Abstract
BACKGROUND: Estimating craniofacial patterns is essential for successful orthodontic treatment. However, conventional static measurements are inadequate for capturing dynamic changes, and manual cephalometric analysis is labor-intensive and requires specialized expertise. In this study, we propose Skel-Net, a novel anatomic prior-guided deep learning network designed to estimate ANB angle changes over five years in children and adolescents aged 8-16. METHODS: In a two-stage approach, Skel-Net combines cephalometric landmark detection via Ceph-Net and multichannel inputs, including two-dimensional heatmaps and ANB priors, to enhance prediction accuracy. A dataset of 612 lateral cephalograms from 245 patients was used to train and validate the model, and its performance was compared against DenseNet121, MobileNetV2, ResNet101, and VGG16. RESULTS: Skel-Net outperformed the other models with the lowest prediction errors (mean absolute error: 1.021 degrees; root mean squared error: 1.338 degrees) and the highest R(2) value (0.517), demonstrating robust predictive capabilities. CONCLUSIONS: By leveraging anatomic priors and longitudinal data, Skel-Net enables dynamic and personalized predictions of craniofacial growth. This framework will facilitate early and precise orthodontic interventions, enhancing treatment efficiency, stability, and overall patient outcomes.