基于DWI和FLAIR的机器学习预测急性脑卒中发病时间的研究

The study of machine learning based on DWI and FLAIR in the prediction of onset time of acute stroke

目的 基于急性脑卒中弥散加权成像(diffusion weighted imaging,DWI)和液体衰减反转恢复序列(fluid attenuated inversion recovery,FLAIR)的影像组学特征,通过机器学习构建急性脑卒中发病时间的预测模型。材料与方法 回顾性分析188例急性脑卒中患者的MRI图像。采用ITK-SNAP软件对DWI上高信号梗死区和FLAIR上对应的急性梗死区进行分割,并应用人工智能应用平台(artificial intelligent kit,A.K.)进行影像组学特征提取和降维,最终使用最小绝对收缩与选择算子算法(least absolute shrinkage and selection operator,LASSO)确定发病时间相关的最佳影像组学特征,通过支持向量机分类器评估其在发病时间预测中的价值,并与人工识别的结果进行比较。结果 共筛选出10个(7个DWI特征及3个FLAIR特征)与卒中发病时间密切相关的影像组学特征。人工识别受试者操作特征(receiver operating characteristic,ROC)分析显示DWI-FLAIR不匹配预测急性脑卒中发病时间的曲线下面积(area under curve, AUC)为0.634,敏感度和特异度分别为0.667、0.622。ROC分析显示该模型预测训练集患者发病时间的AUC为0.975,敏感度和特异度分别为0.932、0.950;预测测试集患者发病时间的AUC为0.915,敏感度和特异度分别为0.868、0.852。结论 基于DWI和FLAIR影像组学的机器学习,能够较为准确地预测急性脑卒中患者的发病时间,为临床静脉溶栓治疗的选择提供影像指导。

Objective: To construct a prediction model of onset time in acute stroke using machine learning based on the radiomic features of diffusion weighted imaging(DWI) and fluid attenuated inversion recovery(FLAIR). Materials and Methods: A total of 188 acute stroke patients receiving MRI were retrospectively enrolled. The ITK-SNAP software was used to segment the high signal areas of DWI and the acute infarct areas of FLAIR. The artificial intelligent kit(A. K.) software was used to extract the radiomic features and reduce the dimensionality. The least absolute shrinkage and selection operator(LASSO) regression analysis was used to determine the radiomic features related to onset time. The support vector machine classifier was used to evaluate its value in onset time prediction, and compared with those of human readings. Results: A total of 10 radiomic features(7 DWI features and 3 FLAIR features) closely related to stroke onset time were screened. The receiver operating characteristic(ROC) analysis of human readings showed that the area under curve(AUC) of DWI-FLAIR mismatch in predicting onset time of acute stroke was 0.634, and the sensitivity and specificity were 0.667, 0.622,respectively. ROC analysis showed that AUC of the prediction model based on the training set was 0.975, the sensitivity and specificity were 0.932 and 0.950 respectively;the AUC of the prediction model based on the test set was 0.915, the sensitivity and specificity were0.868 and 0.852 respectively. Conclusions: Machine learning based on DWI and FLAIR radiomics can accurately predict the onset time of acute stroke patients and provide image guidance for the selection of thrombolytic therapy in clinical.