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:

• We are dealing with difficult cases of orthodontics patients, i.e. the input meshes are obtained by scanning the dental arches of unhealthy patients with various kinds of teeth misplacements, shiftings, etc.
• It should not rely on strong PCs with modern GPUs - it should run on normal consumer computer (imagine for example the computer of your clinitian) in seconds.

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.

Available Testing Dataset

Unfortunately, we are not allowed to share the whole dataset which we’ve used for the training, validation, and testing. However, to compare are results with other frameworks for orthodontic landmark detection, we are sharing a portion of the test set. You can download 20 meshes for the testing here: tbd (NOTE: Currently, we are waiting for the approval from the owner of the dataset).

Results on the Public Testing Subset

For a fair comparison, we’ve measured the landmarking accuracy and teeth detection accuracy solely on the public portion of the test set. Among all the test cases, we’ve tried to share the challenging ones. Our method achieves following results: tbd after the approval.

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.