基于MRI的神经网络模型预测胶质母细胞瘤O^(6)-甲基鸟嘌呤-DNA-甲基转移酶启动子状态

Prediction of O^(6)-methylguanine-DNA ethyhransferase methylation status in glioblastoma based on deep learning

目的基于神经网络模型使用术前常规模态磁共振图像预测胶质母细胞瘤O^(6)-甲基鸟嘌呤-DNA-甲基转移酶(MGMT)启动子甲基化状态。方法回顾性分析兰州大学第二医院经手术病理证实的129例胶质母细胞瘤患者的T2WI序列和T1WI对比增强序列的磁共振图像,其中MGMT甲基化患者66例,MGMT非甲基化患者63例。基于EfficientNet卷积神经网络分别构建T2-net、T1C-net及两种序列联合的TS-net,通过受试者工作特征(ROC)曲线、曲线下面积(AUC)敏感度、准确度及特异度评价模型预测效能。结果TS-net在训练和验证数据集的AUC分别为0.944和0.838,T1C-net模型的AUC分别为0.951和0.830,T2-net的AUC分别为0.929和0.781。结论基于磁共振常规序列图像训练深度学习模型可有效预测胶质母细胞瘤患者的MGMT甲基化状态,帮助制定进一步治疗计划。

Objective To predict the O^(6)-methylguanine-DNA ethyhransferase(MGMT)methylation status of glioblastoma using preoperative normal-scale magnetic resonance imaging based on neural network model.Methods The MR images of T2WI sequence and T1WI contra-enhanced(T1C)sequence of 129 patients with glioblastoma confirmed by surgery and pathology were retrospectively analyzed,including 66 patients with MGMT methylation and 63 with non-methylation.T2-net,T1C-net and TS-net combined with the two sequences were established respectively based on the convolutional neural network EfficientNet.The prediction efficiency of the model was evaluated by the receiver operating characteristic(ROC)curve,sensitivity,accuracy and specificity.Results The AUC values of TS-net were 0.944 and 0.838 in training and validation data sets,0.951 and 0.830 in T1C-net model,and 0.929 and 0.781 in T2-net model,respectively.Conclusion The deep learning model based on MRI routine sequential image training can effectively predict the status of MGMT methylation in GBM patients and help develop further treatment plans.