Abstract

We introduce CrownFusion, a latent diffusion model to generate 3D dental crown designs from geometry images, a 2D-grid encoding of surface coordinates. Departing from traditional approaches that use point clouds, our “CrownImage” representation not only unlocks the full power of image-based generative architectures, but also preserves fine geometric detail by operating at far higher spatial resolutions than point clouds to alleviate smoothing artifacts in the 3D generation. Conditioned on jointly learned embeddings of the surrounding dentition, our probabilistic diffusion model produces anatomically plausible crowns with morphological variations tailored to individual patient cases. To accommodate multiple valid designs, we propose an occlusal fit metric based on proximity to the antagonist teeth. We evaluate our method through quantitative experiments and qualitative studies involving expert testimonials. We also release the FDI 16 Crown Dataset, comprising 11,513 annotated intra-oral scans, to support future research in dental CAD.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2827_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/JohanYe/CrownFusion/tree/main

Link to the Dataset(s)

https://github.com/JohanYe/CrownFusion/tree/main/data

BibTex

@InProceedings{YeJoh_CrownFusion_MICCAI2026,
        author = { Ye, Johan Ziruo AND Yan, Xingguang AND Hauberg, Søren AND Chang, Angel X. AND Zhang, Hao AND Søndergaard, Peter Lempel},
        title = { { CrownFusion: 3D Dental Crown Generation Using Geometry Images and Latent Diffusion } },
        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

    CrownFusion proposes a latent diffusion model for 3D dental crown generation that operates on a novel 2D representation of crown geometry: the CrownImage. A dental crown, as a genus-zero open surface, admits a global parameterisation onto a 2D disk via boundary-fixed harmonic mapping. A fisheye-style radial transform is applied to allocate spatial resolution preferentially to the occlusal surface, which carries the most clinically relevant anatomical detail. The resulting geometry image a 512×512 grid encoding (x, y, z, α), is processed by a fine-tuned FLUX VAE, and a Scalable Interpolant Transformer (SiT) generates CrownImages conditioned on the surrounding dentition (encoded as a 3D point cloud). The paper also releases the FDI 16 Crown Dataset (11,513 annotated IOS cases) and proposes an antagonist-proximity occlusal fit metric as a clinically motivated complement to Chamfer distance.

  • 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.

    VAE reconstruction quality is exceptional: CD=0.02 vs. 1.21 for VF-Net (Table 1) , a 60× improvement in reconstruction accuracy , confirms that the geometry-image representation preserves fine geometric detail far beyond what point-cloud autoencoders can achieve. The heatmap visualisations (Fig. 4) show errors in VF-Net concentrated at cusps and fissures (the clinically critical regions), while CrownFusion reconstructs these features accurately.

  • 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.Missing-context handling for absent neighbouring teeth is unspecified 2.No quantitative diversity metric for the generative model

  • 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.

  • 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.

    (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?

    The paper is of broad interest both to the medical image analysis community (novel 3D-to-2D representation for generative modelling) and to the dental AI community.

  • Reviewer confidence

    Somewhat confident (2)

  • [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 CrownFusion, a latent diffusion-based framework for generating 3D dental crowns using geometry image representations, conditioned on surrounding dentition.

  • 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 paper introduces geometry images (CrownImages) as a 3D-to-2D representation, which leverages mature image-based generative models while preserving fine geometric details. 2.The release of a relatively large-scale, annotated intraoral scan dataset (FDI 16 Crown Dataset) is a valuable contribution that can support future research in dental CAD and generative modeling.

  • 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 model is trained and evaluated only on FDI 16 (upper first molar). This raises concerns about generalizability to other tooth types with different morphology, especially anterior teeth or premolars. 2.This method only focus on single missing tooth, which is limited in real clinical applications. 3.The prediction of tooth crown relies on the input of the tooth preparation. This is unclear as we know that the tooth crowns are mainly determined by its adjacent teeth.

  • 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • 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.

    (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?

    Please refer to the strengthens and weakness of this manuscript.

