Automatic landmark detection in cephalometric lateral radiograph using deep learning one-stage detectors

Cephalometric radiography is the gold standard in analyzing, assessing, and diagnosing the relationship between teeth, jaws, and skeletal structures. In previous studies, artificial intelligence has been widely used to identify anatomical landmarks on lateral cephalometric radiographs, reducing errors and speeding up the analysis. This research aims to explore the automation of the landmarks detection process using You Only Look Once version 5 (YOLOv5). A total of 350 cephalometric radiographs were selected as learning data that trained YOLOv5 and the number of target labels was 17 pairs of landmark. The assessment is evaluated by mean radial error, success detection rate, and matrix evaluation. The system’s accuracy in predicting 17 pairs of landmarks in the image within the error range of 2 mm reaches 90.44% on average and the MRE result is 0.9553 mm. The matrix evaluation results show the confidence value show that the most successful predictions were obtained from Sella (1.00), Menton (0.94), Intermolar (0.93), and the worst from LA (0.04), point B (0.47), and UA (0.14). The YOLOv5 model demonstrates strong landmark detection capabilities but still needs improvement. To achieve the same level of robustness as the gold standard, further research can focus on enhancing datasets and improving YOLO models.