Artificial Intelligence for Automated Cephalometric Analysis in ۲D and ۳D Craniofacial Imaging

Aim: Cephalometry is a fundamental tool for orthodontic diagnosis and treatment planning. Manual landmark annotation is labor-intensive, prone to errors, and exhibits significant intra- and inter-examiner variability. Artificial intelligence (AI) offers a promising solution by providing automated landmark identification at clinically acceptable accuracy levels. This study evaluates the accuracy and reliability of an AI-based system for cephalometric landmark annotation using both lateral cephalograms and cone-beam computed tomography (CBCT) scans.Methods: A convolutional neural network (CNN) model, based on EfficientNet-B۰ pretrained on ImageNet, was developed for end-to-end regression of two-dimensional (۲D) coordinates in lateral cephalograms, while a grid-search optimized Automated Landmark Identification system for CBCT (ALICBCT) was employed for three-dimensional (۳D) landmark localization. The dataset included ۴۰۵ lateral cephalograms and ۵۲ CBCT scans from multiple diagnostic centers across India to enhance heterogeneity and generalizability. Twenty-three landmarks were annotated in lateral cephalograms and ۱۸ in CBCTs. Intra- and inter-examiner reliability was assessed using kappa statistics, and comparative evaluation between modalities was performed using paired t-tests. Accuracy was quantified using mean radial error (MRE) and success detection rate (SDR) at clinically relevant thresholds. P-value<۰.۰۵ was considered as significant.Results: For lateral cephalograms, the AI system achieved an average MRE of ۱.۲۱ ± ۱.۰۰ mm and an SDR of ۸۸.۲% at a ۲-mm threshold. In CBCT scans, the model achieved an average MRE of ۰.۹۵ ± ۰.۷۵ mm and an SDR of ۹۴.۶%, demonstrating superior precision of ۳D imaging. Heatmaps and qualitative assessment confirmed high landmark consistency, particularly for dental and midline skeletal points. Intra- and inter-examiner agreements were substantial (κ = ۰.۷۸ and ۰.۷۲, respectively). Comparative analysis indicated that CBCT-based landmark identification outperformed lateral cephalograms for most landmarks, except select midline points.Conclusion: This study demonstrated that the AI-based system achieved clinically acceptable accuracy and reliability for ۲D and ۳D cephalometric landmark detection, with CBCT outperforming lateral cephalograms. These findings confirm the feasibility of automated landmark identification in orthodontic diagnostics.

کلیدواژه ها:

artificial intelligence ، Automated Landmark Identification ، Cone-beam Computed Tomography ، Lateral Cephalogram