  • 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 #3

  • Please describe the contribution of the paper

    The main contribution of this paper is the introduction of CrownFusion, a methodological framework for 3D dental crown generation. This approach relies on the use of geometry images (CrownImages), which represent tooth surfaces as high-resolution 2D grids encoding full 3D coordinates. This representation enables the application of latent diffusion models for generating complex shapes with a VAE uses for latent space encoding. By combining this formalism with a conditional diffusion transformer that integrates embeddings of adjacent and antagonist teeth, the method can produce multiple anatomically plausible crown designs tailored to individual patients. It thus overcomes the limitations of previous approaches based on point clouds or depth map representations, particularly in terms of geometric quality and anatomical consistency. Furthermore, the paper introduces a clinically relevant occlusal fit metric based on the proximity between the generated crown and the antagonist teeth, providing an objective criterion to evaluate the functional quality of the proposed designs. Finally, the authors release a large annotated dataset of intraoral scans, helping to support and stimulate future research in digital dentistry and anatomical shape generation.

  • 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 paper introduces a formulation of 3D dental crown generation based on the use of geometry images. This approach transforms a complex 3D surface modeling problem into a 2D image generation task while preserving the full geometric information. This is particularly interesting because it allows the direct use of powerful diffusion architectures originally designed for images, without suffering from the smoothing artifacts commonly associated with point cloud-based methods. Moreover, the paper proposes a conditioning strategy, in which representations of neighboring and antagonist teeth are integrated into the model through cross-attention mechanisms. This design enables the generation of crowns that are anatomically realistic and functionally consistent with each patient’s specific anatomy. Furthermore, the method is fully generative and models a distribution of plausible crowns rather than producing a single deterministic solution, aligning with the variability observed in clinical practice, where multiple designs can be acceptable for a given patient. It also demonstrates strong clinical relevance. Expert evaluations as well as a reduction in the need for manual corrections and intersection artifacts compared to existing methods. Finally, the experimental evaluation by introducing a domain-specific occlusal fit metric. This metric better reflects the functional performance of the crowns than standard geometric distances. The release of a large, realistic, and annotated dataset of intraoral scans further strengthens the impact of the paper and the reproducibility of its results.

  • 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 use of geometry images, while effective, is not entirely novel. Prior works already explored geometry images, including in diffusion-based generative settings. As a result, the contribution lies more in adapting this representation to the dental domain rather than introducing a fundamentally new formulation. Moreover, the overall generative framework combining a VAE, a latent diffusion model, and transformer-based conditioning closely follows established pipelines such as Latent Diffusion Models and SiT architectures. This reduces the degree of architectural novelty, even though the integration is well executed. In addition, the reliance on a variational autoencoder for latent space encoding introduces well-known limitations, including reconstruction blur, loss of high-frequency details, and a delicate trade-off between compression and fidelity. These limitations are particularly critical in this setting, as even small latent errors can lead to noticeable geometric distortions in the reconstructed meshes. Alternative approaches such as diffusion models operating directly in pixel could potentially better preserve fine structures. Furthermore, the geometry image parameterization assumes a simple, disk-like topology. This restricts the method’s ability to generalize beyond individual teeth and limits its applicability to more complex anatomical structures. Finally, the generation time is relatively high, around five minutes per case, which may hinder practical deployment in clinical settings. This is significantly slower than deterministic approaches could be a limiting factor in time-sensitive environments. But it remains valuable if it is use for planning where timing is not so sensitive.

  • 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.

  • 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.

    (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?

    please refer to the major strenghts listed above

  • 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 all reviewers for their feedback. Regarding Reviewer 2’s comment #3 under weaknesses: we would like to clarify that the crown generation is primarily conditioned on the surrounding/adjacent teeth, which serve as the main geometric context for determining crown morphology. The only information derived from the tooth preparation is the margin line, which defines the crown-gingiva border and ensures a proper fit at the preparation boundary. We agree that the adjacent teeth are the dominant signal for crown shape, and our model reflects this. We will make the conditioning inputs more explicit in the camera-ready version.




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.

    All three reviewers consistently assigned a score of 4.I agree with the consensus and recommend Accept. The CrownImage representation is a well-motivated adaptation that achieves strong geometric fidelity, the released dataset is a tangible community contribution, and the clinical evaluation is grounded and practical. Reviewers noted limited generalizability beyond FDI 16 and incremental architectural novelty, but these do not constitute fundamental flaws within the scope of a conference paper.



back to top