Automatic detection and localization of all impacted teeth and third molar teeth according to winter classification in panoramic radiographies with deep learning: Interface design for clinical use

Impacted teeth are common and their extent and distribution in different parts of the jaws can vary significantly. Any tooth can be impacted, but the most commonly impacted teeth are the lower third molars. Classification of impacted third molars allows us to recognize the degree of impaction and choose the most appropriate methodology for the surgical procedure, thereby reducing the risk of complications. Famous for its object detection feature, YOLO is a widely used algorithm. The YOLOv8 algorithm is the latest version of the YOLO family as of the date of this study. The aim of this study is to detect all impacted teeth on panoramic radiographs and automatically classify impacted third molars according to the Winter classification system based on deep learning. For this purpose, in this study, panoramic radiographs collected from the faculty database were reproduced with augmentation methods, and 2000 images were obtained for the Winter classification system and 2394 images were obtained for the detection of impacted teeth. Images were labeled by oral radiologists and the coordinates of the labels were downloaded in YOLO format. With the deep network architecture used, 80% of the data was trained, 10% was validated, and 10% was tested with the training weights obtained by improving the model predictions. Results were evaluated according to performance criteria including precision, sensitivity, and F-1 score. According to the Winter classification test results, the average precision value of mesioangular, distoangular, horizontal, vertical, buccolingual and other impacted tooth orientations was obtained as 0.972, the average sensitivity value was 0.967 and the average F1-score value was 0.969. According to the test results in which impacted teeth were detected, the average precision value was obtained as 0.991, the average sensitivity value as 0.995, and the average F1-score value as 0.995. 440 of the impacted teeth were detected correctly, 4 were detected incorrectly, and 2 were not detected by the algorithm. According to the results obtained, the YOLOv8 algorithm showed a successful performance in detecting impacted teeth and classifying third molars according to the Winter classification system.