摘要
目的结合扩散加权成像(DWI)和液体衰减反转恢复序列(FLAIR)的影像组学与机器学习构建醒后卒中的发病时间预测模型并进行验证。方法搜集2017年1月至2020年12月在南京市第一医院就诊的明确发病时间的急性缺血性脑卒中(AIS)患者266例,将患者随机分为训练集(n=185)和测试集(n=81),另搜集在南京医科大学附属江宁医院就诊的明确发病时间的AIS患者(n=56)为验证集,并根据患者发病时间将其分为≤4.5 h和>4.5 h两组。采用A.K.软件提取DWI和FLAIR影像组学特征并构建影像组学标签。通过多变量Logistic回归筛选最佳预测因子并构建诺莫图模型。利用受试者工作特征(ROC)曲线评估模型预测效能,并对模型进行内、外部验证。结果每例患者各提取1584个影像组学特征,降维后筛选出6类26个与卒中发病时间高度相关的特征。ROC显示联合独立预测因子影像组学标签、高血脂症构建的诺莫图模型预测训练集发病时间的曲线下面积(AUC)为0.979(灵敏度和特异度分别为0.923、0.977),预测测试集发病时间AUC为0.968(灵敏度和特异度分别为0.915、0.918),预测外部验证集发病时间AUC为0.897(灵敏度和特异度分别为0.837、0.819),三者间AUC值差异无统计学意义(Delong检验,P> 0.05)。结论基于DWI、FLAIR及临床资料所构建的预测模型能够较为准确地预测AIS患者的发病时间,且拥有较高的可靠性。
Objective To construct and validate a prediction model combine machine learning with radiomics characteristics of diffusion weighted imaging(DWI) and fluid attenuated inversion recovery(FLAIR) in predicting the onset time of wake up stroke. Methods A total of 266 acute ischemic stroke patients(AIS) with clear symptom onset from January 2017 to December 2020 in Nanjing First Hospital were retrospectively enrolled.The patients were randomly divided into training sets(185 cases) and test sets(81 cases).In addition, AIS patient with definite onset time(n=56) in Affiliated Jiangning Hospital of Nanjing Medical University were collected as the validation sets.The patients were divided into ≤ 4.5 h and > 4.5 h according to symptom onset time.The radiomics features of DWI and FLAIR were extracted and the radiomics labels were constructed by A.K.software.Multivariate logistic regression was used to screen the best predictors and construct the nomogram model.Receiver operating characteristic(ROC) curve was used to evaluate the predictive efficacy of the model.Meanwhile, internal and external validation of the model were analyzed. Results A total of 1584 radiomics features were extracted from each patient and 26 features of 6 categories highly related to the onset time of stroke were screened out after dimension reduction.ROC demonstrated that area under curve(AUC) of Nomogram constructed by tags of combined radiomics label and hyperlipidemia in the training sets was 0.979(sensitivity and specificity: 0.923,0.977),the AUC of the test sets was 0.968(sensitivity and specificity: 0.915,0.918),the AUC of external validation sets was 0.897(sensitivity and specificity: 0.837,0.819).There was no statistically significant difference in AUC values among the three groups(Delong test, P> 0.05). Conclusion The prediction model based on DWI,FLAIR and clinical data can predict the onset time of AIS patients accurately and has high reliability.
作者
艾中萍
姜亮
周蕾蕾
张宏
殷信道
王守巨
AI Zhongping;JIANG Liang;ZHOU Leilei(Department of Radiology,The First Affiliated Hospital of Nanjing Medical University,Nanjing,Jiangsu Province 210029,P.R.China)
出处
《临床放射学杂志》
北大核心
2022年第2期241-245,共5页
Journal of Clinical Radiology
基金
国家自然科学基金面上项目(编号:81871420)
国家自然科学基金优秀青年科学基金项目(编号:82022034)。
关键词
醒后卒中
影像组学
机器学习
扩散加权成像
液体衰减反转恢复序列
Wake up stroke
Radiomics
Machine learning
Diffusion weighted imaging
Fluid attenuated inversion recovery
