Identification of Root Resorption on Panoramic X-ray Image Based on EfficientNet

Root resorption is a disease in which the cells of the root of the tooth are destroyed. This disease may be harmless, depending on the cause. When the root resorption has progressed, early detection and treatment is necessary. If root resorption is left untreated, the lesion can progress to pulp exposure, necrosis, sinus and periodontal defects, and perforation. Diagnosis of root resorption is made by clinical examination and a panoramic x-ray machine, however accurate diagnosis requires computed tomography (CT). CT allows full-plane examination of the area of interest and provides an accurate diagnosis. However, CT is less widely used than panoramic radiography in private dental office. Therefore, there is a need for a method to detect root resorption from panoramic radiographs with a high reliability comparable to that of CT images. In this paper, we propose an image analysis method to identify the presence or absence of root resorption from panoramic radiographs. In particular, we propose a method that uses EfficientNet to extract features with fine-tuning and learns them using L2-softmax loss as an error function. We applied the proposed method to panoramic X-ray images of 59 actual cases and obtained TPR: 71% and FPR: 43% respectively.