Diagnostic accuracy of artificial intelligence-assisted caries detection: a clinical evaluation

This clinical study aimed to evaluate the practical value of integrating an AI diagnostic model into clinical practice for caries detection using intraoral images.

In this prospective study, 4,361 teeth from 191 consecutive patients visiting an endodontics clinic were examined using an intraoral camera. The AI model, combining MobileNet-v3 and U-net architectures, was used for caries detection. The diagnostic performance of the AI model was assessed using sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and accuracy, with the clinical diagnosis by endodontic specialists as the reference standard.

The overall accuracy of the AI-assisted caries detection was 93.40%. The sensitivity and specificity were 81.31% (95% CI 78.22%-84.06%) and 95.65% (95% CI 94.94%-96.26%), respectively. The NPV and PPV were 96.49% (95% CI 95.84%-97.04%) and 77.68% (95% CI 74.49%-80.58%), respectively. The diagnostic accuracy varied depending on tooth position and caries type, with the highest accuracy in anterior teeth (96.04%) and the lowest sensitivity for interproximal caries in anterior teeth and buccal caries in premolars (approximately 10%).

The AI-assisted caries detection tool demonstrated potential for clinical application, with high overall accuracy and specificity. However, the sensitivity varied considerably depending on tooth position and caries type, suggesting the need for further improvement. Integration of multimodal data and development of more advanced AI models may enhance the performance of AI-assisted caries detection in clinical practice.