Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Tooth segmentation in Cone-Beam Computed Tomography (CBCT) remains challenging, especially for fine structures like root apices, which is critical for assessing root resorption in orthodontics. We introduce GEPAR3D, a novel approach that unifies instance detection and multi-class segmentation into a single step tailored to improve root segmentation. Our method integrates a Statistical Shape Model of dentition as a geometric prior, capturing anatomical context and morphological consistency without enforcing restrictive adjacency constraints. We leverage a deep watershed method, modeling each tooth as a continuous 3D energy basin encoding voxel distances to boundaries. This instance-aware representation ensures accurate segmentation of narrow, complex root apices. Trained on publicly available CBCT scans from a single center, our method is evaluated on external test sets from two in-house and two public medical centers. GEPAR3D achieves the highest overall segmentation performance, averaging a Dice Similarity Coefficient (DSC) of 95.0% (+2.8% over the second-best method) and increasing recall to 95.2% (+9.5%) across all test sets. Qualitative analyses demonstrated substantial improvements in root segmentation quality, indicating significant potential for more accurate root resorption assessment and enhanced clinical decision-making in orthodontics. We provide the implementation and dataset at https://github.com/tomek1911/GEPAR3D.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/1833_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/tomek1911/GEPAR3D

Link to the Dataset(s)

GEPAR3D dataset: https://zenodo.org/records/15739014

BibTex

@InProceedings{SzcTom_GEPAR3D_MICCAI2025,
        author = { Szczepański, Tomasz AND Płotka, Szymon AND Grzeszczyk, Michal K. AND Adamowicz, Arleta AND Fudalej, Piotr AND Korzeniowski, Przemysław AND Trzciński, Tomasz AND Sitek, Arkadiusz},
        title = { { GEPAR3D: Geometry Prior-Assisted Learning for 3D Tooth Segmentation } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15961},
        month = {September},
        page = {218 -- 228}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The paper address the problem of tooth segmentation from cone-beamed CT. It combines a statistical shape model of a known “atlas” of teeth learned from public data and a 3D deep watershed. Detailed comparison is performed with SOTA methods, excellent results are demonstrated.

  • 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 is a classic application paper that combines known approaches to solve an outstanding challenge and significantly improve the state-of-the-art.
    • the method appears to work really well
    • the paper is very well written, all the necessary implementation details are provided
    • two in-house datasets are mentioned, I hope they’ll be made available
  • 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 novelty is on the application rather than methods side
  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission has provided an anonymized link to the source code, dataset, or any other dependencies.

  • 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/2025/en/REVIEWER-GUIDELINES.html
    • as someone outside the dentistry field, I would appreciate an earlier explanation of classes in the multi-class segmentation problem. From the fact that there are 32 classes, i assume that every tooth makes its own class? And the geometric model helps assure the right classes are assigned if some teeth are missing? Perhaps spell it out in a more obvious way.
  • 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 — could be accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    It’s a good application paper, demonstrating strong segmentation results from combining several approaches.

  • 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.

    Appears to be a potentially impactful contribution, defining the new SOTA.



Review #2

  • Please describe the contribution of the paper

    The proposed Geometric Wasserstein Dice Loss (GeoWDL) effectively encodes anatomical relationships through a penalty matrix derived from inter-tooth distances, enhancing segmentation consistency. The 3D energy basin regression and directional gradient estimation refine boundary localization, particularly for fine root structures, advancing instance-aware segmentation.

  • 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 introduces a novel framework (GEPAR3D) that integrates a Statistical Shape Model (SSM) with deep watershed instance segmentation, offering a promising solution to the challenging problem of root apex segmentation. The integration of geometric priors and energy basin modeling is technically sound and clinically impactful. Extensive experiments across four external test sets (including multi-center data) demonstrate superior performance over state-of-the-art methods, with significant improvements in DSC and recall. Ablation studies validate the contributions of key components.

