Abstract

The design of restorative dental crowns from intraoral scans remains labor-intensive for dental technicians. To address this challenge, we propose a novel voxel-based framework for automated dental crown design (VBCD). The VBCD framework generates an initial coarse dental crown from voxelized intraoral scans, followed by a fine-grained refinement incorporating distance-aware supervision to improve the accuracy and quality. During the training stage, we employ the Curvature and Margin line Penalty Loss (CMPL) to enhance the alignment of the generated crown with the margin line. Additionally, a positional prompt based on a tooth numbering system is introduced to further improve the accuracy of the generated dental crowns. Evaluation on a large dataset of intraoral scans demonstrates that our approach outperforms existing methods, providing a robust solution for personalized dental crown design.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/2280_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/lullcant/VBCD

Link to the Dataset(s)

N/A

BibTex

@InProceedings{WeiLin_VBCD_MICCAI2025,
        author = { Wei, Linda and Liu, Chang and Zhang, Wenran and Zhang, Zengji and Zhang, Shaoting and Li, Hongsheng},
        title = { { VBCD: A Voxel-Based Framework for Personalized Dental Crown Design } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15967},
        month = {September},
        page = {632 -- 641}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This work proposes a personalized voxel-based crown design method that is trained using a large dataset of around 6500 IOS. The machine learning model learns to create a coarse representation of the crown in a voxelized space, which is then converted to a point cloud mesh, and the mesh is refined further to give as close a representation to a “ground truth” crown as possible. The method also uses the FDI tooth number as an input to the algorithm. Shape information is taken into account during the refinement process.

  • 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 is well written and clear. The proposed deep learning model takes into account factors like local shape using curvature and normals information when coming up with the final refined tooth design. Method evaluation is well presented with comparisons to different methods, and using 3 scores showing improvement in all 3. An ablation study is also included showing that tooth prompt, mesh refinement and curvature information inclusion improve the crown design.

  • 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.

    Nice work! I don’t think these are weaknesses, but a few queries from my end:

    • Could you please describe the dataset a bit more? Are these all adult IOS? Where were these done, and which machines? How are the teeth distributed (i.e.are there the same proportion of canines as molars, etc)?
    • Why is the dataset split into 8:1? Did you use the other 10% for validation during training process?
    • The number of iterations in training is 720000. This seems to be extremely high. Could you please comment on why?
    • Can you comment on. how long does an inference take?
    • Does this method assume perfect anatomy around the tooth?
    • In figure 2, Could you please describe where the ground truth volume is fed into the pipeline, and where ConvCoarse needs to go.
    • What are the next steps here? Apart from making it more efficient during inference phase, how would you make it available in practice?
  • 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

    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?

    Overall paper is well written, and structured well. The experiments seem to show the efficacy of the method well. This is nice work. It would be great to see some more insight into the dataset.

  • 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



Review #2

  • Please describe the contribution of the paper

    The paper investigates a new method for automatic crown design based on an intraoral scan and reports on a validation of the approach on a large dataset. Dataset used for training and testing includes crowns for different tooth types, such as incisor, canines, and molars.

  • 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 is well structures and easy to follow. Prior work in this area is discussed and referenced, and the proposed method seems to be a logical extension of these prior work. The authors validated their proposed solution in a large and diverse dataset.

  • 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.

    I don’t see any major weaknesses with the manuscript. However, I would have liked to see a little bit more of a link between the motivation of this work (limitation of current methods to fixed and limited number of points in the generated crown design), and the results and discussion of the proposed method. What was the variation in the number of points in the crown design for your method? It would have been great to see how this proposed method compared to the method suggested in [23]. However, as I understand, the code for [23] is not more available as open source, which makes this more difficult.

  • 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

    I noticed that in your evaluation experiment, the results for DMC [7] seem to be quite a bit larger compared to the results reported in [7] (CD-L2 0.375 vs 0.011). Do you have any thoughts, suggestion, hypothesis regarding this which could be briefly discussed in the manuscript? Minor remarks to improve the readability of the manuscript:

    • Introduction: Define SAP once
    • Results: physical cooridinate -> physical coordinate
  • 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?

    This manuscript is well written and the proposed method is sufficiently validated. The research is a continuation of previous research in this area, and I believe would be of interest for researchers in this field.

  • 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 addressed the questions and concerns of the reviewers with confidence and clearly. Added details, such as data distribution, average nr of mesh points, and average inference time, improved the clarity and reproducibility of the work. I would recommend to accept the work for the inclusion into the MICCAI program.



Review #3

  • Please describe the contribution of the paper

    This paper proposes a fully automated pipeline for generating personalized dental crowns based on intraoral scans. The method consists of a coarse-to-fine strategy of first predicting a voxel-based crown, and then refining the output in using point clouds to finally generate a refined mesh. The method includes tooth position encoding, Unet-based voxel processing, MLP-based pointcloud processing, and DPSR + marching cubes for meshing. The results are obtained on large datasets, against both existing baselines and ablation experiments.

