Accuracy of automated analysis in cephalometry.

Background/purpose: Artificial intelligence (AI) has been widely used in medicine, including orthodontics. The aim of this study was to investigate the training process of a cascaded Convolutional Neural Network (CNN), built for landmark detection on various qualities of lateral cephalograms and to determine the speed, reliability and clinical accuracy of an algorithm for orthodontic diagnosis. Materials and methods: The CNN model was trained on a total of 1600 lateral cephalograms. After each training datasets (input of 400, 800, 1200, 1600 images) were added, the model was evaluated on a test set containing 78 images of varying quality. We measured the accuracy of AI-based landmark detection by statistical analysis of intra- and interexaminer distance errors, as well as examiner versus model predictions, furthermore by prognosis of consecutive diagnostic failures. Results: There was a clear improvement in time efficiency (5.25 min), and substantial improvements were observed during the training process. In terms of accuracy, based on Euclidean distance error measurements, the best model provided more consistent dot tracing than two different examiners or the same examiner on two different occasions. Angular (0.05°-1.86°) and proportional (3.14%) errors, measured by the best model, were considered clinically acceptable. Conclusion: The application of a proper AI-algorithm for orthodontic cephalometric analysis results in lower variability between models than the variability observed among experts. AI predictions supported the examiners in finding the correct location of the specific landmarks more accurately and in less time as the training of the automatic prediction model improved. Further research could investigate the therapeutic consequences.