Deep learning based hybrid system for automatic detection and classification of nasopalatine canal in cone beam computed tomography images.

PURPOSE: The objective of this study was to develop and evaluate a deep learning-based hybrid system for the automatic detection and classification of the nasopalatine canal (NPC) via cone beam computed tomography (CBCT) images. METHODS: CBCT images from 135 patients were analysed retrospectively. The proposed system incorporates a YOLO-based object detection model (YOLOv8, YOLOv9, YOLOv10 and YOLOv11) for NPC localization and advanced deep learning classification models, including MobileNetV3Large, ResNet50, and EfficientNetV2B0. Model performance was evaluated via metrics such as precision, recall, F1 score, and accuracy. Training and testing times were also analysed to assess computational efficiency. RESULTS: The YOLOv10 algorithm demonstrated the most effective detection performance, achieving 100% recall and 99.5% mAP50. Among the classification models, MobileNetV3Large exhibited the highest accuracy (81.48%), precision (82.60%), and F1 score (81.47%). EfficientNetV2B0 closely followed, achieving an accuracy of 79.26%. CONCLUSION: Automatic localization and morphological classification of the NPC on CBCT images appears feasible. The proposed framework may assist clinicians in recognizing anatomical variations and selecting cases that require more detailed radiological evaluation before anterior maxillary procedures. Nevertheless, the findings are preliminary, and expert verification and external validation remain necessary.