Automatized Detection of Periodontal Bone Loss on Periapical Radiographs by Vision Transformer Networks

Area of Science:

• Dentistry
• Artificial Intelligence
• Computer Vision

Background:

• Periodontal bone loss (PBL) detection is crucial for dental health.
• Convolutional neural networks (CNNs) are currently state-of-the-art for PBL detection.
• Transformer networks represent a new frontier in computer vision.

Purpose of the Study:

• To evaluate the diagnostic performance of various vision transformer networks for automated PBL detection.
• To compare the efficacy of different transformer architectures in identifying PBL on periapical radiographs.

Main Methods:

• Utilized a dataset of 21,819 anonymized periapical radiographs.
• Assessed five vision transformer networks: ViT-base/large, BEiT-base/large, and DeiT-base.
• Statistically determined accuracy (ACC), sensitivity (SE), specificity (SP), predictive values (PPV/NPV), and AUC.

Main Results:

• Overall diagnostic accuracy ranged from 83.4% to 85.2% with AUC from 0.899 to 0.918.
• Performance varied by region: lower anterior teeth showed the highest accuracy (94.1-96.7%) and AUC (0.944-0.970).
• Minor differences were observed among the tested transformer networks for PBL detection.

Conclusions:

• Vision transformer networks demonstrate considerable potential for automated PBL detection.
• Diagnostic performance differs across various tooth regions, with lower anterior teeth yielding the best results.
• Further optimization using larger, manually annotated datasets is necessary to enhance diagnostic performance and clinical applicability.