  • 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 construction of the SSM relies on statistical data (e.g., inter-tooth distances, gender-specific models), but the paper does not clarify whether the SSM generalizes to populations with abnormal dentition (e.g., missing teeth, pediatric patients). You should provide demographic details (age, gender, dental conditions) of the cohort used for SSM training and discuss potential biases.
    2. The penalty weights in Q_(q_ij )are empirically set without theoretical justification, risking subjective bias. You should justify the penalty modifiers via cross-validation or anatomical literature (e.g., structural similarity between quadrants).
    3. The core idea of combining a Statistical Shape Model (SSM) with deep watershed instance segmentation lacks sufficient novelty. Prior works have integrated SSMs or anatomical constraints into dental segmentation pipelines, albeit with different technical implementations.
    4. The ablation experiments do not isolate the individual contributions of the SSM and watershed components. For example, it is unclear whether the SSM alone might over-constrain the model or if the watershed method could independently improve performance.
    5. The paper claims significant potential for root resorption assessment, yet fails to validate GEPAR3D’s performance on cases with severe root resorption or missing adjacent teeth. For instance: 1)Root Resorption: When molars undergo pathological root shortening or morphological distortion (e.g., blunted apices), does the SSM-based geometric prior, which assumes “normal dentition,” still provide valid guidance? 2)Missing Teeth: Adjacent tooth loss alters inter-tooth spatial relationships (e.g., drifting molars). The SSM encodes fixed quadrant-wise distances (Sec. 2), but how does GEPAR3D handle deviations from these priors? 3)You should test GEPAR3D on public datasets with annotated resorption cases or simulate resorption via synthetic deformations. Provide ablation studies where teeth are artificially removed from test scans to evaluate robustness to missing adjacent structures. Discuss limitations of the SSM in modeling pathological variability and propose mitigations (e.g., adaptive priors).
  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission has provided an anonymized link to the source code, dataset, or any other dependencies.

  • 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/2025/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — could be accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The author provides the source code and rich experimental content of the paper, but lacks analysis and solutions to the problems of tooth absorption and tooth loss.

  • 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.

    None



Review #3

  • Please describe the contribution of the paper

    This paper proposes a novel Geometry Prior-Assisted Learning approach for 3D tooth segmentation, named GEPAR3D. The method incorporates a Statistical Shape Model of dentition as a geometric prior and leverages a deep watershed technique to model each tooth as a continuous 3D energy basin.

  • 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. Figure 1 of the paper effectively illustrates the overall pipeline of the proposed method.
    2. Extensive ablation studies are conducted to validate the contribution of each component in the framework.
    3. The proposed GEPAR3D method explores the area of integrating segmentation priors and geometry prior-based learning strategies for 3D tooth segmentation.
    4. A particularly noteworthy aspect is the introduction of a penalty matrix, which captures the relational structure among different dental quadrants.
    5. GEPAR3D is compared with a wide range of existing approaches in the field and demonstrates consistently strong performance across various benchmarks.
  • 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. Adjusting the mathematical equations within the text is a good trick of saving the space, however it might be difficult for the reader to absorb them while reading the paper in flow.
    2. It will be nice to add a figure explaining different quadrants, arches, origin and the geometric centre, and how they are related.
    3. The reference to optimal transport theory is missing.
    4. In table-1 the paper mentions TF2 and Cui et al, but its not mentioned anywhere in the paper that what are these?
    5. In Figure-2 of the visual results the ground truth segmentations are not shown.
    6. The architecture details of 3D U-Nets segmentation are missing, also I tried to check the anonymous github reop but most of the codes are not accessible, like the 3D-UNet model and GEPAR3D.
    7. The inference-stage ROI (Region of Interest) network is mentioned but not described in sufficient detail. Missing information includes training strategy, data split used for this network, and its individual performance metrics, all of which are important for reproducibility and understanding its contribution to the final results.
    8. The abstract claims that the method “ensures accurate segmentation of narrow, complex root apices,” but no specific example or quantitative result supports this statement in the paper. Including such evidence would help substantiate the claim.
  • 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.

  • 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/2025/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — could be accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The ideas presented in the paper are novel , and several important claims are made in the abstract. While the first part of the paper clearly outlines the methodology, there is a lack of supporting evidence for some of the claims, particularly those related to the advantages of the proposed method over existing approaches in challenging cases. For instance, the abstract states that the method “ensures accurate segmentation of narrow, complex root apices,” yet no qualitative examples, quantitative evaluations, or case-specific analysis are provided to substantiate this claim. Including such evidence would significantly strengthen the impact and credibility 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.

