摘要
Abstract
Objective:To clinically collect a dataset of plaster models from crown debonding and non-debonding ca-ses,and to use deep learning models to predict the risk of crown debonding.Methods:Clinically collect plaster models of crown debonding and non-debonding cases.Obtain three-dimensional images using an intraoral scanner and generate two-di-mensional images for the dataset.The dataset includes 20 cases(10 cases of debonding and 10 cases without debonding),re-sulting in a total of 16,920 images.These images are randomly allocated into 16,200 training images,360 testing images,and 360 validation images.The performance of three models—CNN,ResNet50,and EfficientNetB0—is compared to determine which model performs best.Results:The accuracy,precision,F-measure for the three models were as follows:CNN(0.995,0.993,0.991),ResNet50(0.957,0.962,0.952),EfficientNetB0(0.885,0.901,0.879),and for AUC:CNN(0.999),Res-Net50(0.997),and EfficientNetB0(0.992).Conclusion:Among CNN,ResNet50,and EfficientNetB0,the CNN model dem-onstrated the best performance in predicting the debonding of full crown restorations.关键词
机器学习/深度学习/卷积神经网络/人工智能/全冠修复体/可视化分析/卷积块注意力模块Key words
Machine learning/Deep learning/Convolutional neural network/Artificial intelligence/Full crown resto-ration/Data visualization analysis/Convolutional block attention module(CBAM)分类
口腔医学
