2026 Volume 68 Issue 2 Pages 48-58
Purpose: Intraoral photographs are widely used for diagnosis, treatment planning, and documentation in the field of dental surgery; however, directly assessing disease activity from these images remains challenging. Bleeding on probing (BOP) is a key clinical indicator of periodontal inflammatory activity, yet conventional assessment requires probing with its attendant discomfort and bleeding risk. This study was aimed at developing an AI-based model to predict BOP from pre-probing intraoral photographs and to compare its performance with that of periodontists.
Methods: A total of 427 intraoral photographs of the maxillary anterior teeth obtained from 254 patients who visited the Asahi University Medical and Dental Center were analyzed, and deep learning using a neural network was established with the presence or absence of BOP as the ground-truth label. AlexNet architecture and an automatically optimized network were employed as the training models. The same images were also independently evaluated by ten dentists, and their assessments were compared with the AI predictions.
Results: The accuracy of AlexNet for predicting BOP was 50.0%, whereas that of the automatically optimized network was 80.7%. The diagnostic accuracy for periodontists ranged from 72.9% (highest) to an average of 67.0%. The AI model demonstrated higher accuracy as compared with the periodontists.
Conclusions: The AI model for predicting BOP from intraoral photographs exhibited diagnostic performance comparable to or exceeding that of the periodontists, suggesting its potential clinical applicability. Future directions include developing a full-arch framework, estimating periodontal inflamed surface area (PISA) and exploring applications in screening and self-care support tools.
