Transformer-based Tooth Alignment Prediction with Occlusion and Collision Constraints

Zhenxing Dong, Jiazhou ChenICCV 2025

TL;DR

This paper introduces a novel transformer-based neural network for automating the complex and time-consuming process of digital orthodontic tooth alignment. By novelly processing 3D tooth point clouds into ordered textures and introducing new occlusion-aware loss functions, the method achieves state-of-the-art accuracy while also addressing critical clinical constraints like collision avoidance.

What Problem Does This Paper Address?

Manually planning orthodontic treatments, which involves repositioning teeth for optimal alignment, is a highly skilled and labor-intensive task. This process relies heavily on orthodontists' clinical experience, making it time-consuming and costly. While AI has shown promise in automating this, existing methods often struggle with accurately representing local tooth features, handling complex occlusal relationships between upper and lower jaws, and ensuring collision-free alignments. This paper aims to bridge this gap by developing a more accurate, efficient, and clinically relevant automated tooth alignment system.

Methods

The core of the proposed method is a Swin Transformer-based network designed for processing 3D tooth point clouds. Key innovations include:

  • Data Reorganization: 3D point clouds of teeth are converted into ordered, multi-channel textures. This is achieved by first fitting a simulated dental arch line and then serializing points based on their distance to this arch, which helps maintain relative positions and improves network performance.
  • Dual-Module Architecture: The network features two branches:
  • A global module (SWTBS) that processes tooth center points to capture overall arch information.
  • A local module (SWTP) that uses a multi-stage hierarchical fusion mechanism within the Swin Transformer pipeline to extract detailed local features from tooth point clouds.
  • Novel Occlusal Loss Functions: Two new loss functions are introduced to quantitatively evaluate the occlusal relationship between the upper and lower jaws:
  • Occlusal Projecting Overlap Loss: Measures how well the predicted interocclusal region matches the ground truth.
  • Occlusal Distance Uniformity Loss: Assesses the consistency of distances between corresponding points in the occlusal regions of posterior teeth, and uses vertex coordinate and angular differences for anterior teeth.
  • Collision Detection Constraint: A BVH collision detection algorithm is integrated during data augmentation to ensure that generated or predicted alignments are collision-free.
  • Jaw Regularization Constraint: This constraint pulls teeth towards the central incisor along the dental arch line to correct excessive gaps, ensuring more reasonable augmented data.

Datasets

The authors introduce a newly collected digital orthodontic dataset comprising 837 sets of dental data. This dataset includes multiple treatment stages for each patient, with ground truth data derived from manually optimized and aligned tooth models. This dataset is noted to be more reliable than post-orthodontic scan data as it avoids topological differences. The dataset and source code will be released to the community.

Key Insights and Ideas

  • Serialization is Key: Organizing 3D point clouds based on a simulated dental arch line significantly improves the transformer's ability to understand inter-tooth relationships and local features, outperforming random or simple sorting methods.
  • Occlusion Matters: Explicitly modeling occlusal relationships with specialized loss functions is crucial for achieving clinically realistic and accurate tooth alignments, going beyond simple geometric reconstruction.
  • Constraints Drive Realism: Integrating clinical constraints like collision detection and jaw regularization during data augmentation is vital for generating realistic and safe orthodontic outcomes, preventing clinically illogical scenarios.
  • Swin Transformer's Adaptability: The Swin Transformer's hierarchical structure and shifted window mechanism prove effective for processing 3D point cloud data, balancing efficiency and feature extraction capabilities.

Key Results

  • The proposed method achieves state-of-the-art performance across various metrics, including ADD/AUC, average rotation error, and average translation error, outperforming methods like TANet, PSTN, TAligNet, and TADPM.
  • The new occlusal loss functions significantly contribute to improved prediction accuracy, particularly in handling complex occlusal relationships.
  • The constrained data augmentation method leads to better training quality and accuracy compared to regular augmentation.
  • The method demonstrates robustness in handling complex cases, including large gaps, malocclusion, fragmentary crowns, missing teeth, and jaw malposition.
  • The user study showed high satisfaction scores from orthodontists, indicating the clinical validity of the method's outputs.

Limitations and Future Work

  • Missing Teeth Handling: While the paper states missing teeth are handled, the impact on serialization and feature extraction is noted as negligible for the sliding window, suggesting potential areas for further improvement in explicit handling.
  • Computational Intensity: Although the method is efficient, the paper mentions that the proposed occlusal losses might lead to longer training times.
  • Future Directions: The authors plan to incorporate other stomatologic constraints and predict the full path of orthodontic treatment, rather than just the final target position.

Summary updated

BibTeX
@inproceedings{dong2025transformer,
  author = {Zhenxing Dong and Jiazhou Chen},
  title = {{Transformer-based Tooth Alignment Prediction with Occlusion and Collision Constraints}},
  booktitle = {IEEE/CVF International Conference on Computer Vision},
  year = {2025},
}
Figure