Accuracy of deep learning models in the detection of accessory ostium in coronal cone beam computed tomographic images.

Accessory ostium [AO] is one of the important anatomical variations in the maxillary sinus. AO is often associated with sinus pathology. Radiographic imaging plays a very important role in the detection of AO. Deep learning models have been used in maxillofacial imaging for interpretation and segmentation. However, there have been no research papers investigating the effectiveness of CNN in detecting AO in radiographs. To fill this gap of knowledge, we conducted a study to determine the accuracy of deep learning models in detecting AO in coronal CBCT images. Two examiners collected 454 coronal section images (227 with AO and 227 without AO) from 856 large field of view [FOV] cone beam tomography [CBCT] scans in the dental radiology archives of a teaching hospital. The collected images were then pre-processed and augmented to obtain 1260 images. Three pre-trained models, the Visual Geometry Group of the University of Oxford-16 layers [VGG16], MobileNetV2, and ResNet101V2, were used as base models. The performance of all the models was analyzed, and ResNet101v2 was selected for classification of images. Fine-tuning approach was employed with L1 (Lasso regression) regularization to avoid overfitting. The test accuracy and loss of the ResNet-101V2 classification model was 0.81 and 0.51, respectively. The precision, recall, F1-score, and AUC of the classification model were 0.82, 0.81, 0.81, and 0.87 respectively. ResNet-101V2 showed good accuracy in the detection of AO from coronal CBCT images. The present study used cropped two-dimensional images of CBCT scans. Future work can be carried out to determine the accuracy of deep learning models in the detection of AO in three-dimensional CBCT scans.