TeethGenerator: A two-stage framework for paired pre- and post-orthodontic 3D dental data generation

Changsong Lei, Yaqian Liang, Shaofeng Wang, Jiajia Dai, Yong-Jin Liu

TL;DR

This paper introduces TeethGenerator, a novel two-stage framework for synthesizing paired 3D dental models representing teeth before and after orthodontic treatment. This addresses the critical bottleneck of data scarcity in digital orthodontics, enabling the training of more robust tooth alignment networks and potentially aiding in the education of dental professionals.

What Problem Does This Paper Address?

Digital orthodontics, a field leveraging computer vision for dental treatment planning, faces a significant challenge: the creation of high-quality, paired 3D models of teeth before and after orthodontic intervention. Acquiring such datasets is labor-intensive, requiring multi-year patient treatments and meticulous scanning. This scarcity hinders the development and validation of automated tooth alignment systems, which rely heavily on large, diverse datasets. Existing 3D generation methods often struggle with the complexity of generating anatomically structured dental arches (24-32 segmented teeth) and maintaining consistency between pre- and post-treatment models, which is crucial for training downstream networks.

Methods

TeethGenerator employs a two-stage approach:

  • Stage I: Teeth Shape Generation Module: This stage focuses on generating post-orthodontic 3D teeth models. It utilizes a VQ-VAE (Vector Quantized-Variational Autoencoder) to compress 3D tooth point clouds into a discrete latent representation. A diffusion model is then trained on this latent space to learn the distribution of tooth shapes and alignments. By sampling from this learned distribution and decoding, the module generates diverse post-orthodontic teeth models. The paper also highlights the importance of organizing teeth into a structured grid (2x2x8 based on FDI numbering) to capture positional and symmetry relationships.
  • Stage II: Teeth Style Generation Module: This stage generates the corresponding pre-orthodontic teeth models. It uses a Transformer architecture that takes the post-orthodontic model's shape information and a "style model" (representing the desired pre-orthodontic state) as input. Feature extractors (E_style and E_shape) are used to capture stylistic and shape information. The Transformer predicts transformation parameters that are applied to the post-orthodontic models to achieve the desired pre-orthodontic appearance, crucially preserving the underlying tooth morphology and ensuring consistency.

Datasets

The primary dataset used is a collection of 1060 pairs of pre/post-orthodontic 3D teeth models from real patients, as described in [36]. This dataset was filtered to 720 pairs for training, 80 for validation, and 120 for testing, offering diverse malocclusion categories. The paper does not introduce a new dataset but focuses on generating synthetic data to augment this existing one.

Key Insights and Ideas

  • Paired Data Generation is Crucial: The paper strongly emphasizes that generating paired pre- and post-orthodontic models is essential for training tooth alignment networks, and TeethGenerator is presented as the first framework to achieve this.
  • Diffusion Models for Shape, Transformers for Style: The separation of concerns into generating the "target" shape (post-orthodontic) using diffusion models and then transforming it to the "initial" state (pre-orthodontic) based on a style using Transformers is an effective architectural choice.
  • Structured Representation Matters: Organizing teeth into a grid (2x2x8) is shown to be beneficial for capturing inter-tooth relationships, leading to better learning of topological and symmetry properties.
  • Synthetic Data Boosts Downstream Performance: The experiments demonstrate that incorporating the generated synthetic data significantly improves the performance of tooth alignment networks, even when the synthetic data size is much larger than the real training set.

Key Results

  • State-of-the-Art Generation: TeethGenerator outperforms existing general 3D shape generation methods (PointFlow, DPM, LION, PVD, DiT-3D) on metrics like Chamfer Distance (CD) and Earth Mover's Distance (EMD) for post-orthodontic teeth generation, achieving a CD of 69.50% and EMD of 71.88% on a 720-sample test set. It also achieves high uniqueness (UCD of 96.25%).
  • Realistic Visuals: The generated models exhibit diverse tooth morphologies, clear boundaries between individual teeth, and realistic alignments, closely resembling real data.
  • Improved Tooth Alignment: Training a tooth alignment network (TANet) with a combination of real and synthetic data leads to significant performance improvements (measured by ADD, PA-ADD, and CSA metrics) compared to using real data alone. Performance continues to improve as more synthetic data is added, up to 10 times the real dataset size.

Limitations and Future Work

  • Data Size Dependency: While performance improves with more synthetic data, the paper notes that it eventually converges, suggesting that the quality and diversity of the synthetic data might become a limiting factor beyond a certain point.
  • Mesh Conversion: The output is point clouds, and mesh conversion for visualization is done via registration, which might introduce minor inaccuracies.
  • Future Work: The authors suggest that the generated data has the potential to assist in the training of orthodontists and technicians, implying future work in educational applications. Further exploration of the synthetic data's impact on other downstream tasks in digital orthodontics is also a natural extension.

Summary updated

Abstract

Digital orthodontics represents a prominent and critical application of computer vision technology in the medical field. So far, the labor-intensive process of collecting clinical data, particularly in acquiring paired 3D orthodontic teeth models, constitutes a crucial bottleneck for developing tooth arrangement neural networks. Although numerous general 3D shape generation methods have been proposed, most of them focus on single-object generation and are insufficient for generating anatomically structured teeth models, each comprising 24-32 segmented teeth. In this paper, we propose TeethGenerator, a novel two-stage framework designed to synthesize paired 3D teeth models pre- and post-orthodontic, aiming to facilitate the training of downstream tooth arrangement networks. Specifically, our approach consists of two key modules: (1) a teeth shape generation module that leverages a diffusion model to learn the distribution of morphological characteristics of teeth, enabling the generation of diverse post-orthodontic teeth models; and (2) a teeth style generation module that synthesizes corresponding pre-orthodontic teeth models by incorporating desired styles as conditional inputs. Extensive qualitative and quantitative experiments demonstrate that our synthetic dataset aligns closely with the distribution of real orthodontic data, and promotes tooth alignment performance significantly when combined with real data for training. The code and dataset are available at https://github.com/lcshhh/teeth_generator.
BibTeX
@inproceedings{lei2025teethgenerator,
  author = {Changsong Lei and Yaqian Liang and Shaofeng Wang and Jiajia Dai and Yong-Jin Liu},
  title = {{TeethGenerator: A two-stage framework for paired pre- and post-orthodontic 3D dental data generation}},
  booktitle = {IEEE/CVF International Conference on Computer Vision},
  year = {2025},
}
Figure