基于MRI影像组学的随机森林模型预测子宫内膜癌肌层浸润深度的初步研究  

Predictive value of random forest based on MRI radiomics in evaluating the invasion depth of endometrial carcinoma

作　　者：郭冉 沈秀芝[1] 信瑞强[1] 石清磊 王金洁 钟佳利[1] 彭如臣[1] GUO Ran;SHEN Xiuzhi;XIN Ruiqiang;SHI Qinglei;WANG Jinjie;ZHONG Jiali;PENG Ruchen(Department of Radiology,Beijing Luhe Hospital,Capital Medical University,Beijing 101149,China;MR Scientific,Siemens Healthcare,Beijing 100102,China;Machine Learning and Data Mining Lab.Software Colledge,Shandong University,Ji’nan 250101,China)

机构地区：[1]首都医科大学附属北京潞河医院放射科,北京101149 [2]西门子医疗系统有限公司磁共振事业部,北京100102 [3]山东大学齐鲁软件学院机器学习与数据挖掘实验室,济南250101

出　　处：《磁共振成像》2020年第12期1133-1137,共5页Chinese Journal of Magnetic Resonance Imaging

基　　金：北京市通州区科技计划项目(编号:KJ2020CX004-19)。

摘　　要：目的基于MRI平扫T2WI和增强T1WI的影像组学特征值,探讨机器学习模型随机森林(random forest,RF)对子宫内膜癌肌层浸润深度预测价值。材料与方法回顾性分析行盆腔MRI平扫及增强检查并经手术病理证实为子宫内膜癌患者的影像资料114例(ⅠA期86例,ⅠB期28例),以4∶1的比例通过分层抽样的方法分为训练集和测试集。采用ITK-SNAP软件分别在矢状面平扫T2WI图像及多期增强T1WI图像第二时相进行手动逐层勾画ROI,分别对T2WI和增强T1WI数据集进行影像组学特征值提取(https://github.com/Radiomics/pyradiomics),并对随机森林模型进行训练和测试(http://scikit-learn.org/),采用ROC曲线评价预测效能。结果基于平扫T2WI图像特征值建立的RF模型预测子宫内膜癌肌层浸润深度在测试集的曲线下面积(AUC)为0.938,其准确度、敏感度、特异度分别为91.3%、87.5%、93.3%,模型中重要性排名前3位的特征分别为形状平坦度(shape flatness,SF)、灰度级带矩阵区域方差(GLSZMzonevariance,GLSZM-ZV)、灰度级长矩阵运行方差(GLRLM run variance,GLRLM-RV);基于增强T1WI图像建立的RF模型在测试集的AUC为0.818,准确度、敏感度、特异度分别为81.8%、100%、75.0%,模型中重要性排名前3位的特征分别为SF、灰度相关矩阵高灰度依赖程度(GLDM large dependencehighgraylevelemphasis,GLDM-LDHGLE)、灰度共生矩阵相关性(GLCM correlation)。结论基于MRI影像组学的随机森林模型在预测子宫内膜癌肌层浸润深度中具有较大应用潜力,其中基于平扫T2WI图像建立模型较增强T1WI显示出更大的诊断价值。Objective:To explore the predictive value of random forest based on MRI plain T2 WI and contrast-enhanced T1 WI radiomics in evaluating the invasion depth of endometrial carcinoma.Materials and Methods:We retrospectively analyzed one hundred and fourteen(eighty-six cases of stageⅠA and twenty-eight cases of stageⅠB)patients with endometrial carcinoma confirmed by surgical pathology and all patients underwent pelvic MRI plain and contrast-enhanced examination.All MRI data were divided into training and testing set by stratified sampling method with the ratio of 4∶1.The ITK-SNAP software was used to manually delineate the region of interest layer by layer on the sagittal T2 WI images and the second phase of the multi-phase T1 WI contrast-enhanced images.The radiomics features were extracted based on an open soured tool named pyradiomics(https://github.com/Radiomics/pyradiomics),and the model was established based on scikit-learn(https://www.sklearn.org/).Predictive performance was evaluated by the receiver operating characteristics(ROC)curve.Results:In the testing set,the area under the curve(AUC)of the RF model based on the plain T2 WI images predicting the depth of myometrial invasion for endometrial carcinoma was 0.938,and the accuracy,sensitivity and specificity were 91.3%,87.5%,and 93.3%,respectively.The top three most important features of the model were shape flatness,GLSZM zone variance,and GLRLM run variance;The AUC of the RF model based on contrastenhanced T1 WI images was 0.818,the accuracy,sensitivity and specificity were 81.8%,100%,and 75.0%,respectively.The top three most important features of the model were shape flatness,GLDM large dependence high gray level emphasis,and GLCM correlation.Conclusions:The algorithm of random forest based on MRI radiomics demonstrated great potential in predicting the invasion depth of endometrial carcinoma,and the model based on T2 WI images demonstrated more diagnostic value than that contrast-enhanced T1 WI images.

关 键 词：子宫内膜肿瘤 肌层浸润 影像组学 磁共振成像 

分 类 号：R445.2[医药卫生—影像医学与核医学] R737.33[医药卫生—诊断学]