Evaluation of Deep Learning for Caries Detection With Fine-Grained Classification and Postprocessing Improvements.

Area of Science:

• Artificial Intelligence in Dentistry
• Deep Learning for Medical Imaging
• Dental Diagnostics

Background:

• Existing deep learning for dental caries detection lacks comprehensive classification and focuses on limited lesion areas.
• There is a need for advanced AI models capable of tooth-instance detection and fine-grained caries classification.

Purpose of the Study:

• To develop and evaluate advanced deep learning models for tooth-instance based dental caries detection.
• To implement fine-grained classification of caries using the International Caries Detection and Assessment System (ICDAS) criteria.
• To introduce two novel correction methods to enhance model stability and accuracy in complex scenarios.

Main Methods:

• Utilized a dataset of 1200 intraoral images, augmented to 8,754 images with detailed tooth annotations.
• Trained and tested three state-of-the-art deep learning models: YOLO-v8, YOLO-v9, and YOLO-NAS.
• Implemented postprocessing correction techniques involving weighted average scoring and adaptive confidence adjustment based on spatial tooth relationships.

Main Results:

• The proposed correction methods significantly improved mean Average Precision (mAP) by 2.8% to 4.7% (p < .01) across the YOLO models, with YOLO-v8 achieving 72.9% mAP.
• Overall precision and recall increased by 3.8% and 5.6%, respectively, with a minor decrease in Frames Per Second (FPS).
• Demonstrated notable improvements in detecting moderate caries and enhanced model robustness.

Conclusions:

• The developed correction methods effectively enhance existing AI frameworks for dental caries detection.
• The fine-grained classification capability holds significant clinical value, assisting dentists in formulating personalized treatment plans.
• This research is expected to advance the application of AI in dentistry and stimulate further investigation.