Evaluation of Root Angulations Through Panoramic Films Using Artificial Intelligence

Background/Objectives: Accurate evaluation of root angulation is essential for assessing root parallelism and orthodontic treatment outcomes. In routine clinical practice, this assessment is often performed by visual inspection of panoramic radiographs, which is subjective and prone to observer variability. The objective of this study was to develop and validate an artificial intelligence (AI)–based algorithm for automated, quantitative assessment of mesiodistal root angulations on panoramic radiographs and to evaluate its accuracy relative to conventional manual measurements. Methods: A total of 214 panoramic radiographs (orthopantomograms), comprising 4280 posterior teeth, were retrospectively selected after applying strict inclusion and exclusion criteria. Individual teeth were automatically segmented using a U2-Net–based deep learning architecture. Tooth long-axis orientation was calculated using principal component analysis, with exclusion of the apical third to minimize the influence of root curvature. Angular deviation was measured relative to fixed horizontal reference lines. Manual measurements performed by experienced examiners using 3D Slicer software served as the reference standard. Intra- and inter-examiner reliability, agreement between AI-based and manual measurements, intraclass correlation coefficients (ICC), and Bland–Altman analyses were calculated. Results: Manual measurements demonstrated excellent reliability, with intra-examiner and inter-examiner ICC values of 0.972 and 0.963, respectively. Agreement between the AI-based algorithm and manual measurements was also excellent (ICC = 0.941). Bland–Altman analysis showed a mean difference of −0.10°, with 95% limits of agreement ranging from −1.60° to 1.41°, indicating minimal bias and no proportional error. Conclusions: The proposed AI-based algorithm provides accurate, objective, and reproducible measurements of posterior tooth root angulations on panoramic radiographs. This approach may support clinical decision-making, reduce observer-related variability, and facilitate efficient assessment of root parallelism in orthodontic practice.