Evaluating the success of artificial intelligence algorithms in determining the maturation period through hand-wrist radiographs

Aim: The aim of our study is to accurately and consistently determine skeletal maturity using AI-based applications with hand-wrist radiographs and, in parallel, to prevent misdiagnosis and mistreatment. For this purpose, the performance of AI algorithms in determining maturation stages, age, and gender from hand-wrist radiographs retrospectively obtained from patient archives will be evaluated. Material and Methods: The material of the study consists of 6572 hand-wrist radiographs obtained from systemically healthy Turkish individuals who presented for orthodontic treatment. The material was classified based on five different maturation methods and additionally categorized by age and gender. Subsequently, the images were annotated using the CranioCatch labeling software (CranioCatch, Eskişehir, Turkey). The datasets were randomly divided into training, validation, and test sets. The performance of the YOLOv8 deep learning architecture, trained using the PyTorch library, was assessed on test datasets for the detection of growth and developmental stages. Results: Using the deep learning model in this study, the accuracy and F1 scores were as follows: for Group I, accuracy was 0.860 and F1 score was 0.850; for Group II, accuracy was 0.920 and F1 score was 0.960; for Group III, accuracy was 0.878 and F1 score was 0.935; for Group IV, accuracy was 0.981 and F1 score was 0.990; for Group V, both accuracy and F1 score were 0.920; for Group VI, both accuracy and F1 score were 0.980; and for Group VII, accuracy was 0.580 and F1 score was 0.570. The highest F1 score, 0.990, was achieved for Group IV, while the lowest F1 score, 0.570, was observed for Group VII. The balanced distribution of the datasets resulted in AUC values close to 1, demonstrating the successful performance of the deep learning model. Furthermore, these values indicate the model's high accuracy in distinguishing between different classes. Conclusion: The successful identification of growth and developmental stages using a deep learning model has the potential to assist clinicians in reducing workload. It is anticipated that AI will initially play an auxiliary role in diagnostic and treatment decision-making. Integrating AI-based applications into clinical workflows will minimize intra- and inter-observer discrepancies and significantly reduce misdiagnoses and mistreatments. Keyword: Artificial intelligence, deep learning, hand-wrist radiographs, Skeletal maturation