关键词: 机器学习, CBCT, 卷积神经网络, 图像分类, 口腔外科

Abstract: PURPOSE: To deeply integrate alveolar surgery with artificial intelligence and change the existing diagnosis and treatment mode to guide clinical work better, especially to provide help for primary oral surgeons. METHODS: CBCT data of 214 patients with impacted mandibular third molar admitted to West China Stomatology Hospital of Sichuan University from September 2022 to December 2022 were randomly selected and divided into training dataset and test dataset. According to the clinical experience of human experts, six-type classification of resistance sources were proposed, and patients’ CBCT data were artificially classified. Then, the classification features were deeply learned and tested through the artificial intelligence oral surgeon(AIOS) model. Finally, the confusion matrix graph and Accuracy-Loss-Epoch curve were used to analyze the learning process and results. RESULTS: All model training datasets achieved 99.07%-100% accuracy. In the test dataset, the accuracy of all models reach more than 80%, and some models can reach 100% accuracy. CONCLUSIONS: AIOS has shown promising prospects for predicting the source of resistance of impacted mandibular third molar on CBCT images and assisting clinical oral surgeons in oral image analysis, laying a good foundation for the development of a full set of AIOS with resistance analysis, protocol formulation, and simulated surgery that can be applied in the clinic of oral surgery in the future.

Key words: Machine learning, CBCT, Convolutional neural network, Image classification, Oral surgery