Efficacy of a Deep Learning System for Automatic Analysis of the Mandibular Canal Type on Panoramic Radiographs

Accurate anatomical variant detection is critical in clinical diagnostics, yet disparities in imaging modalities often challenge reliable assessment. In dentistry, panoramic radiographs (PRs) are widely used for mandibular canal evaluation, but their reported detection rates for bifid variants (0.038%–1.98%) fall far below those of cone-beam computed tomography (CBCT; 10%–66%), highlighting a need for improved diagnostic tools. Here, we address this gap by developing a deep learning-based tri-comparison expertise decision (TED) system to automate mandibular canal variant classification on PRs. Using retrospective data from 442 mandible sides (279 participants, aged 18–32 years), we validated PRs against CBCT ground truth and decomposed multi-class classification into pairwise comparisons with an “Another” class to enhance discrimination of anatomically similar variants. Here we show that the TED system achieved superior diagnostic accuracy (0.701, 95% CI: 0.674–0.728) and AUROC (0.854, 95% CI: 0.824–0.884) compared to assessments by five experienced dentists (highest accuracy: 0.683; AUROC: 0.810), while also revealing strikingly low inter-rater agreement among experts (Fleiss' kappa = 0.046). These results demonstrate that the TED approach not only outperforms manual evaluations but also provides consistent, cost-effective automation of a task prone to human variability. By bridging the performance gap between PRs and CBCT, this tool offers a practical solution for preoperative risk assessment in dental practice. Broader validation across diverse clinical settings could further solidify its role in improving diagnostic workflows and patient outcomes.