3D Teeth Segmentation using PointNet

Table of Contents

• Introduction
• Model Architecture
• Dataset
• Results
• License
• Acknowledgements
• Contact

Introduction

3D Teeth Segmentation using PointNet is an advanced deep learning project designed to automate the segmentation of teeth from 3D dental scans. Leveraging the power of PointNet—a neural network architecture specifically crafted for handling point cloud data—this project addresses the challenges associated with accurate and efficient dental segmentation. Precise segmentation is crucial for various dental applications, including orthodontic planning, prosthodontics, and oral surgery, thereby enhancing both diagnostic capabilities and treatment outcomes.

Model Architecture

the model is based on the PointNet architecture, which is designed to process point cloud data directly. The key components of our architecture include:

1. Input Transformation: A mini-network that aligns the input points to a canonical space.
2. Feature Transformation: Another mini-network that aligns features to a canonical space.
3. Segmentation Network: A series of multi-layer perceptrons (MLPs) that extract global and local features.
4. Output Layer: A final layer that predicts the segmentation label for each point.

The model's ability to handle unordered point sets makes it particularly suitable for 3D dental scans, where the number and order of points can vary between scans.

Dataset

the model was trained and evaluated on a proprietary dataset of 3D dental scans, consisting of:

• 300 high-resolution 3D scans of full dental arches

Results

Visualization of Results

True vs. Predicted Segmentation

Figure 1: Comparison of True vs. Predicted Teeth Segmentation

This visualization demonstrates the high accuracy of our model in segmenting individual teeth from a full 3D dental scan. The left image shows the ground truth segmentation, while the right image displays our model's predictions. Note the precise delineation of tooth boundaries and the correct identification of different tooth types.

License

This project is licensed under the MIT License.

Acknowledgements

• PointNet Authors: Qi et al., 2017
• Open3D: For their excellent library on 3D data processing.
• 3D Smart Factory: For supervision and guidance.

Contact

Mohamed Alaoui Mhamdi

• Email: al.mh.mohamed@gmail.com
• LinkedIn: MohamedalaouiMhamdi
• GitHub: mohamedAlaouimhamdi