基于T2WI的影像组学模型预测胶质母细胞瘤Ki-67的表达水平  

Prediction of Ki-67 expression level in glioblastoma by radiomics model based on T2WI

作　　者：朱雪超 何玉麟[1] 邬莺莺 黎斌 万天意 唐辛 余秋月 ZHU Xuechao;HE Yulin;WU Yingying;LI Bin;WAN Tianyi;TANG Xin;YU Qiuyue(Department of Radiology,the First Affiliated Hospital of Nanchang University,Nanchang 330006,China)

机构地区：[1]南昌大学第一附属医院放射科,南昌330006

出　　处：《磁共振成像》2021年第9期53-56,共4页Chinese Journal of Magnetic Resonance Imaging

摘　　要：目的探讨基于T2WI建立的影像组学模型术前预测胶质母细胞瘤Ki-67表达水平的价值。材料与方法回顾性分析经病理确诊的96例胶质母细胞瘤(glioblastoma,GBM)患者的术前MRI影像,根据Ki-67表达水平分为低表达组(Ki-67<50%)和高表达组(Ki-67≥50)。在T2WI轴位图像上手动勾画感兴趣体积(volume of interest,VOI)并提取影像组学特征,所有病例按照70%∶30%分为训练组和测试组,训练组用于特征筛选和建立机器学习模型,特征筛选由t检验和最小绝对收缩与选择算子(least absolute shrinkage and selection operator,LASSO)完成,特征筛选后的数据建立随机森林(random forest,RF)、Logistic回归和支持向量机(support vector machine,SVM)三种机器学习模型;测试组用于验证建立的模型并绘制ROC曲线,结果表示为准确度、敏感度、特异度和AUC。结果Ki-67低表达组和高表达组的年龄、性别差异无统计学意义;三种机器学习模型中,RF模型的诊断效能最高,准确度、敏感度、特异度和AUC分别为0.72、0.67、0.76、0.72,Logistic回归模型综合诊断效能最低,SVM模型介于二者之间。结论基于T2WI建立的影像组学模型对术前预测胶质母细胞瘤Ki-67表达水平具有一定的价值,其中RF模型预测效能最好。Objective:To explore the value of T2WI imaging model in predicting Ki-67 proliferation level of glioblastoma before operation.Materials and Methods:The preoperative MRI images of 96 patients with glioblastoma diagnosed by pathology in our hospital were retrospectively analyzed.According to the expression level of Ki-67,the patients were divided into two groups:low expression group(Ki-67<50%)and high expression group(Ki-67≥50).The volume of interest(VOI)was manually sketched on the T2WI axial image and the imaging features were extracted.All cases are divided into training group and test group according to 70%∶30%.The training group was used for feature screening and the establishment of machine learning models.Feature selection was completed by t-test and LASSO.After feature screening,three machine learning models of random forest(RF),Logistic regression and support vector machine(SVM)were established.The test group was used to verify the established model and draw ROC curves,and the results were expressed as accuracy,sensitivity,specificity and AUC.Results:There was no significant difference in age and sex between Ki-67 low expression group and high expression group.Among the three machine learning models,the RF model has the highest diagnostic efficiency,and the accuracy,sensitivity,specificity and AUC are 0.72,0.67,0.76 and 0.72 respectively.The comprehensive diagnostic efficiency of the SVM model is the lowest,while the SVM model is between them.Conclusions:The imaging model based on T2WI image has a certain value in predicting the level of Ki-67 expression in glioblastoma before operation,among which RF model is the best.

关 键 词：磁共振成像 胶质母细胞瘤 KI-67 机器学习 影像组学 T2加权成像 

分 类 号：R445.2[医药卫生—影像医学与核医学] R739.41[医药卫生—诊断学]