Deep Learning-Based Dental Caries Diagnosis on Panoramic Radiographies: Performance of YOLOv8 Versus Human Observers

Objectives: To evaluate the diagnostic performance of a YOLOv8x-based deep learning model for detecting approximal, occlusal and buccal caries on paediatric panoramic radiographs and to compare its performance with human observers with different levels of clinical experience. Methods: A total of 1526 panoramic radiographs obtained from children aged 5–12 years were retrospectively analysed. Approximal, occlusal, and buccal caries in primary molars were annotated and used to train a YOLOv8x object-detection model. Model performance was evaluated on an independent test set and compared with three human observers: an intern dentist (ID), a novice specialist student (NSS), and an experienced specialist student (ESS). Diagnostic performance was assessed using precision, sensitivity, F1 score, and true positive counts. Results: The YOLOv8x model demonstrated moderate performance in detecting approximal caries (F1 score: 0.576) but showed limited performance for occlusal caries (F1 score: 0.24) and failed to detect buccal caries. The AI model showed lesion-dependent performance. For approximal caries, it performed comparably to ESS observers (p > 0.05) and better than ID (p < 0.001). Performance was poor for buccal caries (p < 0.001), and intermediate for occlusal caries, with no difference from NSS or ESS (p > 0.05) but lower than ID (p < 0.001). Overall, performance was comparable to experienced observers (p > 0.05) and superior to less experienced observers (p < 0.001). Conclusions: The YOLOv8x model achieved diagnostic performance comparable to less experienced clinicians in detecting dental caries on paediatric panoramic radiographs but did not reach expert-level accuracy. These findings suggest that deep learning models may serve as supportive tools in panoramic caries assessment rather than replacements for expert interpretation.