摘要
目的探讨使用基于MR图像的卷积循环神经网络预测胶质瘤患者异柠檬酸脱氢酶(IDH)蛋白表达。方法选取符合纳入标准的胶质瘤患者236例,IDH蛋白表达阳性83例,IDH蛋白表达阴性153例。应用基于DenseNet-ResNet自动编码器的卷积循环神经网络(CRNN)预测胶质瘤患者IDH蛋白表达,构建基于T_(2)图像(T_(2)-net)、基于T_(1)增强图像(T_(1)C-net)和基于T_(2)+T_(1)增强图像(TU-net)三个独立模型,并用曲线下面积AUC、准确率、召回率、精确率及F1-score对各个模型预测效能进行评价。结果基于DenseNet-ResNet的CRNN网络能够预测胶质瘤IDH的蛋白表达,其中T_(2)-net模型的AUC为0.975、准确率为90.6%、召回率为81.0%、精确率为92.2%、F1-score为89.6%,T_(1)C-net模型的AUC为0.952、准确率为91.1%、召回率为83.0%、精确率为93.2%、F1-score为90.3%,TU-net模型的AUC为0.995、准确率为95.3%、召回率为90.6%、精确率为95.7%、F1-score为95.0%,TU-net模型的AUC、准确率、召回率、精确率及F1-score优于T_(1)-net和T_(2)-net模型。结论DenseNet-ResNet能够准确无创性预测胶质瘤患者的IDH蛋白表达,其中TU-net模型预测效果最佳。
Objective To predict isocitrate dehydrogenase(IDH)protein expression in glioma patients using MR image-based convolutional recurrent neural network.Methods A total of 236 patients with glioma diagnosed were included,among whom 83 were positive for IDH protein expression and 153 were negative for IDH protein expression.We built an autoencoder based on DenseNet-ResNet to predict IDH protein expression in glioma patients through a CRNN(Convolutional Recurrent Neural Network)-based classifier,and constructed three independent models of T_(2)-net(based on T_(2) images),T_(1)C-net(based on T_(1) Enhanced images),and TU-net(based on T_(2)+T_(1) enhanced images).Results The CRNN network based on DenseNet-ResNet could predict the protein expression of glioma IDH.The AUC of T_(2)-net model was 0.975,the accuracy rate was 90.6%,the recall rate was 81.0%,the precision rate was 92.2%,and the F1-score was 89.6%.The AUC of T_(1)C-net model was 0.952,the accuracy was 91.1%,the recall rate was 83.0%,the precision rate was 93.2%,and the F1-score was 90.3%.The AUC of the TU-net model was 0.995,the accuracy rate was 95.3%,the recall rate was 90.6%,the precision rate was 95.7%,and the F1-score was 95.0%.The AUC,accuracy,recall rate,precision and F1-score of the TU-net model were better than those of the T_(1)-net and T_(2)-net models.Conclusion DenseNet-ResNet can accurately and non-invasively predict IDH protein expression in glioma patients,and the TU-net model has the best prediction performance.
作者
施念
许倩
张纯
王贝茹
韩翠平
SHI Nian;XU Qian;ZHANG Chun;WANG Beiru;HAN Cuiping(Department of Medical Imaging,The Affiliated Hospital of Xuzhou Medical University,Xuzhou 221000,China)
出处
《医学影像学杂志》
2023年第10期1745-1749,共5页
Journal of Medical Imaging
基金
江苏省徐州市卫生健康委员会青年医学科技创新项目(编号:XWKYHT20210586)。
关键词
脑胶质瘤
异柠檬酸脱氢酶
卷积循环神经网络
磁共振成像
Brain glioma
Isocitrate dehydrogenase
Convolutional recurrent neural network
Magnetic resonance imaging
