A YOLOv12-based approach for automatic detection of cephalometric landmarks on 2D lateral skull X-ray images

Cephalometric analysis is the quantitative evaluation of skeletal and soft-tissue relationships on lateral skull radiographs; it underlies diagnosis, treatment planning, and growth assessment in orthodontics. The analysis hinges on cephalometric landmarks which are anatomical reference points whose 2-D coordinates are used to derive angles, distances, and ratios that guide clinical decisions. Manual identification of these landmarks is time-consuming where each image can take from 10 to 15 min and is subject to inter- and intra-examiner variability that can exceed 2 mm, propagating error into subsequent measurements. In recent years, artificial intelligence methods have advanced rapidly and are now widely adopted in medical imaging. This paper proposes an automatic landmark-detection pipeline built on YOLOv12, the newest iteration of the You-Only-Look-Once family. Trained and evaluated on a publicly available cephalometric dataset, the YOLOv12 model successfully localized 53.47% of landmarks within 1 mm and 80.57% within 2 mm.