Evaluation of artificial intelligence assisted FCN model performance: Application of automatic segmentation of the nasopalatine canal on CBCT images

Deep learning has become an important framework in medical image analysis, automating and standardizing the segmentation of regions of interest. This process is crucial for computer-aided diagnosis, interventional procedures, and treatment. Although cone-beam computed tomography(CBCT) has revolutionized dental imaging by providing detailed three-dimensional views of maxillofacial structures, several challenges remain in segmenting the nasopalatine canal(NPC) using CBCT. This study aimed to evaluate the automatic segmentation of the NPC on CBCT images based on the performance of a deep learning–based nnU-Net v2 model. The study was conducted retrospectively, and the data were divided into two groups as training (n=82) and test (n=8) sets. Automatic segmentation was evaluated using performance metrics commonly employed in clinical radiology and artificial intelligence, including true positives, false positives, false negatives, accuracy, precision, recall, F1-score, Dice coefficient, Jaccard index (IoU), 95% Hausdorff distance (mm), and Hausdorff distance (mm). In addition, the receiver operating characteristic (ROC) curve was generated, and the area under the curve (AUC) was calculated. According to the confusion matrix obtained with the nnU-Net v2 architecture for NPC segmentation, the values of true positives, false positives, false negatives, accuracy, precision, recall, and F1-score were found to be 914, 282.13, 131.25, 3.09, 0.11, 0.99, 0.74, 0.89, and 0.81, respectively. Furthermore, the Dice coefficient, Jaccard index (IoU), 95% Hausdorff distance (mm), and Hausdorff distance (mm) were calculated as 0.80, 0.67, 1.02 mm, and 1.90 mm, respectively. The AUC value derived from the ROC curve was 0.94. These findings demonstrate that the nnU-Net v2 model provides high accuracy and reliability in the segmentation of the nasopalatine canal.