    The rebuttal has addressed the majority of my concerns; however, I still notice missing details regarding the ROI Network training and related details.



Author Feedback

We thank reviewers for constructive feedback, recognizing our strong results and comprehensive experiments. Suggested clarifications will be incorporated in the final version if accepted.

R1–R3 (Generalization to abnormal dentition) The SSM in GeoWDL acts as a soft loss prior during training; it does not impose hard constraints or override ground-truth (GT) labels. The SSM weights the loss by anatomical plausibility (penalizing anatomically invalid errors such as cross quadrant swaps) yet still lets the model fit every labeled case, including pathological variations like missing or drifted teeth. Thus, GEPAR3D learns abnormal anatomies alongside normal ones. The SSM is based on an unbiased cohort [15]; we will report its demographics per Kim et al.

R2 (Resorption Assessment Prerequisite) Resorption analysis requires comparing sequential scans to a reliable baseline segmentation. As no public CBCT datasets include resorbed annotations, we first validate general apex segmentation accuracy. Under‐segmentation would mask any subsequent shortening. Future work will test resorbed cases to assess model performance on pathological roots; we will discuss it in the conclusions.

R2, R3 (Challenging cases evaluation) Existing methods under-segment apices (Fig. 2), a limitation we explicitly quantify. We report i.a. apex-sensitive metrics - binary NSD and RCB (see Evaluation section). We also overlay Hausdorff distance (HD) heatmaps on GT to localize errors (Fig. 2). GEPAR3D achieves higher DSC, NSD, and RCB across public and private CBCT datasets (4 centers, 2 ethnicities, 46 scans), including missing-teeth cases (Fig. 2, In-house B). For example, NSD1 improves +3.3% over SGANet, and RCB +9.5% over TSG-GCN (Tab. 1), confirming better root coverage.

R1, R2 (Limited novelty and non-isolated ablation) GEPAR3D is the first fully automatic 3D CBCT tooth segmentation method embedding an SSM into the loss. Prior dental SSM approaches (Evain et al., 2017; Keustermans et al., 2012; we will refer to them in related works) are non-differentiable and require manual initialization. Unlike recent SGANet and TSG-GCN that impose priors as rigid architectural constraints, GeoWDL uses soft regularization to preserve flexibility for rare morphologies. We introduce a novel 3D watershed transform adaptation, replacing classification with distance regression and adding 3D angular loss to improve apex precision. Ablation (Tab. 2) shows the SSM alone (#3) boosts DSC by +1.28, watershed (#4) by +1.31, and both (#7) yield (+1.74).

R2 (Clarify limitations) Our training was limited to adult dentition, reducing generalization to pediatric cases. SSM adaptation would require extending the penalty matrix to deciduous teeth. Static SSM priors (GEPAR3D, SGANet) outperformed dynamic ones (TSG-GCN, Tab. 1), though adaptive priors remain promising but data-demanding; we defer this to future work.

R2 (Penalty weights) GeoWDL’s penalty weights are anatomically motivated [17,29]: maxillary and mandibular teeth differ morphologically. The weights retain the SSM’s distance-based core but emphasize anatomical similarity beyond geometry. Misclassifying a tooth as its distant contralateral in the same arch incurs a lower penalty than confusing it with a nearby tooth from the opposite arch - an anatomy-driven design, rather than hyperparameter tuning.

R1, R3 (Reproducibility) Upon acceptance, we will release full code (architecture, ROI detector) and IRB-approved in-house test sets.

(Camera-ready clarifications) Fig. 2 shows surface HD heatmaps overlaid on GT labels (green = low error, purple = high), highlighting apex deviations; we will clarify the caption. We will add quadrant layouts, a 32-teeth legend, its earlier introduction, and improve equation formatting. We will specify “TF2” refers to ToothFairy2 [2,5] and “Cui et al.” to [6]. Optimal transport [9] defines the Wasserstein loss as the minimal cost of moving probability mass between prediction and target.



Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.

    N/A

  • 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    Reviewers agree that the work has merit and is of interest for MICCAI community. Weaknesses identified during reviewing were addressed in the rebuttal.



back to top