Abstract: To address the problems of low efficiency and insufficient consistency in manual landmark positioning during cephalometric analysis, this study investigates an automatic detection method for key anatomical landmarks in lateral cephalograms. Based on 1035 lateral cephalometric images obtained between 2020 and 2025 from Beijing Stomatological Hospital, Capital Medical University, a high-density annotated dataset containing 116 craniofacial anatomical landmarks is constructed to establish an automatic landmark-detection model, RTMPose. Systematic experiments are performed to evaluate the performance advantages of the model in complex multi-landmark-detection tasks. Additionally, a consistency analysis of clinical cephalometric measurements and multicenter real-world validation experiments are conducted to comprehensively verify the stability and reproducibility of the model under different imaging conditions and data distributions. Experimental results show that the proposed method achieves favorable performance in terms of both localization accuracy and generalizability, thereby providing reliable technical support for the automation of cephalometric analysis and intelligent orthodontic diagnosis.