparisons [Meta reviewer, R2.3, R4.2]. We appreciate the suggestion of stronger voxel baselines. Our method still outperforms recent architectures. SwinUNETR yields 0.48 ± 0.16 mm MAE and 37.1 ± 17.6% IoU@0.3mm, and 3D UXNet yields 0.46 ± 0.15 mm MAE and 38.8 ± 17.4% IoU@0.3mm. These will be added to Table 1.R4.2 also requested a surface manifold baseline. This is already present in Table 2 as “w/o Cross surface (K=1)”, which uses the same PointMLP encoder on the same tooth point cloud with coordinates and normals only, i.e. direct surface classification. Its IoU is 35.7% vs 56.2% for the full model, isolating CSR. “w/o Gaussian Band” reaches 53.2%, isolating BGR. (3) Reporting, ethics, and reproducibility [R1.2, R2.2, R2.4, R4.4, R4.5]. The five fold cross validation is patient level with no patient overlap, on 28 patients and 684 teeth spanning gingivitis to Stage IV periodontitis. The study was approved under SH9H-2026-T98-1, with anonymized data and informed consent. We will state the split protocol, IRB approval, and anonymization explicitly. We will expand the Introduction and Related Work to cover tooth, gingiva, and alveolar bone segmentation together with surface, mesh, and graph based learning, including the suggested ITMatch and NKUT references. We will also report training and evaluation details for the upstream nnU-Net tooth instance segmentation that precedes Marching Cubes and Laplacian smoothing, and provide a per fold MAE and IoU breakdown. (4) Clinical boundaries, robustness, and interpretability [R1.3, R4.1, R4.3, R4.5]. For multi rooted teeth the gingival margin lies at the CEJ on the crown above the furcation, so root multiplicity does not affect surface normals in the sampled −0.3 to 1.7 mm band. Molars achieve 0.39 mm MAE and 55.6% IoU, consistent with single rooted teeth. On error propagation, we excluded target teeth with full crowns or in tooth implants while retaining minor restorations. On this cohort the upstream nnU-Net reaches 96% Dice on teeth, sufficient to support stable downstream prediction. Stratified evaluation on severe metal artifact cases is left as future work. The protocol uses standard intraoral aids and is well tolerated, though severe gag reflex, restricted mouth opening, or extensive tooth loss can constrain placement, and we will state these boundaries in the Discussion. For interpretability we will add figures showing the 1D intensity profiles from the three anatomical regions and the predicted Gaussian confidence field rendered on the tooth surface. We believe these clarifications and additions address the central concerns and respectfully hope the reviewers will consider raising their recommendation
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 addresses an important clinical problem and introduces a thoughtful combination of a physical scanning protocol and a geometry-prior-based formulation. Reviewers generally find the method promising, but also note concerns about generalizability beyond the specialized protocol, limited baseline comparisons, and incomplete reporting. The authors are encouraged to address these points in rebuttal.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
This paper addresses an important clinical problem and proposes a thoughtful protocol-plus-method co-design for gingiva prediction in CBCT. The cross-surface representation and band-aware Gaussian regression are well motivated and align naturally with the anatomy of the task. Although one reviewer raised concerns about generalizability, I am more persuaded by the reviewers who saw the protocol as an enabling prerequisite rather than a confounder, and the rebuttal further clarified the split protocol and baseline fairness. Overall, I believe the methodology is convincing, and the reported gains are meaningful, so I am inclined to accept this paper.
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.
Although some limitations remain regarding the dependence on a specialized acquisition protocol and the need for broader clinical validation, the rebuttal adequately addresses the main concerns by clarifying the necessity and scope of the protocol, confirming patient-level evaluation and reproducibility details. Therefore, the AC recommends 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.
Two reviewers explicitly upgraded to Accept (R2=3→Accept, R4=4→Accept). Rebuttal effectively addressed main concerns: confirmed patient-level splits and ethics approval. Innovation is outstanding: physical scanning protocol + cross-surface representation + Gaussian regression.