Skip to Main content Skip to Navigation
Theses

Automatisation du traitement des imageries tridimensionnelles dento-maxillo-faciales par apprentissage profond : application à la segmentation et à la céphalométrie

Abstract : The clinical use of three-dimensional (3D) dentomaxillofacial imaging has developed significantly in recent years, allowing for improved diagnosis and planning of some orthodontic and orthodontic-surgical treatments. However, the processing of these 3D images remains restrictive and relies on many manual steps, requiring several levels of validation, time and trained operators. The clinical routine is still largely based on the use of 2D methods, which are not well adapted for patients with complex facial deformities such as important asymmetries or craniofacial syndromes.The main objective of this work was to implement deep learning models in order to automate two steps in the processing of these 3D images: (1) the reconstruction of 3D surface models, a process called "segmentation", and (2) the placement of anatomical landmarks for 3D cephalometric analysis. The evaluation of these models was performed on an original database of patients with varied and marked facial deformities, comparing the performance of the algorithm with that of experts on the basis of clinically relevant metrics.In a test database of 153 CT scans, the automated segmentation had a surface Dice Similarity Coefficient at 1mm of 98.03 ± 2.48%, with 148 scans having a mean score which cleared the 95% limit for clinical significance. In a test database of 37 CT scans, the mean error of cephalometric landmark localization was 1.0 ± 1.3 mm, and 90.4% of predictions were within 2 mm of the reference. A broader validation, including data from other clinical centers, will need to be performed to assess the generalizability of these results. Three clinical cases illustrate the perspectives of clinical applications of these results.
Complete list of metadata

https://pastel.archives-ouvertes.fr/tel-03772880
Contributor : ABES STAR :  Contact
Submitted on : Thursday, September 8, 2022 - 3:54:50 PM
Last modification on : Friday, September 9, 2022 - 3:43:54 AM

File

112604_DOT_2022_archivage.pdf
Version validated by the jury (STAR)

Identifiers

  • HAL Id : tel-03772880, version 1

Citation

Gauthier Dot. Automatisation du traitement des imageries tridimensionnelles dento-maxillo-faciales par apprentissage profond : application à la segmentation et à la céphalométrie. Biomécanique [physics.med-ph]. HESAM Université, 2022. Français. ⟨NNT : 2022HESAE041⟩. ⟨tel-03772880⟩

Share

Metrics

Record views

0

Files downloads

0