Cephalometric analysis is the primary diagnosis method in orthodontics. In our previous study, the algorithm was developed to estimate cephalometric landmarks from lateral facial photographs of patients with normal occlusion. This study evaluates the estimation accuracy by the algorithm trained on a dataset of 2320 patients with added malocclusion patients and the analysis values. The landmarks were estimated from the input of lateral facial photographs as training data using trained CNN-based algorithms. The success detection rate (SDR) was calculated based on the mean radial error (MRE) of the distance between the estimated and actual coordinates. Furthermore, the estimated landmarks were used to measure angles and distances as a cephalometric analysis. In the skeletal Class II malocclusion, MRE was 0.42 ± 0.15 mm, and in the skeletal Class III malocclusion, MRE was 0.46 ± 0.16 mm. We conducted a cephalometric analysis using the estimated landmarks and examined the differences with actual data. In both groups, no significant differences were observed for any of the data. Our new algorithm for estimating the landmarks from lateral facial photographs of malocclusion patients resulted in an error of less than 0.5 mm; the error in cephalometric analysis was less than 0.5°. Therefore, the algorithm can be clinically valuable.
Keywords: CNN-based algorithm; Cephalometric analysis; Cephalometric landmarks; Estimation cephalometric landmarks; Lateral facial photograph; Malocclusion data.
© 2024. The Author(s).