Maxilla–Mandible Occlusion Analysis

A simple, Python-based occlusion estimation for maxilla–mandible models captured by intraoral scanners.

Requirements (for all script)

• Python 3.12
• PyTorch >= 2.7
• CUDA >= 12.8
• see requirements.txt for additional dependencies
• Both the maxilla and mandible models should be aligned in the bite position.
• Models should be in .stl file format.

How to use this repository:

This repository includes three script iterations:

1. OCCcompare.ipynb: computes tooth occlusion between the mandible and maxilla using libigl's signed_distance function.
2. OCCcompare+ML.ipynb: computes tooth occlusion between the mandible and maxilla + tooth segmentation with DilatedToothSegNet and highlights teeth that have no contact.
3. OCCcompare+ML+fast_simplification.ipynb: computes tooth occlusion between the mandible and maxilla + tooth segmentation and highlighting teeth that have no contact + mesh simplification usinpyvista (replacing the orignialy used Fast Quadric Mesh Reduction from Fast Quadric Mesh Reduction from link).

To run these scripts, simply open Jupyter Notebook and execute them. Parameters that can be adjusted are defined inside the scripts.

The results are show as 3D image and tables:

----Number of file ----
All contact area: XXX mm²
Good contact (green): XXX mm²
Bad contact (red): XXX mm²
Intermediate contact (orange): XX mm²

Docker

To build Docker

docker build -t OCCcompare_docker .

Conatiner can be started (Windows PowerShell)

docker run -p 8888:8888 -e JUPYTER_TOKEN=12345 -d --gpus all -v "${PWD}:/app" --name dilatedtoothsegnet dilatedtoothsegnet

A Jupyter notebook can be started by opening the following in a browser: http://localhost:8888/?token=12345

Separate teeth highlight using DilatedToothSegNet

• For segmentation DilatedToothSegNet was used.
  • Model was train using Teeth3DS dataset with mixed maxilla and mandible 1200 models (referred to by the author as split 'S2'). Training parameters: n_bit_precision=16, train_batch_size=2 and epochs=100.
  • Training was performed NVIDIA RTX 4080 SUPER GPU.
  • Model weigts are in ML_param.ckpt file

Caution

Be aware that your 3D teeth models better be aligned with the Teeth3DS dataset models. Both ML-based scripts include a correction step (get_pca_rotation), but it is not stable. Future improvements should include a better implementation to address this issue (eg. AlignNet part).