Detection of dens invaginatus in panoramic radiographs with artificial intelligence algorithms developed with deep learning method

Objective: The aim of this study was to evaluate the success and reliability of an artificial intelligence application developed using YOLOv5 and YOLOv8 deep learning models with two different labeling for the detection of teeth with dens invaginatus in panoramic radiographs. Materials and Methods: In this study, 656 panoramic radiographs of patients aged 8 to 18 were labeled for teeth with dens invaginatus in the upper anterior region using segmentation and detection methods. The labeling of the images was performed using the CranioCatch software (Eskişehir, Turkey) by a research assistant with 2.5 years of experience and a pediatric dentist with 15 years of experience. Each model was evaluated based on performance criteria, including precision, accuracy, and F1 score. Results: In this study, the detection method of the YOLOv5 model yielded precision, recall, and F1 scores of 0.945, 0.887, and 0.915, respectively. For the segmentation method of the same model, these values were 0.905, 0.928, and 0.916, respectively. In the case of the YOLOv8 model, the detection method achieved precision, recall, and F1 scores of 0.950, 1, and 0.974 while for the segmentation method, these values were 0.940, 0.994, and 0.966, respectively. Conclusion: Based on the results of this study, it has been observed that the deep learning models developed were successful in detecting dens invaginatus. It is believed that deep learning-supported systems could be integrated into pediatric dentistry practice and serve as a decision support mechanism for clinicians. Keywords: Dens invaginatus, Panoramic radiography, Artificial intelligence, YOLO