Detection of lesions on panoramic radiographs using deep learning method

Aim: The aim of this study was to evaluate the performance of a deep learning method for the detection of lesions on panoramic radiographs. Material and Method: In this retrospective study, the panoramic radiographs of patients over the age of 18 who had applied to the Department of Oral and Maxillofacial Radiology at Ordu University Faculty of Dentistry were analyzed. A total of 2371 panoramic radiographs with lesions were divided into two separate data sets: radiolucent and radiopaque. The radiolucent lesion dataset comprised 1325 panoramic radiographs and 1561 labels, whereas the radiopaque lesion dataset included 1046 panoramic radiographs and 1323 labels. The lesion borders were delineated and labeled by identifying the lesion area. The YOLOv8 algorithm was used to detect lesions. To evaluate the images, three different scenarios were developed. In the first scenario, the detection of radiolucent and radiopaque lesions was evaluated. In the second scenario, radiolucent and radiopaque lesions were evaluated based on their localization. In the third scenario, the detection performance of radiolucent and radiopaque lesion in the entire dataset was evaluated. 70% of the dataset was utilized for training, 15% for validation, and 15% for testing. The precision, recall, and F1-score measures were employed to assess model success. Results: The precision, recall, and F1-score for the detection of radiolucent lesions were 0.91, 0.72, and 0.80, respectively. When detection scores were analyzed based on the localization of radiolucent lesions, it was discovered that the posterior mandible had the highest precision, recall, and F1-scores, with values of 0.75, 0.74, and 0.74, respectively. The lowest scores were 0.57, 0.17, and 0.26, respectively, in the posterior maxilla. The precision, recall, and F1-score for the detection of radiopaque lesions were 0.73, 0.69, and 0.71, respectively. In addition, similar to radiolucent lesions, the highest values in radiopaque lesions were in the posterior mandible. It was observed that the performance of the model was negatively affected when the entire data set was evaluated collectively. Conclusion: The results of the study indicate that the YOLOv8 algorithm can be utilized to detection lesions in panoramic radiographs.