EVALUATION AND AUTOMATED SEGMENTATION OF THE NASOPALATINE CANAL USING CBCT AND DEEP LEARNING (U-NET) ARCHITECTURE

Background: Background and Objective: The nasopalatine canal (NPC) is an important structure in maxillofacial
surgery and dentistry, and has implications for surgical and dental treatments. This study aimed to assess
morphological differences and anatomical dimensions of the NPC with cone-beam computed tomography (CBCT)
and to investigate the usefulness of deep learning with U-Net architecture, in the automated segmentation of the
canal.
Methods:441 adult CBCT scans were studied in a retrospective cross-sectional manner. NPC was evaluated in
sagittal view to determine the shape, length, and width at three levels, as well as age-related changes. After evaluation,
sagittal images were extracted. An artificial intelligence (AI) model based on U-Net, trained from manually annotated
data, was then applied for segmentation. The model was evaluated using metrics such as the Dice score, accuracy,
precision, recall, and F1 score.
Results: Cylindrical and funnel shapes were the most common across all age groups, with no significant difference
in shape distribution between the two groups. NPC width did not significantly change with age, but canal length
significantly increased with age (P = 0.0003). For the U-Net model, Class 2 demonstrated higher training performance compared to the others (Dice = 0.531, Recall = 0.682). In the testing phase, Class 3 demonstrated the highest Dice value (0.868) and accuracy (1.000); whereas Class 4 had the lowest accuracy (Dice = 0.263, Accuracy = 0.790).
Conclusion: CBCT can provide detailed evaluation of the NPC morphology. The U-Net-based AI model
demonstrated substantial performance in segmenting the NPC. Combining AI with CBCT could improve diagnostic
accuracy and treatment planning in maxillofacial surgeries․