Effectiveness of Artificial Intelligence in the Automatic Diagnosis of Lesions Detected in Panoramic Radiographs

This study was aimed at evaluating the effectiveness of artificial intelligence (AI) in detecting jaw cysts and tumors, analyzing lesion content, and establishing diagnoses on panoramic radiographs. To construct the dataset, panoramic radiographs (PRs) of 124,800 patients aged 10-90 years were retrospectively evaluated. A total of 1300 PRs meeting the inclusion criteria, confirmed by histopathological reports, were included. Five primary tasks were investigated: lesion detection, histopathological diagnosis, lesion content assessment, jaw region segmentation, and benign/malignant differentiation. Lesion boundaries were manually delineated using CranioCatch software (CranioCatch, Eskisehir, Turkey), generating 1166 annotations. The YOLOv8 architecture was used for model development. Results obtained via the confusion matrix were evaluated using performance metrics including sensitivity, specificity, and F1 score. For lesion detection, sensitivity, specificity, and F1 score were 0.85, 0.80, and 0.82, respectively; for histopathological diagnosis, these values were 0.77, 0.86, and 0.81, respectively. The highest overall values for histopathological diagnosis were 0.88 for mucous retention cysts and 0.83 for ameloblastoma. For lesion content detection, the model achieved sensitivity, specificity, and F1 score values of 0.73, 0.72, and 0.73, respectively, while jaw region segmentation yielded values of 0.99, 1.00, and 0.99, respectively. For benign/malignant differentiation, scores were 0.90, 0.74, and 0.82, respectively. The YOLOv8 algorithm demonstrated high accuracy in lesion detection, content classification, and benign/malignant differentiation, with particularly outstanding performance in jaw region identification. The findings highlight the strong potential of this algorithm to support clinicians in radiological assessments; however, comparative model analyses and multicenter validation studies are required before it can be recommended as a clinical decision-support tool.