Objective: The aim of this study was to evaluate the accuracy of a multi-stage convolutional neural network (CNN) model-based automated identification of posteroanterior (PA) cephalometric landmarks Methods: The multi-stage CNN model was implemented with a personal computer. A total of 430 synthesized PA cephalograms from conebeam computed tomography (CBCT-PA) were selected as samples. Twenty-three landmarks were manually identified on the all CBCT-PA images by a single examiner. Intra examiner’s reproducibility was confirmed by two times identification on the 85 randomly selected CBCT-PA images which also set as test data with two-week interval before model training. We input the 345 data out of 430 CBCT-PA images to the multi-stage CNN model at the initial learning stages. The multi-stage CNN model was tested with 85 images. The first manual identification on 85 images was set as a truth ground. The mean radial error (MRE) and successful detection rate (SDR) were calculated to evaluate the errors between manual identification and AI prediction. Results: The accuracy of AI showed MRE = 2.23 mm ± 2.02 mm and range of 2 mm below SDR achieved 60.88%. For comparison of the repetitive task, AI predicted landmarks on the same position, and MRE in two times manual identification showed 1.31 ± 0.94 mm. Conclusion: The prediction for CBCT synthesized PA cephalometric landmarks with automated identification has not sufficiently achieved less than 2 mm error range of clinically favorable level. Consistency of AI landmark identification on PA cephalograms was better than manual identification

Keywords: CT, Artificial intelligence, Convolutional Neural Networks, Diagnosis and Treatment planning, Facial asymmetry, Automatic identification, PA Cephalometrics, Cone-Beam Computed Tomography