Three-dimensional U-Net-based segmentation of the maxillary incisive canal on cone-beam computed tomography.

INTRODUCTION: The objective of this study was to evaluate a 3-dimensional (3D) U-Net-based framework for the automated segmentation of the maxillary incisive canal on cone-beam computed tomography (CBCT) images. METHODS: This retrospective study included CBCT scans of 200 patients aged 12-62 years. Manual segmentation of the maxillary incisive canal was performed to generate reference annotations. The dataset was randomly divided into training (n = 140) and testing (n = 60) sets. A 3D U-Net architecture was trained on the training set and evaluated on the testing set using the Dice similarity coefficient, accuracy, precision, recall, 95% Hausdorff distance, segmentation time, consistency measures, and morphologic subgroup-based quantitative measurements of canal volume and surface area. RESULTS: The model achieved a mean Dice similarity coefficient of 0.905, a mean precision of 0.892, a recall of 0.917, an accuracy of 0.997, and a mean 95% Hausdorff distance of 1.551 mm. Automated segmentation required approximately 11 seconds per case, compared with 30 minutes and 16 seconds for manual segmentation. The model demonstrated high algorithmic consistency across repeated runs. A higher recall relative to precision indicated that most of the canal structures were successfully identified, although some false-positive segmentations occurred. Quantitative measurements of canal volume and surface area derived from the 3D U-Net segmentation showed high agreement with manual segmentation across morphologic subtypes. CONCLUSIONS: A 3D U-Net-based framework achieved automated segmentation of the maxillary incisive canal on CBCT images with satisfactory accuracy and time efficiency, enabling reproducible 3D anatomic delineation and quantitative morphologic assessment. This framework represents a preliminary methodologic exploration and may be further optimized to reduce segmentation errors.