UCLA

Since its advent in 1931, analysis of lateral cephalometric radiographs has been an important aspect of orthodontic diagnosis and treatment planning. This type of orthodontic analysis provides a quantitative evaluation of the positions of various anatomical structures of the face, including skeletal, dental, and soft tissue features. Cephalometric analysis, however, is a time-consuming process and is not always performed in clinical practice. Cephalometric analysis from Cone-Beam Computed Tomography (CBCT) imaging is particularly cumbersome, requiring both reconstruction of a two-dimensional lateral cephalogram as well as subsequent computer-assisted tracing analysis using a program such as Dolphin. As the use of CBCT continues to become more prominent in orthodontics, methodologies to expedite cephalometric analysis may be valuable in facilitating their use by clinicians. Several different commercially available (e.g. CephX) and proprietary products have been created to analyze cephalometric radiographs using automated artificial intelligence (AI), but the accuracy of these products remains incompletely established. Our study examined the accuracy and workflow impacts of CephX’s ABO analysis directly from CBCT compared to manual reconstruction and tracing by human examiners on 40 CBCT volumes. We hypothesized that automatically generated cephalometric analyses from CBCT volumes would differ significantly from those generated by human examiners, and that workflow time would be statistically significantly decreased in the automated analysis. In contrast, overall workflow time was found to be significantly greater for the CephX software than for the human examiners. Furthermore, our study showed that seven of the eleven measurements in the ABO analysis differed statistically significantly from the human examiners: SNB, ANB, SN-MP, L1-NB, L1-MP, LL to E-Plane, and UL to E-Plane. Nine of the eleven measurements showed average measurement error within clinically acceptable limits, while two showed greater average error than is clinically acceptable – U1-SN and L1-MP were the least accurate measurement in our study. Bland-Altman plots were constructed showing that FMA may show a slight tendency towards greater accuracy at high values, while LL to E-Plane may display a slight tendency to underestimate more frequently at low values and overestimate more frequently at high values. Only one measurement (ANB) showed limits of agreement within the maximum allowed difference; ten out of the eleven variables exceeded the maximum allowed difference. Overall, CephX may offer clinically acceptable performance for most variables in the ABO analysis, but needs further improvement overall, particularly with the inclination of the upper and lower incisors (U1-SN and L1-MP).