机械取栓治疗前后扩散加权成像影像组学结合机器学习预测卒中的预后

Combining Machine Learning with Radiomics Based on Diffusion Weighted Imaging Before and After Mechanical Thrombectomy in Predicting Outcome of Stroke

目的探讨卒中机械取栓治疗前后扩散加权成像(DWI)影像组学的机器学习预测预后的效果。资料与方法回顾性分析2017年1月—2020年9月在南京市第一医院接受机械取栓治疗的卒中患者,其中训练集157例,测试集68例。采用A.K.软件分别提取治疗前后DWI梗死区影像组学特征,并应用最低绝对收缩和选择算子回归模型筛选最佳影像组学特征,基于所选特征通过支持向量机分类器建立卒中机械取栓后预后的预测模型,利用受试者工作特征曲线评价模型的预测效能。收集常州市第二人民医院卒中患者(验证集75例)对模型进行外部验证。结果每例患者治疗前后DWI图像共提取792个影像组学特征,降维后筛选出20个与预后高度相关的特征。受试者工作特征曲线分析显示支持向量机分离器建立的模型预测训练集患者预后的曲线下面积(AUC)为0.984,准确度达0.974;预测测试集患者预后的AUC为0.960,准确度达0.928;预测验证集患者预后的AUC为0.901,准确度达0.898;并具有较高的外部验证一致性(P>0.05)。结论基于治疗前后DWI的影像组学特征构建的模型对卒中机械取栓后预后预测具有较高的效能和较好的泛化能力。

Purpose To investigate the value of machine learning based on radiomics of diffusion weighted imaging(DWI)to predict the outcome after mechanical thrombectomy in acute stroke.Materials and Methods Acute stroke patients in Nanjing First Hospital from January 2017 to September 2020(training set 157 patients,test set 68 patients)were retrospectively analyzed.The imaging omics features were extracted from lesions on DWI before and after therapy using A.K.software,least absolute shrinkage and selection operator regression model was used to screen the features,and subsequently,the selected features were used to construct the prediction model by support vector machine classifier.Receiver operating characteristic curve was used to evaluate the predictive efficacy of the model.The stroke patients in Changzhou Second People’s Hospital(Validation set,n=75)were enrolled to verify the model.Results A total of 792 imaging omics features of each patient were extracted from DWI before and after therapy,and 20 features highly related to outcome after mechanical thrombectomy in acute stroke were screened after dimension reduction.Receiver operating characteristic curve analysis showed that the area under curve(AUC)of support vector machine classifier based on training set in predicting outcome was 0.984 and the accuracy was 0.974;the AUC of SVM model based on test set was 0.960 and the accuracy was 0.928;the AUC of SVM model based on validation set was 0.901 and the accuracy was 0.898.There was high consistency in external validation(P>0.05).Conclusion The imaging omics features and machine learning model based on DWI before and after therapy has high predictive efficiency and good generalization ability for patient outcome after mechanical thrombectomy in acute stroke.