Left-Right Consistency and Quantification of Craniofacial Asymmetry by Two Fully Automatic AI-Powered...

Background/Objectives: Automatic three-dimensional (3D) cephalometry on cone beam computed tomography (CBCT) is increasingly used in orthodontic and orthognathic practice, but whether fully automatic systems reproduce bilateral measurements with equal fidelity on both sides, and whether the right-left difference they report can be trusted as a clinical measure of asymmetry, has not been studied systematically. This study aimed to quantify side-specific agreement and asymmetry quantification fidelity of two commercial AI cephalometry systems against calibrated manual 3D tracing. Methods: Thirty-one CBCT scans were analysed by calibrated manual tracing (InVivo software), an integrated AI module (AI InVivo), and a cloud-based system (Diagnocat). Side-specific agreement was quantified using the two-way random effects intraclass correlation coefficient [ICC(2,1)], Bland-Altman analysis, and two one-sided equivalence tests (±2°/±2 mm). Asymmetry fidelity was assessed by ICC of the automatic right-left difference (Δ = R - L) against the manual Δ, Pearson correlation, and sensitivity/specificity for detecting clinically relevant asymmetry. Results: Per-side agreement closely mirrored overall accuracy; any systematic error was applied almost equally to both sides (bias difference ≤0.3° or ≤0.9 mm between sides). The mandibular plane angle achieved excellent ICC (0.95-0.97) and equivalence on both sides for both systems. The gonial angle was equivalent for AI InVivo (ICC 0.92-0.93) but not for Diagnocat (ICC 0.40; bias -6.95° and -6.97° on the right and left, respectively). Asymmetry agreement was generally weaker than per-side accuracy. Diagnocat showed no correlation with manual gonial asymmetry (ICC -0.12), while AI InVivo demonstrated poor agreement for body length asymmetry (ICC 0.18). Both systems under-detected clinically relevant gonial asymmetry (sensitivity 0.00-0.17), with errors predominantly in the false-negative direction. Conclusions: Fully automatic AI-powered 3D cephalometry applies systematic errors symmetrically to both sides, so per-side accuracy holds equally for right and left measurements. However, asymmetry fidelity is a separate, more demanding property that does not follow from absolute accuracy and remains unreliable for landmark-sensitive parameters such as the gonial angle. Clinician verification is required before automatic analysis is used for asymmetry assessment.