Clinical Reliability of AI-Based Cephalometric Analysis Using WebCeph: A Comparative Agreement Study.
Authors
Affiliations (2)
Affiliations (2)
- Dentistry Department, Medicine and Health Science Faculty, Catholic University of Valencia, 46001 Valencia, Spain.
- Stomatology Department, University of Valencia, 46010 Valencia, Spain.
Abstract
<b>Background/Objectives</b>: Artificial intelligence has accelerated cephalometric analysis by enabling rapid and standardized measurements. However, whether these automated outputs can be considered clinically interchangeable with expert manual tracing remains unresolved, particularly for routinely used analyses such as Steiner. <b>Methods</b>: A comparative experimental study was conducted on 100 lateral cephalometric radiographs analysed using two parallel approaches: expert manual tracing and fully automated analysis with the WebCeph platform. Seven Steiner variables (SNA, SNB, ANB, 1-NA, 1-NB, interincisal angle, and FMA) were evaluated. Paired <i>t</i>-tests were used to assess differences between methods, while agreement was evaluated using intraclass correlation coefficients and Bland-Altman analysis. Particularly low agreement was observed for clinically relevant parameters such as ANB and FMA. <b>Results</b>: Six of the seven variables showed statistically significant differences between methods. Automated measurements systematically tended to overestimate both skeletal and dental parameters. Agreement was inconsistent and frequently poor, with ICC values ranging from 0.01 to 0.60 for clinically relevant variables such as ANB and FMA. Importantly, small or non-significant mean differences did not translate into acceptable agreement. Bland-Altman analysis confirmed the presence of systematic bias and wide limits of agreement, especially for dental measurements. <b>Conclusions</b>: Despite its speed and automation, WebCeph does not achieve clinically acceptable agreement with expert manual tracing across several key cephalometric variables. The observed discrepancies-particularly in parameters critical for sagittal and vertical diagnosis-may compromise clinical interpretation and treatment planning. These findings support the use of AI-based cephalometric analysis as an adjunctive tool rather than a substitute for clinician-guided evaluation.