Abstract
Background. The accurate timing of growth modification treatments is crucial for optimal results in orthodontics. However, traditional methods for assessing growth status, such as hand-wrist radiographs and subjective interpretation of lateral cephalograms, have limitations. This study aimed to develop a semi-automated approach using machine learning based on cervical vertebral dimensions (CVD) for determining skeletal maturation status.
Methods. A dataset comprising 980 lateral cephalograms was collected from the Department of Orthodontics, Shahid Beheshti Dental School in Tehran, Iran. Eight landmarks representing the corners of the third and fourth cervical vertebrae were selected. A ratio-based approach was employed to compute the values of C3 and C4, accompanied by the implementation of an auto_error_reduction (AER) function to enhance the accuracy of landmark selection. Linear distances and ratios were measured using the dedicated software. A novel data augmentation technique was applied to expand the dataset. Subsequently, a stacking model was developed, trained on the augmented dataset, and evaluated using a separate test set of 196 cephalograms.
Results. The proposed model achieved an accuracy of 99.49% and demonstrated a loss of 0.003 on the test set.
Conclusion. By employing feature engineering, simplified landmark selection, AER function, data augmentation, and eliminating gender and age features, a model was developed for accurate assessment of skeletal maturation for clinical applications.