AI-based 3D Measurement of Root Canal Curvature from CBCT: Validation of an Automated Schneider Angle Analysis

Conventional two-dimensional curvature assessment methods are insufficient to capture the complex three-dimensional (3D) geometry of root canals. This study aimed to develop and validate a deep learning-based pipeline for automated 3D Schneider angle measurement from cone-beam computed tomography images of mandibular molar mesial roots and to evaluate its agreement with expert manual measurements.

A 3D No New U-Net segmentation model was developed using 331 training/validation cone-beam computed tomography volumes and evaluated on an independent held-out test set of 50 cases comprising 127 canal segments. Automated Schneider angles were compared with reference measurements obtained by 2 experienced endodontists performing blinded, repeated measurements across 2 sessions. Agreement was assessed using intraclass correlation coefficients (ICCs), Bland-Altman analysis, and quadratic weighted kappa. The equivalence margin was set at ±5°.

Overall, artificial intelligence (AI)-manual agreement was good (ICC = 0.890; 95% CI: 0.840-0.920), with a mean bias of +0.88° (limits of agreement: -6.51° to +8.27°). Agreement was highest for common segments in Vertucci Type II configurations (ICC = 0.947). Mesiobuccal canals showed a small but statistically significant positive bias (+1.45°; P < .05). Schneider category agreement was substantial (quadratic weighted κ = 0.717). Mean-difference equivalence was demonstrated within the prespecified ±5° margin for all canal types (all P < .05), although the 95% limits of agreement extended beyond the ±5° margin at the individual-measurement level.

This internal validation study suggests that the AI-assisted workflow can approximate expert 3D Schneider angle measurements at the mean-difference level; however, external validation and stronger individual-level agreement are required before clinical implementation.