基于脑梗死和脑脊液的影像组学预测急性脑卒中机械取栓术后恶性脑水肿发生的分析  

The study of machine learning based on DWI infarction and cerebrospinal fluid in predicting malignant edema after thrombectomy for acute stroke

作　　者：徐敬敬 王志楠 徐翔 马跃虎[2] 彭明洋 陈国中 谢光辉[2] XU Jingjing;WANG Zhinan;XU Xiang;MA Yuehu;PENG Mingyang;CHEN Guozhong;XIE Guanghui(Department of Radiology,Nanjing Integrated Traditional Chinese And Western Medicine Hospital,Nanjing 210000,China;Department of Medical Imaging,Nanjing First Hospital,Nanjing Medical University,Nanjing 210006,China)

机构地区：[1]江苏省南京市中西医结合医院放射科,江苏南京210000 [2]南京医科大学附属南京医院(南京市第一医院)医学影像科,江苏南京210006

出　　处：《医学影像学杂志》2024年第1期7-10,共4页Journal of Medical Imaging

基　　金：国家自然科学基金项目(编号:82001811)。

摘　　要：目的探讨基于急性脑卒中弥散加权成像(diffusion weighted imaging,DWI)核心梗死区及脑脊液的影像组学特征,通过机器学习构建急性脑卒中机械取栓术后恶性脑水肿发生的预测模型。方法选取155例急性脑卒中患者的MRI图像,基于软件自动分割DWI高信号梗死区及脑脊液区,应用AK软件进行影像组学特征提取和降维,采用最小冗余最大相关特征选择方法筛选最佳影像组学特征,通过支持向量机分类器构建恶性水肿发生的预测模型。结果经筛选后具有10个特征的子集(包含7个DWI脑梗死特征和3个脑脊液特征)获得了最高的平均AUC值,并被指定为最终的特征子集纳入模型分析。ROC分析显示训练集患者预测恶性脑水肿的AUC为0.975,敏感度、特异度和准确度分别为90.3%、69.8%、93.5%;测试集患者预测恶性脑水肿的AUC为0.893,敏感度、特异度、准确度分别为86.8%、90.3%、87.1%。结论基于DWI脑梗死和脑脊液影像组学特征的机器学习,能够较为准确地预测急性脑卒中机械取栓术后恶性脑水肿的发生,为临床早期干预治疗提供指导。Objective To construct a prediction model of malignant edema after thrombectomy for acute stroke using machine learning based on the radiomic features of diffusion weighted imaging(DWI)infarction and cerebrospinal fluid.Methods A total of 155 acute stroke patients receiving MRI were retrospectively enrolled.The DWI infarction and cerebrospinal fluid were segmented using software.The AK software was used to extract the radiomic features and to reduce the dimensionality.The the minimum redundancy and maximum correlation feature selection method was used to determine the radiomic features.The support vector machine classifier was used to construct prediction model of malignant edema.Results The subset with 10 features(including 7 DWI infarction features and 3 cerebrospinal fluid features)obtained the highest average the area under curve(AUC)value after screening,and was designated as the final feature subset for model analysis.The ROC analysis showed that the AUC of predicting malignant edema in training set was 0.975,and the sensitivity and specificity were 0.903,0.968,respectively,and the AUC of predicting malignant edema in test set was 893,and the sensitivity and specificity were 0.868,0.903,respectively.Conclusion Machine learning based on DWI infarction and cerebrospinal fluid can accurately predict the malignant edema after thrombectomy for acute stroke,and can provide guidance for clinical early intervention.

关 键 词：卒中 磁共振成像 脑脊液 恶性脑水肿 机器学习 

分 类 号：R743.3[医药卫生—神经病学与精神病学] R445.2[医药卫生—临床医学]