A Deep Learning Application for Identifying Cavities in Dentistry: Utilizing Smartphone-Taken Intraoral Photos

The potential of digital examinations and advanced tele-dentistry applications remains limited due to the lack of universal access to dental care, particularly in remote regions with scarce financial resources. This study aims to standardize the use of digital dental cavity evaluation and progression tracking through image-based techniques. The primary objective of this research was to design an innovative, cost-effective virtual tool capable of predicting dental cavitation from non-standardized images with clinically acceptable reliability. A dataset comprising 1,903 photographs of flattened tooth surfaces, captured using an iPhone 7 by 696 participants, was utilized. The study evaluated four deep learning strategies: YOLOv3, Faster R-CNN, RetinaNet, and Single-Shot Multi-Box Detector (SSD), to detect early cavities and dental infections. The benchmark was set against dentists' evaluations based on image inspections aligned with the Global Cavities Classification and Oversight Method. Among the models, YOLOv3 and Faster R-CNN demonstrated the highest sensitivity for cavitated caries, achieving 88.5% and 72.5%, respectively. However, their sensitivity for visually non-cavitated lesions was lower, at 37.9% and 27%, respectively. All four models achieved specificity levels exceeding 86% for cavitated cavities and over 72% for visually non-cavitated lesions. This research integrates model-based categorization, test-time enhancement, and portable deep learning techniques into a virtual AI-driven system capable of predicting dental cavities from irregular images with adequate clinical reliability. The proposed framework has the potential to enhance access to oral healthcare in underserved and resource-constrained areas, supporting automated diagnoses and advancing tele-dentistry application.