Risk Assessment Method for Dental Condition Based on Intraoral Imaging and Deep Learning

Deep learning models are increasingly being applied to dental imaging, enabling automated assessment of oral health in environments where traditional clinical imaging equipment is not available. This paper presents a tooth-level risk assessment method based on intraoral photographs captured with consumer-grade smartphone cameras. A dataset of 1,272 images from 330 participants was annotated by licensed dentists with common dental conditions, including caries, plaque, calculus and gingivitis. Based on these annotations, each visible tooth was assigned to one of two clinically motivated risk categories. A YOLOv8-based segmentation and classification model was trained to localize each tooth and predict its associated risk level. The results show that the model provides consistent performance across risk categories, with a macroaveraged F1-score of approximately 0.75 at the optimal confidence threshold. Most classification errors occurred between neighbouring risk levels, while confusions between the extreme classes were uncommon. Aleatoric uncertainty, estimated via Test-Time Augmentation, correlated strongly with visually ambiguous or low-quality images, providing an internal indication of the reliability of each prediction. These findings demonstrate that smartphone-based intraoral imaging, combined with deep learning, can be used as a measurement pipeline for tooth-level risk assessment, producing meaningful risk estimates together with uncertainty information relevant for screening and self-monitoring applications.