Clinical accuracy of cephalometric analysis using deep learning–based automated landmark identification on CBCT in class I and class II malocclusions

This study aimed to perform three-dimensional (3D) cephalometric analysis based on automatically identified landmarks, to evaluate their clinical accuracy, and to investigate the relationship between algorithmic precision, measured by mean radial error (MRE), and clinical validity. Retrospective cone-beam computed tomography (CBCT) scans representing diverse dentition stages and malocclusion types were used to develop an automated landmark identification model incorporating an optimized U-Net architecture with an Efficient Global Attention module. Seventy-one 3D cephalometric measurements derived from manually annotated landmarks and AI-generated landmarks were compared across 75 randomly selected CBCT scans of Class I/II malocclusion patients. Statistical analyses included paired t-tests/wilcoxon signed-rank test, intraclass correlation coefficients (ICC), and Bland–Altman analysis. Analysis revealed that 9 out of the 71 measurements (12.68%) showed statistically significant differences; however, all mean differences between AI-derived and ground truth measurements were clinically negligible (≤ 1 mm/°). ICC analysis demonstrated excellent agreement overall, with only two parameters (PP–HF–MSP and Me–MSP; 2.82%) showing ICC values below 0.90. Bland–Altman analysis indicated that 59.15% of AI-based cephalometric measurements achieved clinical interchangeability with ground truth, defined by limits of agreement within ± 2.0 mm/°. Among 36 linear measurements, all 26 parameters associated with landmarks exhibiting an MRE below 2 mm fell within clinically acceptable limits, whereas angle-based measurements did not demonstrate a clear correlation with MRE. The precision of automated 3D cephalometry is contingent upon the magnitude and directional vector of landmarking error. Angular measurements are particularly susceptible to unconstrained directional errors. Consequently, the MRE metric alone is insufficient to comprehensively evaluate the accuracy of automated cephalometric analysis, particularly in regions lacking definitive anatomical contours. Clinically applicable automated 3D cephalometry may therefore benefit from minor manual refinement at specific landmarks, such as the gonion and incisor root apex, particularly in patients with mixed dentition.