Automatic detection of interproximal, occlusal and secondary caries in bi̇tewing radiographies with artificial intelligence: İnterface design for clinical use

Early detection of dental caries can prevent invasive treatment and thus save healthcare costs; however, interproximal caries are difficult to detect by clinical examination alone, and bitewing radiography is the gold standard for diagnosing demineralized interproximal caries. YOLO algorithm is a convolutional neural network (CNN) based deep learning model, and YOLOv8 is the latest YOLO model that can be used for object detection, image classification, and instance segmentation tasks. This study aimed to determine the success of the YOLOv8 algorithm in detecting secondary, occlusal and interproximal (d1-d2-d3) caries on bitewing radiographs. In this study, 860 bitewing radiographs were collected from the faculty database. The number of bitewing images was increased to 3440 with augmentation methods. Oral radiologists labeled caries lesions on bitewing radiographs as d1, d2, d3, secondary, and occlusal. The data was randomly split into 80% training, 10% validation, and 10% testing. The weight file for dental caries was obtained by training the YOLOv8 algorithm with the transfer learning method. Using the obtained weights, tooth decay that the algorithm had not seen before was automatically detected in the test radiographs. Test results were evaluated by two oral radiologists and performance criteria were calculated. In the test images, the average precision, average sensitivity and average F1 score values for secondary, occlusal and interproximal caries were obtained as 0.977, 0.932 and 0.954, respectively. Trained on bitewing radiographs, the YOLOv8 algorithm detected different types of dental caries lesions with high success rates.