Evaluation of the clinical utility of lateral cephalometry reconstructed from computed tomography extracted by artificial intelligence.

Abstract

This study assessed the accuracy and reliability of artificial intelligence (AI)-reconstructed images of two-dimensional (2D) lateral cephalometric analyses of facial computed tomography (CT) images, which is widely used for the diagnosis of craniofacial deformities and in the planning of their treatment. Facial CT datasets from 40 patients were collected. Original 1 mm slices were reformatted to 3 mm, and then an AI algorithm reconstructed the 3 mm slices and converted them back to 1 mm to generate lateral cephalometric images. Three observers traced the craniofacial landmarks manually and with autotracing. Landmark discrepancies were quantified between the original 1 mm CT slice and the AI-reformatted 1 mm CT slice, as well as between the original 1 mm CT slice and the 3 mm CT slice. Landmark discrepancies were then compared and reliability tests conducted. AI-reconstructed 1 mm CT images showed significantly fewer discrepancies than the 3 mm CT images, particularly in critical landmarks such as the sella, basion, nasion, orbitale, and pogonion (p < 0.05). The greatest discrepancies in the 3 mm images were observed in the sella and basion. Interobserver and intraobserver reliability analyses showed high consistency (Cronbach's α > 0.7 in nearly all cases). These results support the use of AI-reconstructed images for more accurate diagnosis and treatment planning in craniofacial surgery, potentially reducing the need for additional imaging.