基于MRI影像组学特征构建脑胶质瘤IDH1突变预测模型  

Prediction model construction based on MRI radiomics features for evaluating IDH1 mutation of glioma

作　　者：江少凡[1] 宋阳 薛蕴菁[1] 蒋日烽[1] JIANG Shao-fan;SONG Yang;XUE Yun-jing(Departments of Radiology,Fujian Medical University Union Hospital,Fuzhou 350003,China)

机构地区：[1]福建医科大学附属协和医院放射科,福建福州350003 [2]华东师范大学上海磁共振重点实验室,上海200241

出　　处：《放射学实践》2023年第12期1493-1499,共7页Radiologic Practice

基　　金：福建医科大学启航基金项目(2019QH1034)。

摘　　要：目的:探讨基于不同MRI序列的影像组学特征构建的机器学习(ML)模型预测胶质瘤IDH1突变的价值。方法:回顾性搜集经手术病理证实的161例胶质瘤患者(70例IDH1突变型/91例野生型)的临床资料,主要包括年龄、性别、Karnofsky功能状态(KPS)评分和肿瘤的病理分级。所有患者术前行MRI检查获得T2WI、T2-FLAIR、ADC图及对比增强T1WI图像,术后病理标本均行IDH1检测。将161例患者按照7∶3的比例随机分配为训练集和测试集。由2位影像医师利用Image J软件共同对病灶在配准过的T2-FLAIR或对比增强T1WI序列上进行逐层ROI的勾画,最后形成感兴趣区容积(VOI),然后使用FAE软件在各序列图像上提取VOI的影像组学特征,基于训练集的数据,通过均值归一化、方差分析的特征选择方法、皮尔逊相关系数的特征降维方法、4种ML分类器(线性判别分析、LASSO回归、逻辑回归、支持向量机)以及十折交叉验证法构建15种ML模型,并采用ROC曲线和Delong检验在测试集中筛选出最佳序列或序列组合;再基于最佳序列或序列组合,使用均值归一化、2种特征选择方式(方差分析、特征权重算法)、2种降维方式(皮尔逊相关系数和主成分分析法)、4种ML分类器构建了16种ML模型,通过FAE软件的one-standard error法、Delong检验和ROC曲线筛选拟合度较好且AUC最高的ML模型,并将此模型联合临床参数构建联合模型,评价各联合模型的诊断效能。结果:基于ADC图+对比增强T1WI序列组合提取的组学特征构建的4种ML模型在测试集中的AUC分别为0.888、0.872、0.896和0.877,均高于其它序列或序列组合构建的ML模型;利用此序列组合构建的16种ML模型中,以方差分析特征选择法、主成分分析方法的降维方式及分类器为LASSO回归时所构建的ML模型具有较好的拟合度且测试集中的AUC最高,为0.829(95%CI:0.658~0.966),此模型结合KPS评分和肿瘤病理分级所构建的联合模Objective:To investigate the value of different machine learning models based on radiomics features of different MRI sequences in predicting IDH1 mutations in glioma.Methods:The age,sex,Karnofsky Performance Status(KPS)score and pathological grading data of 161 patients with glioma(70 IDH1 mutant and 91 wild-type)confirmed by operation and pathology were retrospectively collected.All patients underwent preoperative MRI to obtain T 2WI,T 2-FLAIR,ADC and contrast-enhanced T 1WI images,and postoperative pathological specimens were detected by IDH1.161 patients were randomly assigned to a training set and a test set in a 7∶3 ratio.Two radiologists delineated the each lesion to obtain 2D-ROIs with Image J software on each slice of the lesion on registered T 2-FLAIR or enhanced T 1WI sequences and volume of interest(VOI)was formed.Then the imaging radomics features of VOI are extracted from each sequence of images using FAE software.Based on the data from the training set,Fifteen ML models were constructed by means normalization,feature selection method of ANOVA,feature dimensionality reduction method of Pearson correlation coefficient,four ML classifiers(linear discriminant analysis,LASSO regression,logistic regression,support vector machine)and ten-folds cross-validation method,and then the best sequence or sequences combination was selected out in test set by ROC curve analysis and Delong test.Based on the best sequence or sequence combination,16 ML models were constructed through mean normalization,two feature selection method of ANOVA and feature weight algorithm,two feature dimensionality reduction Pearson correlation coefficient/principal component analysis method,and 4 ML classifiers.The ML models with good fit and highest AUC value were selected out by one standard error method,Delong test and ROC curve.Finally,clinical parameters were combined to evaluate the diagnostic efficacy of each combined model.Results:The AUC values of the four ML models constructed based on the radiomics features extracted from the

关 键 词：胶质瘤 异柠檬酸脱氢酶突变 影像组学 磁共振成像 预测模型 

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]