Mobile dental diagnostics : deep learning for teeth numbering and plaque detection

Abstract

Oral diseases represent a significant global health burden, with barriers such as cost and limited access that hinder timely care. With the widespread availability of smartphones and advancements in deep learning, this research explores low-cost mobile dental diagnostics. The primary objective was to develop and evaluate deep learning models for key dental tasks: teeth localization and numbering, and plaque detection, using images captured by ordinary smartphone cameras. Employing a quantitative and experimental design, the study trained models using the Digileap dataset, which contains smartphone images collected under minimally controlled conditions. Significant preprocessing was incorporated to handle inherent image variability. A customized Mask R-CNN model was utilized for tooth localization and numbering, while both a Mask R-CNN and a custom CNN binary classifier were investigated for plaque detection. Model performance was assessed using mAP for tooth localization and numbering, and PR-AUC and macro F1-score for plaque binary classification. The first model's robustness was evaluated via 10-fold cross-validation and testing on an external dataset. A proof-of-concept application, OCID, was also developed to demonstrate the technical feasibility of using two models in a streamlined dental diagnostics application. The findings indicate that integrating AI with smartphones holds promise for teledentistry, public screening, and patient empowerment, demonstrating successful tooth localization and numbering as a crucial foundational step for automated dental systems. While challenges in plaque detection from undyed images persist, the research highlights the potential of AI to support disease detection and emphasizes the need for high-quality, representative datasets for advancing mobile dental diagnostics