AI-Based Cephalometric Landmark Detection via Deep Learning Networks Evaluation

Cephalometric X-rays are essential in orthodontics for assessing skeletal relationships and treatment planning. However, manual tracing is time-consuming and prone to interobserver variability. AI-driven analysis has evolved from early rule-based methods to deep learning models like Convolutional Neural Networks (CNNs), improving accuracy but still facing challenges in standardizing imaging protocols and addressing measurement discrepancies. YOLO-based models, such as DR. ANNIE™, enhance efficiency by detecting multiple landmarks simultaneously, eliminating anchor boxes, and reducing manual calibration. While AI shows promise in increasing precision and reproducibility, dataset diversity, calibration accuracy, and imaging standardization remain critical challenges for clinical adoption. The sample was composed of 80 lateral cephalometric radiographs taken from the UIC-COD Dolphin Imaging Database. Manual tracings formed the Manually Detected Landmark Group (MDL), while AI-generated landmarks created the AI Identified Landmark Group (ADL). Both groups were aligned using Sella (0,0) as the origin and the Frankfort Horizontal (FH) plane to ensure direct spatial comparisons. Statistical analysis assessed differences in X and Y coordinates, revealing: Mean differences: X: 2.2 ± 1.8 mm, Y: 3.0 ± 2.6 mm with an overestimation trend in measurements as well with Linear: 1.4 ± 1.1 mm, Angular: 2.3 ± 1.8°. No significant mean difference in X-coordinates of Orbitale, ST A Point, Menton, L1 Tip, U1 Root, Basion, and PNS. All Y-Axis coordinates were statistically significant. Nine cephalometric measurements showed discrepancies, including Convexity, Lower Lip to E-Plane, Molar Relation, Porion Location, U1-FH, U1-APo, Facial Axis, Mandibular Arc, U1-FH, and U Incisor Inclination SUMMARY (continued) AI-assisted cephalometric analysis demonstrated promising accuracy but tended to overestimate linear and angular measurements. While AI showed consistency in select X-coordinates, Y-axis errors remained significant, and multiple cephalometric measurements exhibited discrepancies. These findings highlight AI's potential in orthodontic diagnostics but emphasize the need for further refinement in landmark detection, calibration, and dataset standardization to enhance clinical reliability.