Automated Detection of Taurodontism in Panoramic Radiographs Using a YOLOv8-Based Deep Learning Model

Abstract

This study aimed to develop and evaluate a deep learning-based object detection system for the automated detection of taurodontism on panoramic radiographs using the You Only Look Once version 8 (YOLOv8) architecture and to compare the diagnostic performance of different model variants. This retrospective diagnostic accuracy study included 247 panoramic radiographs from patients aged 13 years and older. After applying exclusion criteria, 1631 molar teeth were analyzed. Teeth were classified as taurodontic or normal according to the taurodontism index described by Shifman and Chanannel, which served as the reference standard. All teeth were manually annotated by experienced oral and maxillofacial radiologists. Three YOLOv8 variants (nano, small, and medium) were trained and evaluated under identical conditions. Model performance was assessed using mean average precision at an intersection over union threshold of 0.50 (mAP50) and across thresholds of 0.50-0.95 (mAP50-95), along with precision, recall, F1-score, and inference time. The YOLOv8 medium model achieved the highest performance, with an mAP50 of 0.988 and mAP50-95 of 0.849, precision of 0.956, recall of 0.971, and F1-score of 0.963. The small model provided a favorable balance between accuracy and computational efficiency, while the nano model demonstrated faster inference with lower accuracy. Qualitative evaluation showed accurate localization and classification, including in anatomically complex regions. YOLOv8-based models demonstrated high diagnostic accuracy for automated taurodontism detection on panoramic radiographs. This approach may improve diagnostic consistency and workflow efficiency and supports the integration of artificial intelligence into routine dental imaging practice.