Detection of Cervical Vertebrae Using Object Detection and Semantic Segmentation Methods in Lateral Cephalometric Radiographs

Abstract

This study proposes an artificial intelligence-based method for the detection and semantic segmentation of C2, C3, and C4 cervical vertebrae in lateral cephalometric radiographs. The dataset used in the research consists of 3,085 lateral cephalometric radiographs provided by the Department of Orthodontics, Faculty of Dentistry, Van Yüzüncü Yıl University. Following evaluations by expert clinicians, 2,520 radiographs that met the criteria for diagnostic accuracy and clinical suitability were included in the study. In the initial stage, vertebral regions were identified using YOLOv8 and YOLOv11 object detection models, and these areas were meticulously annotated using QuPath software. The labelled data were then subjected to segmentation using advanced deep learning models such as Attention-UNet, Attention-ResUNet, SEEA-UNet, and ResAt-UNet.The study revealed that the object detection models achieved a high performance with an accuracy of 99.8%. Among the segmentation models, Attention- ResUNet demonstrated the best performance with an accuracy of 99.25%, while the ResAt-UNet model stood out with its balanced generalization capacity. The generated binary masks provided a reliable dataset for bone age estimation and skeletal maturity analysis. This study aims to reduce radiation exposure and streamline clinical workflows by eliminating the need for additional imaging. The findings indicate that AI-supported methods minimize errors caused by manual assessments and ensure standardization in skeletal analysis. It is anticipated that these methods could be widely utilized in orthodontic and pediatric medical applications in the future.