  • 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 is well-written. The method includes interesting contributions, e.g. coarse-to-fine strategy using various geometry representations, tooth position encoding, curvature and CMPL losses, etc. The performance is strong against baselines, on a large dataset. Ablation studies are well designed and clearly show the benefits of each component.

  • 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 description does not mention a validation set, which is interesting. How is overfitting detected, if at all? It is unclear how 2x2x2 cm^3 cropping was performed, and whether different crop centers could affect the results. It would be interesting to include a time scale for designing crowns using conventional methods vs. this method.

  • 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

    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.

    (6) Strong Accept — must be accepted due to excellence

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

    This paper is well-organized, well-written, and shows strong performance on a large dataset.

  • 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.

    After reading through all of the reviews and rebuttal, I stand by my original review. Actually, I’m not sure why this paper was not early accepted in the first place. It seems like all reviewers agree on the paper’s merits. The few minor questions I had were answered well by the rebuttal. Incorporating these responses in the final camera-ready version would be helpful.




Author Feedback

We appreciate the reviewers’ consistent recognition of our work and their constructive feedback. We summarize the comments with great care and address each of the points.

  1. Dataset (R1, R3) All data are from adults and provided by a partnered commercial company. Due to privacy policies, further details such as acquisition equipment or company name cannot be revealed. The data distribution is: 16% incisors, 3% canines, 25% premolars, and 56% molars. This imbalanced data distribution is consistent with the clinical observation that premolar/molar damage is more common in practice. Original train/val/test split is 7:1:1. In final evaluation, we merged the training and validation sets and retrained the model, resulting in an 8:1 ratio in section 3.1. We use this data split to ensure sufficient incisors and canines in the training set. All cases in our dataset assume perfect anatomy around the tooth. Cases with adjacent or antagonist missing teeth will be addressed in future work.

  2. Method
    • Region of interest (R1) Regarding Section 3.1, we crop a 2 cm-sided cube centered on the crown of the target tooth, similar to [23]. Considering the average tooth size (~10mm), the ROI with 2 cm side is sufficient to encompass target tooth, the antagonist and adjacent teeth, which are vital to crown design. Displacement of the ROI center resulting in partial exclusion of these contents may compromise the generation results.
    • Comments on Fig 2 (R3) As described in Section 2.2, we use the ground truth volume (V_GT) for voxel-level coarse supervision. In Fig.2, V_GT corresponds to the red upward arrow after the UNet backbone (not shown for concision). ConvCoarse denotes the output layer of UNet, which aims to provide logits for L_BCE computation.
    • Motivation (R2) The number of vertices in the crown mesh varies depending on the type of the target tooth. For example, the shape of the molar crown is more complex than incisor. Thus, the number of vertices required for a molar crown mesh (>10k) exceeds that of an incisor (<8k). Previous methods based on point cloud frameworks could only generate points with a fixed number, lacking the capability to dynamically adjust the output vertex count according to tooth type. VBCD could generate a variable number of points (7k-9k points for incisor, >10k points for molar, etc.), allowing for higher-quality crown meshes, as shown in Fig.3.
  3. Experiment Experiment Setting (R3) Regarding Section 3.1, the dataset includes 6499 cases, with 5776 used for training. We trained for 720,000 iterations (~125 epochs), a common setup for training a relative large backbone. Iterations were used instead of epochs to better support the cosine learning rate scheduler, enabling smoother adjustment.
  • Discrepancy of result report in table1 and [7] (R2)
  • There are two main reasons:
    1. Our dataset is more diverse and extensive, which may expose robustness limitations in DMC.
    1. Regarding section 3.2, our metrics are computed in real-world coordinates (coordinates measured in mm), while the source code of DMC evaluates CD-L2 in normalized coordinates, which understate the error. Visual results in [7] also indicate notable differences from the ground truth, suggesting their reported CD-L2 may not fully reflect real-world accuracy.
  • Inference time (R1,R3)
  • We regret omitting it in the manuscript for brevity. Our log shows it takes 4.3 min to run through the test set (723 cases) under the settings in Section 3.1 (about 357 ms/case), which is much faster than the manual method (5-10 min/case)[14].
  1. Future work (R1,R3) Regarding Section 4. The short-term goal is to improve computational efficiency and memory usage. The mid-term goal is generate the inner surface with the extracted cervical margin to compose the whole crown for practical application. The long-term goal is to address crown generation for cases with multiple adjacent missing teeth.




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’

    I have read the manuscript, review comments, rebuttal letter. All reviewers recommend acceptance (after rebuttal). This meta reviewer believes that the authors did a good job in addressing concerns.



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’

    The author has addressed the concerns of the authors successfully



back to top