Purpose:
To evaluate the accuracy and reliability of a newly developed artificial intelligence (AI)-driven system in identifying cephalometric landmarks, comparing its performance to manual tracing by orthodontists.

Material and methods:
A total of 506 lateral cephalometric radiographs of patients aged 14 to 40 years were collected from Dubai Dental Hospital, United Arab Emirates. After applying exclusion criteria, 473 radiographs were included in the final analysis. Manual tracings were performed using Dolphin Imaging software, identifying key landmarks including Sella, Nasion, A-point, B-point, Menton, Glabella, and others. The AI-based cephalometric system (AI Ceph) was trained on these radiographs to locate the same landmarks. Its performance was compared to manual tracings to assess accuracy and reliability.

Results:
The AI Ceph system demonstrated high concordance with manual tracing. The Menton point showed the highest agreement (96.6%), while Nasion had the lowest (86.6%). The majority of landmark deviations were within 2 mm, the predefined clinically acceptable range, while a small proportion exceeded this threshold and were classified as discordant (> 2 mm) according to the prespecified binary criterion. Sensitivity was above 80% for all landmarks, except Nasion (72.7%) and A-point (78.8%). A statistically significant improvement in concordance after model retraining was observed only for the Glabella landmark (p < 0.0213).

Conclusions:
The AI system provided accurate and reliable cephalometric landmark identification. While most deviations fell within the clinically acceptable 2-mm tolerance, landmarks exceeding this threshold were clearly identified and quantified using the binary discordance classification (> 2 mm), ensuring transparent reporting of all errors.