This study identified the most accurate model for predicting longitudinal craniofacial growth in a Japanese population using statistical methods and machine learning. Longitudinal lateral cephalometric radiographs were collected from 59 children (27 boys and 32 girls) with no history of orthodontic treatment. Multiple regression analysis, least absolute shrinkage and selection operator, radial basis function network, multilayer perceptron, and gradient-boosted decision tree were used. The independent variables included 26 coordinated values of skeletal landmarks, 13 linear skeletal parameters, and 17 angular skeletal parameters in children ages 6 to 12 years. The dependent variables were the values of the 26 coordinated skeletal landmarks, 13 skeletal linear parameters, and 17 skeletal angular parameters at 13 years of age. The difference between the predicted and actual measured values was calculated using the root-mean-square error. The prediction model for craniofacial growth using the least absolute shrinkage and selection operator had the smallest average error for all values of skeletal landmarks, linear parameters, and angular parameters. The highest prediction accuracies when predicting skeletal linear and angular parameters for 13-year-olds were 97.87% and 94.45%, respectively. This model incorporates several independent variables and is useful for future orthodontic treatment because it can predict individual growth.

Background/Objectives: Accurate landmark identification underpins reliable cephalometric analysis. This study evaluated a ResNet50-based, single-stage regression convolutional neural network for direct automatic localization of 15 landmarks on lateral cephalograms. Methods: This retrospective study included 669 lateral cephalograms (669 patients): 619 for training and 50 randomly selected for an internal holdout test set. One of five orthodontists annotated each cephalogram, and coordinates served as the reference standard. One orthodontist independently re-annotated all test images after more than 2 weeks to assess reproducibility. Model performance was evaluated using Euclidean localization errors and success detection rates (SDR). Results: Across 750 landmark predictions, mean localization error was 1.25 ± 1.39 mm (95% confidence interval, 1.15–1.35 mm) and median error was 0.72 mm. SDRs within 1.0, 2.0, and 4.0 mm were 64.5%, 81.5%, and 94.4%, respectively. A statistically significant overall difference was observed among the 15 landmarks (Friedman χ²(14) = 28.84, p = 0.011). Point A had the numerically largest mean error (1.66 mm) and the mandibular central incisor the smallest (1.04 mm). The mean difference between original and repeated annotations was 0.52 ± 0.92 mm. Conclusions: In this single-center internal holdout test set, mean localization error was below the 2.0 mm benchmark. However, 18.5% of predictions exceeded 2.0 mm, and external validity remains unconfirmed. Artificial-intelligence-generated landmarks should be verified by orthodontists, and external validation is required before broader clinical use.