Comparison of the performance of deep learning-based artificial intelligence models and manual methods in lateral cephalometric analyses

Aim: Artificial intelligence (AI) enhances cephalometric diagnostic procedures in orthodontics and improves the efficiency of patient evaluations. The aim of this study was to compare and evaluate the cephalometric measurement results obtained from the Dolphin Imaging software, which performs computer-assisted digital cephalometric analysis, and the CranioCatch software, which performs fully automatic cephalometric analysis supported by artificial intelligence. Methods: This study included lateral cephalometric radiographs generated from cone-beam computed tomography (CBCT) records and digital lateral cephalometric radiographs of 571 patients. Cephalometric analyses were performed using a total of 36 measurements, comprising 17 angular, 18 linear, and 1 ratio parameter. To analyze the differences between the two methods, Independent Sample t-Test and Mann–Whitney U tests were applied, and Bland–Altman plots were constructed. The level of statistical significance was set at p < 0.05. Results: Following the comparative analysis of the obtained cephalometric measurement values, statistically significant differences were identified in the following parameters: FMA, SN/GoGn angle, gonial angle, articular angle, sum of posterior angles, facial axis angle, U1–SN angle, interincisal angle, maxillary length, maxilla–mandible difference, corpus length, N–Me, N–ANS, S–Go, S–Ar, ramus height, U1 protrusion, and Jarabak ratio. Furthermore, assessment of the Bland–Altman plots demonstrated that the FMA, SN/GoGn angle, gonial angle, interincisal angle, and corpus length parameters exhibited clinically significant discrepancies between the compared methods. Conclusions: Deep learning-based fully automatic cephalometric analyses can provide clinically acceptable diagnostic accuracy. However, it is considered that additional manual adjustments are required to achieve higher accuracy and more reliable results. Keywords: Cephalometric analysis, Deep learning, Artificial intelligence