The present study aimed to provide and evaluate the efficiency of an artificial intelligence mechanism for detecting cystic lesions on cone beam computed tomography (CBCT) scans.

The CBCT image dataset consisted of 150 samples, including 50 cases without lesions, 50 dentigerous cysts (DC), and 50 periapical cysts (PC) based on both radiographic and histopathological diagnosis. The dataset was divided into a development set with 70 % of samples for training and validation and a final test set with the other 30 % of samples. Four images were obtained for each case, including panoramic, manually segmented panoramic, axial, and manually segmented axial images. A deep convolutional neural network (CNN) architecture was used for automatic lesion detection and diagnosing the type of cystic lesion. To increase the number of image samples and avoid overfitting, a data augmentation procedure was applied. Recall, precision, F1-score, and average precision (AP) values were measured for lesion detection performance, and sensitivity, specificity, and accuracy indicators from the confusion matrix were calculated for the lesion classification performance of the CNN model.

Mean average precision, recall, and F1-score for the detection of DCs and PCs were respectively, 0.87, 0.92, and 0.89 before data augmentation, and 0.93, 0.95, and 0.93, after the augmentation process. For the classification of DCs with data augmentation, sensitivity, specificity, accuracy, and AUC values were 96.4 %, 99.5 %, 97.3 %, and 0.98, respectively, and for PCs with augmentation, these values were 89.6 %, 98.9 %, 98.1 %, and 0.94, respectively. Lastly, for no lesion samples, sensitivity, specificity, accuracy, and AUC values were 100 %, 99.1 %, 99.4 %, and 0.99, respectively, by application of data augmentation.

Our developed deep learning-based CNN algorithm showed high accuracy, sensitivity, and precision values (more than 90 %) for detecting and classifying dentigerous and periapical cysts on CBCT images using data augmentation.