Automated cervical vertebral maturation staging using deep learning: an anatomically guided framework for orthodontic skeletal maturity assessment

Accurate assessment of skeletal maturity is essential for orthodontic diagnosis and treatment planning, particularly when growth-modifying interventions depend on the timing of the pubertal growth spurt. Cervical vertebral maturation (CVM) analysis provides a practical method for skeletal maturity assessment using routine lateral cephalometric radiographs without additional radiation exposure. However, conventional CVM staging remains susceptible to examiner variability, particularly during transitional maturation stages. Artificial intelligence (AI) has shown potential in improving diagnostic consistency in medical imaging, although many previously reported automated CVM systems have relied on relatively small datasets or limited annotation protocols. This study aimed to develop and evaluate an anatomically guided artificial intelligence framework for automated CVM staging. A retrospective dataset of 1,047 anonymised lateral cephalometric radiographs from patients aged 6–17 years was used for model development. CVM stages were determined using the modified Baccetti classification system by calibrated orthodontic examiners following structured consensus procedures. Chronological age and sex were not used as input variables during model development. The dataset was partitioned into training (80%) and internal testing (20%) subsets at the patient level. A two-stage framework consisting of YOLOv8 vertebral localisation followed by Light Gradient Boosting Machine (LightGBM)-based morphological classification was implemented. Model performance was evaluated using accuracy, precision, recall, F1-score, receiver operating characteristic analysis, and Cohen’s kappa statistics. Independent testing was performed using 30 previously unseen radiographs. Inter- and intra-examiner reliability demonstrated excellent agreement, with ICC values ranging from 0.946 to 0.967 and Cohen’s kappa values ranging from 0.852 to 0.891. The proposed framework achieved an adjacent-stage accuracy of 96.67% and an exact-stage accuracy of 80.00% on the independent testing dataset. Most misclassifications occurred between adjacent maturation stages, particularly between Stages III and IV. Cohen’s kappa analysis demonstrated substantial to near-perfect agreement between automated predictions and expert consensus. The proposed framework demonstrated promising performance for automated CVM staging using an anatomically guided two-stage architecture trained on calibrated expert annotations. While substantial agreement with expert consensus was observed within the study dataset, the findings should be interpreted as preliminary due to the retrospective design and limited independent testing sample. Further multicentre validation and prospective evaluation are required before routine clinical implementation.