Virtual surgical planning-to-actual surgical outcome discrepancies following artificial intelligence-assisted and manual three-dimensional cephalometric tracing in orthognathic surgical planning

Abstract

This study compared discrepancies between virtual surgical planning (VSP) and actual surgical outcomes (ASO), as well as time efficiency, among manual and AI-assisted three-dimensional (3D) cephalometric tracing workflows in OGS planning. Fifteen patients undergoing bimaxillary orthognathic surgery were retrospectively reviewed. Three preoperative virtual surgical plans were created for each patient: manual tracing in Invivo, manual tracing in ON3D, and AI-assisted tracing in ON3D. The primary comparison was conducted between the manual and AI-assisted ON3D workflows, with Invivo included as a platform-level reference workflow. The primary endpoint was the 3D discrepancy between VSP and ASO at eight skeletal and dental landmarks along the X-, Y-, and Z-axes. Manual and AI-assisted ON3D workflows showed no statistically significant differences in VSP-to-ASO discrepancies for most skeletal landmarks along the X- and Y-axes, with absolute mean differences generally within clinically acceptable limits. Both ON3D workflows showed statistically significant Z-axis discrepancies at mandibular landmarks, including Point B, Pogonion, and Menton; however, several were approximately 1–1.5 mm and were not specific to AI-assisted tracing. Maxillary landmark discrepancies were generally within ± 1 mm across ON3D workflows and remained within clinically acceptable limits. AI-assisted tracing reduced mean preoperative construction time from approximately 10.5 min to 4.2 min and reduced total planning time from approximately 17–18 min to 7–9 min. Inter-examiner reliability for time measurements was high, with ICC values exceeding 0.75. In contrast, inter-examiner reliability for ASO landmark identification was only fair-to-moderate (ICC = 0.47). AI-assisted 3D cephalometric tracing showed no statistically significant differences in VSP-to-ASO discrepancies compared with manual tracing workflows, while improving planning efficiency. However, these findings should be interpreted as downstream workflow-based outcome data rather than direct validation of coordinate-level AI landmark identification accuracy.

The authors' abstract, as published at the source. Scientific Reports, 2026 · DOI ↗