A Semi-Supervised Transformer-Based Deep Learning Framework for Automated Tooth Segmentation and Identification on...

Area of Science:

• Artificial Intelligence in Dentistry
• Medical Imaging Analysis
• Deep Learning Architectures

Background:

• Automated tooth segmentation and identification are vital for digital dental workflows.
• Current deep learning models struggle with accuracy in partially edentulous individuals.
• A need exists for robust methods and standardized datasets for dental image analysis.

Purpose of the Study:

• To introduce SemiTNet, a novel semi-supervised Transformer-based framework for enhanced tooth segmentation and identification.
• To specifically address performance limitations in partially edentulous cases.
• To establish the open-source TSI15k dataset as a unified benchmark for dental radiography studies.

Main Methods:

• Development of SemiTNet using a semi-supervised learning approach with label-guided teacher-student knowledge distillation.
• Utilizing a Transformer-based architecture for improved feature extraction.
• Training and validation on the large-scale TSI15k dataset comprising 16,317 panoramic radiographs (1589 labeled, 14,728 unlabeled).

Main Results:

• SemiTNet achieved superior performance in tooth segmentation and identification compared to five state-of-the-art networks.
• Near-perfect accuracy (over 99.69%) for fully dentate individuals and excellent accuracy (over 93%) for partially edentulous individuals.
• Statistically significant improvements in tooth identification for edentulous cases, with minimal model size.

Conclusions:

• SemiTNet demonstrates high efficacy, outperforming existing complex models, particularly in challenging partially edentulous cases.
• The developed SemiTNet framework offers a robust solution for automated dental image analysis.
• The open-source TSI15k dataset provides a valuable resource for advancing research in dental radiography.