Bone age assessment through the analysis of cervical vertebrae in lateral cephalometric radiographs using semantic segmentation and object detection

Cervical vertebral maturation (CVM) is an important and reliable indicator for accurately identifying growth stages in orthodontic treatment. Although hand-wrist radiographs are conventionally used for this purpose, they impose additional radiation exposure, are prone to inter-observer variability, are tedious, and require extra clinical time. This thesis aimed to develop an automated artificial intelligence workflow capable of assessing CVM from routine lateral cephalograms without additional imaging and within clinical speed limits. For model training, 3,750 cephalograms were selected from the archives of the Department of Orthodontics at Van Yüzüncü Yıl University. The proposed modular system consists of three stages. First, YOLO-v11 rapidly localizes the C2-C4 region within milliseconds, achieving 99.8 % precision, 99.8 % recall, and an mAP₀․₅ of 0.9935, ensuring highly accurate detection. Second, Attention-ResUNet, utilizing multi-scale attention blocks, precisely segments the bone contours at the pixel level, obtaining Dice = 0.956 and Precision = 0.954 scores that surpass clinical thresholds for anatomical accuracy and contour sensitivity. Third, NFNet classifies skeletal maturation according to the six-stage CVMS system, yielding 90.0 % accuracy, a ROC-AUC of 0.990, and specificity of 0.980 on the independent test set, effectively minimizing classification errors. Data were partitioned using five-fold stratified cross-validation; mean fold accuracy was determined as 88.5 % with an RSD of 1.79 %, demonstrating consistent model performance. These results provide competitive values over larger datasets compared to literature reports using smaller samples. Additionally, consolidating detection, segmentation, and classification into one step significantly shortens processing time and reduces clinician workload. The developed system offers a rapid, low-dose, interpreter-independent, innovative and effective alternative for CVM-based age estimation and is ready for clinical implementation. Future work should investigate multitask learning to concurrently estimate dental age or frontal sinus development and integrate three-dimensional low-dose imaging data.