Comparison of nasopalatine canal detection in CONE-BEAM computed tomography images with artificial intelligence

NPC is an anatomical formation localized in the midline posterior to the maxillary central incisors. The morphology of the canal varies considerably. NPC can also be visualized with CBCT, an advanced imaging modality. In addition, CBCT is used in many studies in the field of artificial intelligence. Artificial intelligence is systems developed with the aim of solving problems by machines that can learn through data. YOLO (You Only Look Once) is a rapid object detection framework. In the last official update of YOLOv5, version 5 of YOLO, segmentation models were introduced. The aim of this thesis study is to compare the observer and artificial intelligence in terms of segmentation of NPC and branching status assessment in CBCT images. In the study, the CBCT images were retrospectively scanned. A total of 200 CBCT images of 100 canals with branching in NPC and 100 canals without branching were selected. The axial sectional frame view of these CBCT images was used for labeling. They were labeled and classified according to the presence or absence of NPC branching in the images. This thesis study was done with YOLOv5x-seg model. 80% of the images were divided into 160 training datasets, 10% validation datasets, and 10% test datasets. The training was done by making 800 epochs (training tours) with the YOLOv5x-seg model. Using the Confusion Matrix, the performance of the AI was evaluated. For intra-observer reliability, after 2 weeks, 20% of the images were reclassified for presence or absence of branching. Sensitivity, precision, F1 score, and IoU values for NPC detection and classification of the YOLOv5x-seg model; It was found as 0.9680, 0.9953, 0.9815, 0.9636 for the group with NPC branching, and as 0.9827, 0.9975, 0.9900, 0.9803 for the group without NPC branching, respectively. mAP and AUC values are found as 0.7930 and 0.8841 for the group with NPC branching, respectively, 0.9637 and 0.9510 for the group without NPC branching. We think that even when the YOLOv5x-seg model is trained with the presence of NPC branching and fewer datasets, it achieves a good, if not perfect, prediction accuracy above acceptable limits. We think that the segmentation feature of the YOLOv5 algorithm, which is basically an object detection algorithm, has achieved quite successful results despite its recent development, but is open to development and promising.