(42) Comparing Lateral Cephalometric Tracing Between Artificial Intelligence and Human: Agreement, Reliability, and Systematic Differences for Skeletal Deformities

Abstract: Lateral cephalometric tracing, introduced by Broadbent and Hofrath, remains the primary method for diagnosing and planning treatment of dental malocclusions and skeletal discrepancies.¹ Manual tracing requires accurate identification of anatomical landmarks with millimeter precision, which often leads to a time-consuming and error-prone process.² With the development of Artificial Intelligence (AI) programs for cephalometry, their reliability and accuracy have come into question when compared to manual tracing. Our study aims to assess the agreement, reliability, and systematic differences between AI and humans lateral cephalometric tracings across skeletal classifications.
Five lateral cephalograms per skeletal deformity – Asymmetry, ClassII, ClassIII, Open-bite, and Pseudo-ClassIII – were analyzed by CephX (AI-powered software) and 9 oral and maxillofacial surgery senior residents at Thomas Jefferson University Hospital (Figure 1). The residents were provided training on lateral cephalometric tracing with Steiner+Wits analysis using Dolphin Imaging & Management Solutions®11.95(Chatsworth, CA,USA).
The primary predictor variable was the tracing method – AI versus residents. The outcome variables consisted of 14 cephalometric measurements (SNA, SNB, ANB, occlusal plane to SN, interincisal angle, MP-SN, Pog-NB, Wits appraisal, U1-NA, L1-NB, upper lip to E-plane, lower lip to E-plane), analyzed by skeletal classification. The secondary outcome variables were age, gender, and comorbidities.
Descriptive statistics were calculated. Normality was assessed using the Shapiro-Wilk test. If normally distributed, one-sample t-test was used to compare resident tracings to AI reference values. AI served as the reference due to CephX’s deterministic nature, producing consistent, identical outputs. Wilcoxon signed-rank test with normal approximation was used for non-normal variables. To assess the extent of agreement and systematic differences, mean/median differences, and 95% confidence intervals were calculated (α = .05).
AI tracings demonstrated statistically significant differences consistently across all skeletal deformities in the following:
Interincisal angle: AI consistently over-projected angulations, with significantly higher mean values compared to senior residents in all skeletal classifications (P < .05), especially in open-bite patients (108.4° [AI mean], 94.63° ± 5.7 [human mean]).
Soft tissue Measurements: AI over-projected both the upper and lower lips to the E-plane across all deformities (P < .05). The largest upper lip discrepancy was observed in pseudo-class III (mean difference = 8.0 mm), while Class III showed the greatest lower lip deviation (mean difference = 3.3 mm).
Skeletal angular measurements: AI systematically underestimated SNB and overestimated MP-SN (P < .05). The greatest discrepancies were seen in pseudo-class III, where AI under-projected SNB by 3.1° (77.49°, 80.59° ± 1.04) and over-projected MP-SN by 12.27° (39.45°, 27.18° ± 1.84). In asymmetry, AI overestimated MP-SN by 11.67°.
Across all skeletal classifications, pseudo-Class III exhibited the greatest discrepancy (92.9%), with 13 out of 14 cephalometric variables significantly differing between AI and human tracings. Open-bite and Class II followed with 78.6% discrepancy, while asymmetry and Class III had 64.2%.
The results suggest that AI tracings differed significantly from human tracings, particularly in interincisal angle, upper and lower lip to E-plane, SNB, and MP-SN. The discrepancy rates range across all skeletal classifications from 64.2% to 92.9%. Greater deviations were observed in more complex skeletal deformities – open-bite and pseudo-Class III – suggesting AI lacks reliability for precise analysis. Therefore, while AI offers efficiency and standardization, AI is not yet a viable replacement for manual cephalometric tracing especially in complex cases. Manual tracing remains the more reliable standard in cases requiring precise, individualized cephalometric measurements.

References
[1] Broadbent, H. B. (1931). A new X-ray technique and its application to orthodontia. Angle Orthodontist, 1(2), 45–66.
[2] Kolokitha, O.-E. G., & Topouzelis, N. (2011). Cephalometric methods of prediction in orthognathic surgery. Journal of Maxillofacial and Oral Surgery, 10(3), 236-245. https://doi.org/10.1007/s12663-011-0228-7
Figure 1: Lateral cephalometric radiographs for (A) Asymmetry, (B) Class II, (C) Class III, (D) Open-Bite, and (E) Pseudo-Class III analyzed by AI and senior residentsTable 1: One-sample Tests and 95% Confidence Interval of AI Versus Senior Residents For Each Skeletal Classification (*Wilcoxon signed-rank test; Statistical significance is in red.)