Artificial Intelligence-Assisted Quadrant-Level Treatment-Site Estimation for Patient Safety

Introduction and aims Safety interventions in dentistry, including checklists, have been reported; however, it remains difficult to demonstrate sufficient effectiveness in preventing treatment site misidentification, including wrong tooth extraction. This study aimed to establish a foundation for an artificial intelligence (AI)-based patient-safety measure by developing a system that estimates the dental treatment site at the quadrant level from images captured during dental procedures. Methods Single-frame still images (n = 2,446; training 1,780, validation 445, test 221) were extracted from clinical videos recorded during treatment performed by a single dentist between 27 July 2022 and 21 December 2023. Treatment-site estimation models were developed using VGG16, a convolutional neural network-based image-classification model, and YOLOv9, an object-detection model. Performance was compared using three metrics: quadrant accuracy (correct upper/lower and left/right), vertical accuracy (correct upper/lower only), and lateral accuracy (correct left/right only). Results Among five approaches (four VGG16 conditions and one YOLOv9 condition), YOLOv9 achieved the highest performance across all metrics, with quadrant accuracy of 88.7%, vertical accuracy of 96.8%, and lateral accuracy of 91.9%. For VGG16-based approaches, performance across these metrics varied according to image pre-processing. Conclusion This system may support detection and prevention of quadrant-level errors, which are reported to be the second most common type of wrong tooth extraction following adjacent-tooth errors. Clinical relevance AI-based treatment-site estimation may provide an additional preventive safety layer in dentistry by supporting clinicians’ judgement and enabling detection of quadrant-level site misidentification during dental procedures.