Abstract Automated diagnostic charting on panoramic radiographs is essential for optimizing clinical workflows and reducing diagnostic variability. This study aims to perform a comparative performance analysis of two next-generation deep learning architectures, YOLOv11 and YOLOv12, for the automated detection of 13 different dental conditions. A hybrid dataset consisting of 2,297 panoramic radiographs was compiled, comprising 1,579 images from a single institution and 718 external images from Roboflow Universe to test generalization. The models were trained to detect conditions including caries, implants, bone loss, and impactions using a standardized training protocol. Performance was evaluated using mean Average Precision (mAP@0.5), Precision, Recall, and F1-score on both internal and external test sets. YOLOv11 demonstrated superior performance compared to YOLOv12, achieving a mAP@0.5 of 0.857 on the internal test set. Furthermore, YOLOv11 exhibited robust generalization capabilities, maintaining a mAP@0.5 of 0.806 on the unseen external dataset. While detection rates were high for well-defined objects like crowns and implants, performance was lower for subtle pathologies such as bone loss and caries. The findings identify YOLOv11 as the more robust and reliable architecture for multi-class dental object detection compared to the newer YOLOv12. The proposed model holds significant potential as a clinical decision support tool, serving as a reliable "second opinion" to enhance diagnostic accuracy and efficiency in daily practice.
Mutlu et al. (Mon,) studied this question.