Research And Application Of Deep Learning Algorithm In C-shaped Root Canal Identification Of mandibular second molars

Abstract

Abstract Objective:C-shaped root canal morphology is complex and usually appears in the mandibular second molar. Accurate identification of its morphology facilitates clinical decision making.Deep learning has great advantages in image identification and classification by learning the inherent rules and representation levels of sample data.The purpose of this study is to evaluate the classification effect of deep learning algorithms in CBCT cross-sectional morphology of C-shaped root canals of mandibular second molars, and to develop the artificial intelligence recognition system using the best algorithm. Method:A total of 2266 CBCT cross-sectional images of mandibular second molars,including 1174 C-shaped and 1092 non C-shaped root canals were collected. Nine deep learning algorithms including GoogleNet,InceptionV3,InceptionResNetV2, MobileNet, NASNetMobile, ResNet152, ResNet50V2, ResNet101V2 and VGG were used to train these data models.The accuracy,model parameters, training time, confusion matrix and AUC were used to evaluate model performance. Result:The training accuracy of nine deep learning algorithms was 100.0%, 100.0%, 99.9%, 100.0%, 99.8%, 99.6%, 100.0%, 100.0%, 91.3%,the testing accuracy was 96.3%, 94.0%, 97.2%, 92.6%, 70.4%, 44.4%, 94.4%, 98.1%, 44.9%, the AUC value was 0.99, 0.99, 0.99, 0.98, 0.92, 0.88, 1.00, 0.99, 0.83, the training time was 102, 227, 862, 135, 189, 1009, 286, 502, 480min, the model parameters were 5975602, 21806882, 54339810, 3230914, 4271830, 58375042, 23568898, 42630658, 20025410 .The results show that NASNetMobile, ResNet152 and VGG algorithms have low test accuracy and poor generalization ability. ResNet101V2 has the highest test accuracy and the best effect. Conclusion: The deep learning algorithm can quickly and effectively identify the C-shaped root canals of the mandibular second molars, and ResNet101V2 algorithm works best. The prototype of artificial intelligence recognition system based on this algorithm can reduce the work intensity and subjectivity of doctors' recognition, and has a good application prospect.