Abstract:

Objective To develop a deep learning network based on attention mechanism to identify the number of the vertical mandibular third molar(MTM) roots(single or double) on panoramic radiographs in an automatic way. Methods The sample consisted of 1 045 patients with 1 642 MTMs on paired panoramic radiographs and Cone-beam computed tomography(CBCT) and were randomly grouped into the training(80%), the validation(10%), and the test(10%). The evaluation of CBCT was defined as the ground truth. A deep learning network based on attention mechanism, which was named as RN-MTMnet, was trained to judge if the MTM on panoramic radiographs had one or two roots. Diagnostic performance was evaluated by accuracy, sensitivity, specificity, and positive predict value(PPV), and the receiver operating characteristic(ROC)curve with the area under the ROC curve(AUC). Its diagnostic performance was compared with dentists’ diagnosis, Faster-RCNN, CenterNet, and SSD using evaluation metrics. Results On CBCT images, single-rooted MTM was observed on 336(20.46%) sides, while two-rooted MTM was 1 306(79.54%). The RN-MTMnet achieved an accuracy of 0.888, a sensitivity of 0.885, a specificity of 0.903, a PPV of 0.976, and the AUC value of 0.90. Conclusion RN-MTMnet is developed as a novel, robust and accurate method for detecting the numberof MTM roots on panoramic radiographs.

Key words: mandibular third molar, panoramic radiography, CBCT, deep learning, attention mechanism