Detection of Jaw Lesions on Panoramic Radiographs Using Deep Learning Method

Abstract

This study aimed to evaluate and compare the performance of state-of-the-art deep learning models for detecting and segmenting both radiolucent and radiopaque jaw lesions on panoramic radiographs. A total of 2371 anonymized panoramic radiographs containing jaw lesions were retrospectively collected and categorized into radiolucent and radiopaque datasets. Expert annotation was performed to delineate lesion boundaries and assign anatomical localization (anterior/posterior maxilla and mandible). Four deep learning architectures-YOLOv8, YOLOv11, Mask R-CNN, and RT-DETR-were trained and evaluated under three experimental scenarios: (I) training without spatial labels, (II) data augmentation with unlabeled background images, and (III) inclusion of spatial localization annotations. Performance metrics included precision, recall, F1-score, and mean average precision (mAP@0.5 and mAP@0.5-0.95), with paired t-tests used for statistical comparison. In Scenario I, YOLOv11x-seg and YOLOv8x-seg achieved the highest segmentation performance for radiolucent and radiopaque lesions, respectively. For detection, YOLOv8x performed best on radiolucent lesions, while RT-DETR-L outperformed others on radiopaque lesions. In Scenario II, while YOLOv8x-seg achieved the best segmentation results across both lesion types, RT-DETR-L demonstrated superior detection performance, particularly for radiolucent lesions. In Scenario III, RT-DETR-L consistently outperformed all models across both lesion types. This study demonstrates the potential of state-of-the-art deep learning models for effective detection of lesions in panoramic radiographs. The developed models may offer valuable support to clinicians in lesion evaluation; however, it is recommended that they be employed primarily as decision support tools within clinical workflows, rather than as standalone diagnostic systems.