A deep learning approach for radiological detection and classification of radicular cysts and periapical granulomas

Abstract

ObjectivesDentists and oral surgeons often face difficulties distinguishing between radicular cysts and periapical granulomas on panoramic imaging. Radicular cysts require surgical removal while root canal treatment is the first-line treatment for periapical granulomas. Therefore, an automated tool to aid clinical decision making is needed. MethodsA deep learning framework was developed using panoramic images of 80 radicular cysts and 72 periapical granulomas located in the mandible. Additionally, 197 normal images and 58 images with other radiolucent lesions were selected to improve model robustness. The images were cropped into global (affected half of the mandible) and local images (only the lesion) and then the dataset was split into 90% training and 10% testing sets. Data augmentation was performed on the training dataset. A two-route convolutional neural network using the global and local images was constructed for lesion classification. These outputs were concatenated into the object detection network for lesion localization. ResultsThe classification network achieved a sensitivity of 1.00 (95% C.I. 0.63–1.00), specificity of 0.95 (0.86–0.99), and AUC (area under the receiver-operating characteristic curve) of 0.97 for radicular cysts and a sensitivity of 0.77 (0.46–0.95), specificity of 1.00 (0.93–1.00), and AUC of 0.88 for periapical granulomas. Average precision for the localization network was 0.83 for radicular cysts and 0.74 for periapical granulomas. ConclusionsThe proposed model demonstrated reliable diagnostic performance for the detection and differentiation of radicular cysts and periapical granulomas. Using deep learning, diagnostic efficacy can be enhanced leading to a more efficient referral strategy and subsequent treatment efficacy. Clinical significanceA two-route deep learning approach using global and local images can reliably differentiate between radicular cysts and periapical granulomas on panoramic imaging. Concatenating its output to a localizing network creates a clinically usable workflow for classifying and localizing these lesions, enhancing treatment and referral practices.