Generative Artificial Intelligence for Orthognathic Planning: Patient-Specific 3-Dimensional Jaw Reference Forms Conditioned on the Cranial Base

BACKGROUND: Conventional cephalometry relies on measurements derived from a limited set of landmarks and population-based norms, which may not fully capture individual skeletal form. Although evaluated independently, many cephalometric variables reflect shared anatomical relationships that influence overall facial form. These constraints motivate patient-specific, anatomically complete reference forms for personalized planning.

PURPOSE: The study purpose was to compare surgical plans created using a generative artificial intelligence (AI) model with those created by expert surgeons.

STUDY DESIGN, SETTING, AND SAMPLE: This was a prospective, in-silico (computer-based), paired-equivalence validation study using retrospectively acquired preoperative computed tomography or cone-beam computed tomography scans. The sample included virtual models of subjects who underwent orthognathic surgery.

PREDICTOR VARIABLE: The predictor variable was the planning method, grouped into 2 categories: AI-guided template-based planning and surgeon-derived planning using conventional cephalometry supplemented by clinical expertise.

MAIN OUTCOME VARIABLE: The main outcome variable was in-silico agreement, defined as surgical plan similarity between AI-guided and surgeon-generated plans, quantified using the dice similarity coefficient (DSC), which ranges from 0 to 1, with higher values indicating greater agreement between plans.

COVARIATES: Age and sex were documented but not included in the equivalence testing.

ANALYSES: Each case was planned once with the AI-guided template and once by each of 3 expert surgeons independently. Similarity between plans was measured using DSC. Statistical equivalence between plans was evaluated using the 2 one-sided tests at a 0.05 significance level, and distributional similarity was assessed using the Kolmogorov-Smirnov test.

RESULTS: The sample was composed of 30 subjects (20 [67%] female, 10 [33%] males) with a median age of 23.5 years (interquartile range, 18 to 32). The mean intersurgeon DSC was 0.78 ± 0.11 (95% CI, 0.75 to 0.80), and the mean AI-guided-surgeon DSC was 0.77 ± 0.13 (95% CI, 0.74 to 0.80). The two one-sided test procedure confirmed statistical equivalence within a ±0.10 DSC margin (P < .001). The Kolmogorov-Smirnov test showed no significant difference in distributions (P = .6), indicating that the AI-generated plans fell within natural expert variation.

CONCLUSIONS: AI-guided surgical plans demonstrated statistical equivalence in overall plan similarity compared with surgeon-generated plans, indicating that the proposed method produces results within the range of expert variability.