Growth stage detection using machine learning methods from cephalometric radiographs

Growth and development stages can be determined through the cervical vertebra stages defined on cephalometric radiographic images. Traditionally, accurately identifying the Cervical Vertebra Maturation (CVM) stage can be time-consuming and requires expertise. Today, artificial intelligence systems, utilizing various algorithms, have the capability to perform various tasks such as prediction and classification across multiple fields. These systems can make diagnoses with high accuracy, thus facilitating the work of specialists while minimizing the error rate.The aim of this thesis is to analyze the cervical vertebrae observed in cephalometric radiographs and to detect the stages of CVM using machine learning algorithms. The dataset used in this study was obtained from the records of patients who applied to the Department of Orthodontics, Faculty of Dentistry, Selçuk University. The dataset, created using these records, consists of cephalometric radiographs from 120 patients aged between 7 and 17 years, corresponding to the three different stages of CVM. The key points for determining the stages were suggested by an expert dentist, and 18 features were automatically extracted from the individuals' cephalometric radiographs (a total of 120×18 samples). In the first phase, a classification process was carried out using the dataset with 18 features, applying 5-fold cross-validation and 7 different machine learning algorithms. The best accuracy obtained from this classification was 82.5%. Then, 30 samples were separated from the dataset, and the remaining samples were augmented using the Variational Autoencoder (VAE) method. Afterward, numerical data standardization was performed, and 7 significant features were selected using the Lasso method. The same cross-validation technique was applied to the new dataset, and the training parameters of the algorithms were recorded. The previously separated test dataset was presented to the trained models, and the final performance metrics were noted. As a result, the XGBoost algorithm achieved the best results across all metrics, reaching an accuracy of 90%. Thus, a unique and successful perspective was contributed to the literature.