Abstract
BACKGROUND: Bifid mandibular canals (BMCs) represent a clinically significant anatomical variation that increases the risk of iatrogenic inferior alveolar nerve injury during oral and maxillofacial procedures. Accurate identification of BMCs on cone-beam computed tomography (CBCT) remains challenging due to their subtle morphology and reliance on observer expertise. This study aimed to evaluate the diagnostic performance of a customised deep learning–based three-dimensional (3D) U-Net model for the automated delineation and detection of BMCs on CBCT volumes. METHODS: A total of 208 CBCT volumes acquired using the Planmeca ProMax 3D MID system (Planmeca Oy, Helsinki, Finland) were retrospectively selected. The dataset was stratified into a training set, intentionally enriched with anatomical variants, and an independent test set. Ground truth annotations were established through manual voxel-wise segmentation by experienced oral and maxillofacial radiologists. A customised 3D U-Net architecture was utilised via the MONAI framework with five-fold cross-validation. Model performance was evaluated on the independent test set using voxel-wise metrics, including the Dice Similarity Coefficient (DSC) and Intersection-over-Union (IoU), as well as case-level detection metrics (sensitivity and specificity). RESULTS: The model achieved mean DSC values of 0.91 for standard mandibular canals and 0.77 for bifid variants. On the independent test set, enriched with complex morphological subtypes, the model demonstrated 100% sensitivity, 81.8% specificity, and 90.5% overall diagnostic accuracy for detecting BMCs. CONCLUSIONS: This study demonstrates that a customised 3D U-Net framework can accurately identify BMCs on CBCT volumes within a single-centre setting. When used as a decision-support tool, the proposed approach has the potential to assist radiologists in recognising anatomically complex canal variants. Further multicentre validation is required to confirm broader clinical applicability.