View on GitHub

Robust Teeth Detection in 3D Dental Scans by Automated Multi-View Landmarking

DL-based framework for the detection of landmarks in intraoral scans of orthodontics cases

Download this project as a .zip file Download this project as a tar.gz file

Welcome to GitHub Page for the project I’ve created as my Bachelor’s Thesis and later presented at BIOIMAGING ‘22 conference.

The primary objective of this page is to present summary of the paper as well as to hold all the necessary links at one place (code, testing dataset, …).

Method Summary

Our method aims to bring automation in the process of orthodontics treatment. Some facts that stem from this:

Outline

To meet the aforementioned needs, the learning task is to regress heatmaps of Gaussians in 2D. Afterward, using multiple viewpoints, the result postion is calculated from valid viewpoints only – incorrect ones are eliminated using RANSAC and least-squares fit.

Overall Results

On a testing dataset of complicated orthodontic cases, we report the landmarking accuracy of 0.75 +- 0.96 mm. As for the missing teeth, our method detects correctly the presence of teeth in 97.68% cases. These results are achieved using Attention U-Net, 100 viewpoints and RANSAC post-processing.

Dataset Availability

Unfortunately, we are unable to share the dataset used for training, validation, and testing, as it contains private medical data for which we did not obtain the necessary approval for public release. Consequently, the trained model weights cannot be made publicly available either.

Contact

Do not hesitate to ask in case you have any questions – my LinkedIn profile: click or write me an email: click.

Special Thanks

Special thanks goes to Michal Španěl, the best supervisor, and to TESCAN 3DIM, s.r.o. for providing the dataset and funding.