Deep Learning Automates Cervical Vertebral Maturation Staging

Given that manual cervical vertebral maturation (CVM) analysis on lateral cephalograms is a standard method for assessing skeletal maturity in orthodontics​, this study aimed to evaluate whether deep convolutional neural networks (CNNs) can reliably automate CVM staging. This retrospective study examined 1,000 lateral cephalometric radiographs from adolescent patients (approximately 10–20 years old, both sexes) collected between 2015 and 2021 at a Class A tertiary hospital, with institutional ethical approval. Multiple deep CNN models (including a ResNet-50 architecture) were trained and fine-tuned for automated classification of CVM stages. Model performance was evaluated using classification accuracy and Cohen’s kappa statistic to measure agreement with reference standards provided by experienced orthodontists. Additionally, the effects of integrating 3D cone-beam CT (CBCT) data with the 2D radiographs and leveraging generative adversarial networks (GANs) and recurrent neural networks (RNNs) were explored to improve sensitivity and robustness. All CNN models achieved high accuracy (exceeding 90%), with an optimized ResNet-50 model performing best at 94.6% accuracy (Cohen’s κ = 0.935), indicating excellent agreement with expert CVM stage assessments. This CNN-based approach closely matched expert scoring and outperformed other tested architectures. Furthermore, fusion of 3D CBCT data with 2D images yielded improved sensitivity, and incorporating GAN- and RNN-based techniques enhanced the robustness of the CVM stage predictions. he optimized CNN model enables accurate, reliable CVM staging and shows strong potential as an AI-assisted tool for orthodontic timing, with enhanced performance through 3D image integration and advanced neural networks.