Deep Learning-Based Detection of Overhanging Restorations and Calculation of Radiographic Alveolar Bone Loss Percentage with YOLOv11 and YOLOv12.

The objective of this study is to evaluate the YOLOv11 and YOLOv12 deep learning models for the detection of overhanging restorations in panoramic radiographs and the calculation of the associated percentage of alveolar bone loss. Overhanging restorations were labeled using bounding boxes on 550 panoramic radiographs, while alveolar bone loss was labeled using position estimates of the enamel-cementum junction, alveolar crest, and root apex points. Two distinct models were developed, each based on the Ultralytics YOLOv11 and YOLOv12 architectures, and these models underwent a training, validation, and testing process on disparate datasets. The overhanging restoration detection model for YOLOv12 achieved a mean average precision at the 50-object level of 98.3%, a precision of 97.2%, and a recall of 96.5%. For YOLOv11, these values were 98.2%, 97.1%, and 96.3%, respectively. In the context of the bone loss position prediction model for YOLOv12, the values were determined to be 94.5%, 95.1%, and 96.1%, while for YOLOv11 these values were 94.1%, 94.5%, and 91.1%, respectively. The calculation of alveolar bone loss as a percentage is of paramount importance for the purpose of grading and staging according to the prevailing periodontal disease classification system. Nevertheless, the process is time-consuming, and it is susceptible to error, which can lead to incorrect or incomplete diagnoses. In this study, the percentage of alveolar bone loss was estimated by the proposed models with a technical inference speed of approximately 100 ms under controlled experimental conditions. These results demonstrate that deep learning-based approaches may offer potential advantages in supporting clinical decision-making systems. However, further studies involving clinical time-motion analyses and diagnostic consistency comparisons are needed. It is therefore hypothesized that, with appropriate validation, the models developed in this study could eventually support clinical decision-making processes.