Deep Learning-Driven Cephalometric Landmark Validation In Orthognathic Surgery

This study aimed to evaluate the accuracy of artificial intelligence (AI)-assisted automatic 3D cephalometric analysis in orthognathic surgery. Ten patients who underwent bimaxillary orthognathic surgery were enrolled. 3D images from craniofacial cone-beam computed tomography (CBCT) pre- and post-surgery were measured using both an AI-driven automated system (CMF-Net) and a software-assisted manual method (Mimics 19.0, Materialise). Three experienced specialists performed the manual analysis of 24 predefined anatomical landmarks. The mean radial error (MRE) was calculated to assess the accuracy of automatic landmark localization. A paired t-test was used to compare the differences between automatic and manual analysis. The intraclass correlation coefficient (ICC) was used to assess the agreement between the two methods. Asymmetric rate(Q) was used to evaluate the hard tissue symmetry. A Paired t-test was used to analysis the asymmetric rate changes between pre- and post-surgery. The average MRE of the automatic analysis was within the clinically acceptable accuracy threshold (<2mm). The ICC test between the automatic and manual analyses revealed good agreement for all measurements. Both analyses showed improved chin deviation and asymmetry in mandibular body length post-surgery. Additionally, the paired t-test of asymmetry rates revealed a statistically significant difference in two linear measurements (P < 0.05). These findings demonstrate that the automatic analysis of 3D cephalometric measurements shows great accuracy in hard tissue evaluation in orthognathic surgery. The performance of the CMF-net indicates its potential for use in virtual planning in orthognathic surgery.