Development and validation of an automated deep learning-based cervical vertebral maturation classification system using object detection and morphological analysis

Accurate assessment of skeletal maturity is fundamental in orthodontics, as treatment timing during growth has a decisive influence on skeletal response, treatment efficiency, and long-term stability. The cervical vertebral maturation (CVM) method provides a practical, radiation-free approach to skeletal maturity assessment using routine lateral cephalograms; however, its clinical utility remains constrained by interpretive subjectivity, transitional morphological ambiguity, and inconsistent reproducibility, particularly at pubertal boundary stages. Although artificial intelligence (AI) has been proposed as a means to standardise CVM staging and support growth-based orthodontic decision-making, progress has been limited by the absence of robust, well-calibrated ground truth datasets and by automated systems that insufficiently reflect clinical reasoning. This study aimed to develop and validate a fully automated CVM classification system that integrates object detection and morphology-driven analysis using deep convolutional neural network architectures, and to benchmark its diagnostic performance against calibrated expert consensus. Recognising that annotation quality is a critical determinant of AI performance, a structured multi-expert calibration and consensus framework was implemented to establish reproducible reference labels. Under these calibrated criteria, 1,047 anonymised lateral cephalograms from patients aged 6 to 17 years were annotated into six CVM stages by two orthodontic experts. The dataset was partitioned into training (80%), validation (10%), and test (10%) subsets. A two-stage automated pipeline was developed, combining YOLOv8 for precise localisation of the second to fourth cervical vertebrae (C2–C4) with LightGBM for classification based on extracted morphological features. This design decoupled anatomical localisation from stage classification, enabling targeted analysis of biologically relevant structures while maintaining model interpretability. Model performance was evaluated using accuracy, precision, recall, F1-score, and agreement-based metrics. Inclusive accuracy and unweighted kappa (κ) were employed to reflect clinically acceptable adjacent-stage variability, while intraclass correlation coefficients were used to quantify intra-observer reliability. The proposed system demonstrated high diagnostic performance, achieving expert-comparable strict classification accuracy and improved inclusive accuracy consistent with clinical staging tolerance. Agreement with expert consensus was substantial to near-perfect across performance metrics, particularly for morphologically distinct pubertal and post-pubertal stages. Residual misclassifications were predominantly confined to biologically transitional CVM stages, reflecting intrinsic morphological continuity rather than systematic algorithmic error. External validation further confirmed patterned behavioural consistency with expert staging and yielded interpretable morphological cues that aligned with established clinical reasoning. Taken together, these findings indicate that reliable, automated CVM staging is achievable when AI development is grounded in expert-calibrated annotation and biologically informed model design. The proposed framework demonstrates the potential for CVM assessment to become more reproducible and clinically consistent through integration of expert-calibrated annotation and AI-assisted analysis. By integrating consensus-based ground truth with object detection and morphology-driven classification, this study demonstrates the translational potential of AI to enhance consistency and reliability in skeletal maturity assessment within orthodontic practice.