An Artificial Intelligence Model for the Automatic Classification of the Cervical Vertebral Maturation Stages

To develop an artificial intelligence (AI) model for automatic classification of cervical vertebral maturation (CVM) stages on lateral cephalometric radiographs (LCR).

Overall, 1140 LCRs were labeled and classified according to the CVM user guide by McNamara and Franchi using a 2-stage pipeline. In the first stage, the LCRs were cropped to extract the region of interest (ROI) using the YOLOv11s.pt model. In the second stage, the extracted ROI were used to train the ResNet-101 deep learning (DL) model to classify CVM stages. Model performance was evaluated using accuracy, recall, precision, and F1-score. Performance comparison between human and AI was conducted by calculating the quadratic weighted kappa and ICC values.

Intra- and inter-examiner reliabilities were measured using weighted kappa values of 0.992 and 0.953, respectively. The ResNet-101 DL model showed 84% overall accuracy, and the quadratic weighted kappa and ICC values between human and AI readings were 0.93 and 0.95, respectively.

The proposed DL model achieved 84% accuracy in automated CVM stage classification. This AI-based framework enhances objectivity and efficiency of skeletal maturation assessment through standardized and reproducible classification. Accordingly, it reduces the influence of subjective evaluation, particularly where clinical workload is high.

Automated CVM assessment improves objective evaluation of skeletal growth. To our knowledge, this study is the first to integrate YOLOv11-based localization with ResNet-101 for CVM classification, enabling an automated workflow supported by a curated and balanced dataset and a reproducible framework.