T2WI机器学习在鉴别高级别胶质瘤和脑单发转移瘤中的应用  

Application of T2WI machine learning in the identification of high-grade gliomas and brain solitary metastases

作　　者：刘哲 刘小彤[3] 尹畅畅 梁挺 卞益同[2] 王斐 钱步月[3] 杨健 LIU Zhe;LIU Xiao-tong;YIN Chang-chang;LIANG Ting;BIAN Yi-tong;WANG Fei;QIAN Bu-yue;YANG Jian(School of Life Science and Technology, Xi'an Jiaotong University, Xi'an 710049;Department of Diagnostic Radiology, The First Affiliated Hospital of Xi'an Jiaotong University, Xi'an 710061;School of Electronic and Information Engineering, Xi'an Jiaotong University, Xi'an 710049, China)

机构地区：[1]西安交通大学生命科学与技术学院,陕西西安710049 [2]西安交通大学第一附属医院放射科,陕西西安710061 [3]西安交通大学电子与信息工程学院,陕西西安710049

出　　处：《西安交通大学学报（医学版）》2019年第5期794-799,814,共7页Journal of Xi’an Jiaotong University（Medical Sciences）

基　　金：国家重点研发计划(No.2016YFC0100300);国家自然科学基金资助项目(No.81471631,81771810,51706178);2011年教育部“新世纪优秀人才支持计划”(No.NCET-11-0438);西安交通大学第一附属医院临床研究课题重点项目(No.XJTU1AF-CRF-2015-004)资助~~

摘　　要：目的探讨T2WI机器学习鉴别高级别胶质瘤和脑单发脑转移瘤的价值。方法收集我院2016年1月至2018年11月经病理或随访证实为高级别胶质瘤(41例)或脑单发转移瘤(34例)的患者。所有患者术前均行常规MRI检查,其中包含轴面T2WI。采用ITK-SNAP软件在T2WI上对全瘤进行逐层手动勾画水肿范围感兴趣区(region of interest,ROI),病灶最上和最下两层除外。通过Python对每层ROI进行纹理和形态学特征提取。特征分类统计方法采用支持向量机(support vector machine,SVM)、Logistic回归和朴素贝叶斯机器学习算法,其中70%数据作为训练集,30%作为验证集。同时邀请2名具有3年以上影像诊断经验的医师对病灶水肿区/对侧正常脑组织信号比(nSI)进行半定量受试者工作特征曲线(receiver operating characteristic,ROC)分析,结果均以曲线下面积(area under the curve,AUC)、敏感性和特异度表示,并比较不同方法的诊断效能。结果基于T2WI图像特征机器学习分析,SVM算法具有较高的诊断效能,训练集和测试集AUC分别为0.79(95%confidence interval,CI:0.75~0.83)和0.71(95%CI:0.64~0.77),敏感性和特异度分别为70.88%(95%CI:65.0%~76.3%)和77.13%(95%CI:70.5%~82.9%);Logistic回归和朴素贝叶斯诊断效能较低,训练集和测试集AUC分别为0.77(95%CI:0.73~0.81)/0.67(95%CI:0.60~0.73),0.72(95%CI:0.67~0.76)/0.62(95%CI:0.55~0.69),相比SVM差异有统计学意义;医师半定量分析诊断效能最低,AUC为0.58(95%CI:0.44~0.70)。结论 T2WI图像特征机器学习在术前鉴别高级别胶质瘤和脑单发转移瘤中具有一定的优势,其中SVM模型最具潜力。Objective To explore the value of T2WI machine learning in identifying high-grade gliomas and solitary brain metastasis.Methods We recruited 41 patients with high-grade glioma and 34 patients with brain single-shot metastasis from January 2016 to November 2018 in our hospital confirmed by pathology or follow-up examination.All the patients underwent routine MRI before surgery, including axial T2WI.The ITK-SNAP software was used to manually delineate the region of interest (ROI) of whole tumor in the edema range on T2WI, which did not include the top and bottom images of the lesion.The texture and morphological features of each layer of ROI were achieved through Python software.The feature classification statistical method used the support vector machine (SVM), Naive Bayes and logistic regression machine learning;70% of the data were used as training set and 30% as verification set.At the same time, two resident radiologists with more than 3 years of clinical experience were invited to perform semiquantification receiver operating characteristic (ROC) analysis based on the edema area/contralateral normal area signal ratio (nSI).The area under the curve (AUC), sensitivity and specificity were expressed, and the diagnostic performance of different methods was compared.Results Based on the T2WI image features machine learning analysis, the SVM algorithm had higher diagnostic performance.The AUCs of training set and test set were 0.79 (95% CI :0.75-0.83) and 0.71 (95% CI :0.64-0.77), respectively.The sensitivity and specificity were 70.88%(95% CI :65.0%-76.3%) and 77.13%(95% CI :70.5%-82.9%), respectively.Logistic regression and Naive Bayes had lower diagnostic efficacy, the AUCs of logistic regression and Naive Bayes training and test set were 0.77 (95% CI :0.73-0.81)/0.67 (95% CI :0.60-0.73) and 0.72 (95% CI :0.67- 0.76 )/0.62 (95% CI :0.55-0.69), respectively, which showed significant difference compared to SVM model.The doctor s semiquantification analysis showed the lowest diagnostic performance and the AUC was 0

关 键 词：T2WI 机器学习 胶质瘤 转移瘤 

分 类 号：R73-3[医药卫生—肿瘤]