Page 123 - 南京医科大学学报自然科学版

第42卷第8期 南京医科大学学报（自然科学版）
2022年8月 Journal of Nanjing Medical University（Natural Sciences） ·1165 ·

·临床研究·

机器学习结合影像组学特征预测急性脑卒中机械取栓预后

陈罕奇 ，张 浩 ，葛晓敏 ，彭明洋 ，谢光辉 ，陈国中 ，殷信道 ，许 瑜 3
1
1
2
2
2
1
2
南京医科大学附属常州市第二人民医院放射科，江苏 常州 213003；南京医科大学附属南京医院（南京市第一医院）医学
1 2
影像科，江苏 南京 210006；常州金东方医院放射科，江苏 常州 213000
3
［摘 要］ 目的：结合机器学习与影像组学特征构建预测急性缺血性脑卒中（acute inschemic strohe，AIS）机械取栓治疗后预后
的模型并进行验证。方法：回顾性分析在南京市第一医院就诊的AIS患者，按随机数字表法分为训练集（n=105）和测试集（n=
50），另收集在南京医科大学附属常州市第二人民医院就诊的AIS患者（n=45）作为外部验证。采用A.K.软件提取弥散加权成
像（diffusion weighted imaging，DWI）和灌注加权成像（perfusion weighted imaging，PWI）病变区的影像特征，应用最低绝对收缩和
选择算子（least absolute shrinkage and selection operator，LASSO）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（support vector machine，SVM）分类器建立预测急性脑卒中预后预测模型，使用受试者操作特征（receiver operating character⁃
istic，ROC）曲线评价模型的预测效能，并应用验证集对模型进行外部验证。结果：每例患者DWI和PWI图像各提取1 316个影
像组学特征，降维后筛选出40个与卒中预后高度相关的特征。ROC曲线分析显示联合DWI+PWI的模型预测训练集和测试集
的曲线下面积（area under curve，AUC）（训练集：0.981；测试集：0.891）均高于单序列模型（DWI或PWI），其准确度分别达0.943、
0.900。外部验证结果显示基于 DWI+PWI 的模型同样优于单序列（DWI 或 PWI）的预测模型，灵敏度和特异度分别为 0.864、
0.783，其准确度可达0.822。结论：结合机器学习与影像组学特征构建的模型可预测AIS机械取栓治疗预后，并具有较好的泛
化能力。
［关键词］ 卒中；机器学习；弥散加权成像；灌注加权成像；预后
［中图分类号］ R445.2 ［文献标志码］ A ［文章编号］ 1007⁃4368（2022）08⁃1165⁃06
doi：10.7655/NYDXBNS20220818

A model combined machine learning with radiomics characteristics in predicting outcome
after mechanical thrombectomy in acute stroke
CHEN Hanqi ，ZHANG Hao ，GE Xiaomin ，PENG Mingyang ，XIE Guanghui ，CHEN Guozhong ，YIN Xindao ，
1
2
2
2
1
2
1
XU Yu 3
1 Department of Radiology，the Second People’s Hospital of Changzhou，Nanjing Medical University，Changzhou
213003； Department of Radiology，Nanjing First Hospital，Nanjing Medical University，Nanjing 210006；
2
Department of Radiology，Changzhou Jindongfang Hospital，Changzhou 213000；China
3
［Abstract］ Objective：To establish and validate a prediction model combined machine learning with radiomics features in predicting
outcome after mechanical thrombectomy in acute stroke. Methods：Imaging data of acute stroke patients in Nanjing First Hospital were
retrospectively collected. These patients were divided into a training set（n=105）and a test set（n=50）according to random number
table method. Acute stroke（n=45）in the Second People’s Hospital of Changzhou were collected as the validation set. A.K. software
was used to extract radiomics features on diffusion weighted imaging（DWI）and perfusion weighted imaging（PWI）. Least absolute
shrinkage and selection operator（LASSO）regression model was used to screen the features，and then，the selected features were used to
establish the prediction model by support vector machine（SVM）classifier. Receiver operating characteristic（ROC）curve was used to
evaluate the predictive efficacy of the model，and the validation set was used to verify the generalization ability of the model. Results：
One thousand three hundred and sixteen radiomics features of each patient were extracted from DWI and PWI，and 40 features highly
related to outcome after mechanical thrombectomy in acute stroke were screened after dimension reduction. ROC analysis showed that
the area under curve（AUC）of DWI+PWI model（training set：0.981；test set：0.891）was higher than those of DWI or PWI model，and
the accuracy were 0.943 and 0.900，respectively. The results of validation of the model showed that the prediction model based on