Populational influence on cephalometric landmark identification: performance of two AI-driven software programs in...

To assess the performance of cephalometric landmark identification performed by two AI-driven software programs in images from different populations (Brazilian and Korean).

Sixty lateral cephalometric radiographs (30 Brazilian and 30 Korean) were analyzed. The Brazilian images were acquired using the Orthophos XG 5/Ceph device, while the Korean images were obtained from the International Symposium on Biomedical Imaging 2015 database. Images of patients with permanent dentition were included, excluding those with poor head positioning or severe craniofacial deformities. Twenty cephalometric landmarks were identified by two examiners used as the reference standard. Two AI-driven software programs, CefBot™ (Brazil) and WebCeph™ (Korea), automatically identified the same landmarks. Coordinate values for each landmark were measured using ImageJ, and the data were analyzed with Analysis of Variance and Dunnett's post-hoc test.

The Brazilian software showed high accuracy in identifying landmarks on Brazilian images (90%) but was less precise on Korean images (80%), with significant discrepancies in the Glabella, Menton L, Basion, and Orbitale landmarks. Similarly, the Korean software had a higher accuracy in its own population (95%) than in another population (85%), with notable inaccuracies in the Menton L, Basion, and Porion landmarks.

Discrepancies in the identification of specific landmarks, such as Glabella and Menton L, suggest that the accuracy of the software may be influenced by the training process itself and by the population origin of the training data.