基于机器学习的大脑白质影像组学标签识别帕金森病早期阶段的应用研究  

Application of brain white matter radiomics signature based on machine learning in prediction of early-stage Parkinson's disease

作　　者：徐建国[1] 宋侨伟 沈莹 舒震宇[1] XU Jianguo;SONG Qiaowei;SHEN Ying;SHU Zhenyu(Department of Radiology,Zhejiang Provincial People’s Hospital(Affiliated People’s Hospital of Hangzhou Medical College),Hangzhou 310014,China)

机构地区：[1]浙江省人民医院(杭州医学院附属人民医院)放射科,杭州310014 [2]绍兴文理学院医学部

出　　处：《浙江医学》2020年第18期1954-1959,1964,共7页Zhejiang Medical Journal

基　　金：浙江省中医药科技计划项目(2020ZA009)。

摘　　要：目的使用机器学习开发并验证一种基于大脑白质的影像组学标签用于预测帕金森病(PD)的早期阶段。方法从PD进展标记倡仪数据库(PPMI)中收集340例受检者的影像和临床资料,包括171例健康对照人群和169例PD患者。所有受试者按7:3随机分为训练组237例和测试组103例。在训练集的基线MRI中,切割出三维大脑白质以提取每例患者的影像组学特征,进行降维后使用机器学习方法构建影像组学标签。利用ROC曲线评估影像组学标签在训练组和测试组中的诊断效能,用Hosmer-Lemeshow检验分析标签的拟合优度。将所有数据集按照ROC曲线的截断值分为高危组和低危组,比较两组PD患者例数,以确定标签的临床效果。利用ROC曲线和决策曲线(DCA)分别评估标签在所有PPMI数据中的准确性和净效益。结果训练组和测试组的影像组学标签AUC分别为0.849和0.824,灵敏度分别为0.75和0.78,特异度分别为0.87和0.87。Hosmer-Lemeshow检验表明,标签在训练组和测试组中的拟合优度比较差异无统计学意义(P>0.05)。将所有数据集按模型的最佳诊断阈值(截断值=0.1338)分为高危组144例和低危组196例。其中高危组PD患者123例,HC人群21例;低危组PD患者44例,HC人群152例;两组PD患者例数比较差异有统计学意义(P<0.05)。影像组学标签在所有数据集中的诊断准确性为0.823,DCA曲线也显示了良好的净效益。结论基于常规MRI的大脑白质影像组学标签对PD患者表现出良好的鉴别性能,证明了其在PD患者识别中的临床应用价值。Objective To develop and validate a radiomics model for predicting Parkinson's disease(PD)based on brain white matter by machine learning.Methods Brain images of 340 subjects from Parkinson's progress markers Initiative(PPMI)database,including 169 PD patients(most of whom in early stage)and 171 healthy controls(HC).All subjects were randomly divided into the training(n=237)and test(n=103)sets with a ratio of 7:3.In the baseline MR images,the white matter was segmented to extract and score the radiomic features of each patient,and to establish radiomics model using machine learning after the dimensionality of data of training group were reduced.ROC curve was used to evaluate the diagnostic efficiency of radiomics signature in training and test sets,and Hosmer-Lemeshow test was used to analyze the goodness of fit of signature.In addition,all data sets were divided into high-risk group and low-risk group according to the cut-off value of ROC curve.The number of patients with PD in the two groups was compared to determine the clinical effect of signature.Finally,ROC curve and decision curve analysis(DCA)curve were used to evaluate the accuracy and net benefit of signature in all PPMI data.Results The area under curve(AUC)of radiomics model in the training and testing sets were 0.849 and 0.824,the sensitivity were 0.75 and 0.78,the specificity were 0.87 and 0.87.Hosmer-Lemeshow test showed that there was no significant difference in the fitting degree of signature in the training and testing sets(P>0.05).According to the best diagnostic threshold(cut-off value:0.1338),there were 144 subjects in the high-risk group,including 123 PD patients and 21 HC,196 subjects in the low-risk group,including PD 44 patients and 152 HC,the number of patients with PD in the two groups was significantly different(P<0.05).The diagnostic accuracy of the signature in all subjects was 0.823,and the DCA curve also showed a good net benefit.Conclusion The radiomics model based on white matter from convention MRI is a non-invasive tool for pre

关 键 词：影像组学 白质 磁共振成像 机器学习 帕金森病 

分 类 号：R742.5[医药卫生—神经病学与精神病学]