基于影像组学特征构建机器学习模型鉴别原发性中枢神经系统淋巴瘤与胶质母细胞瘤  

Construction of a machine learning model based on radiomics features for differentiating between primary central nervous system lymphoma and glioblastoma

作　　者：郑成君 刘汉杰 丁一鸣[1] 何磊 宋欣宇 潘宇初 安松培 于书卿[1] Zheng Chengjun;Liu Hanjie;Ding Yiming;He Lei;Song Xinyu;Pan Yuchu;An Songpei;Yu Shuqing(Neurosurgery Center,Beijing Tiantan Hospital,Capital Medical University,Beijing 100070,China)

机构地区：[1]首都医科大学附属北京天坛医院神经外科学中心,北京100070

出　　处：《中华神经外科杂志》2023年第1期30-34,共5页Chinese Journal of Neurosurgery

摘　　要：目的通过基于磁共振T1加权成像(T1WI)增强序列提取的影像组学特征,结合机器学习方法构建预测原发性中枢神经系统淋巴瘤(PCNSL)与胶质母细胞瘤(GBM)的模型并验证。方法回顾性收集首都医科大学附属北京天坛医院神经外科学中心2020年1月至2021年12月收治的120例经病理学确诊的PCNSL和GBM患者的临床及影像学资料,并按7∶3的比例随机分为训练组(84例)和测试组(36例)。使用3D-Slicer软件在患者术前T1WI增强序列上勾画肿瘤增强边界,使用Python软件中"Pyradiomics"包提取训练组的影像组学特征数据,使用独立样本t检验及LASSO回归筛选出训练组中鉴别两种肿瘤最佳的影像组学特征;通过随机森林分类器构建基于影像组学特征的诊断模型,使用测试组数据及5折交叉验证方法进行验证。绘制受试者工作特征(ROC)曲线,通过曲线下面积(AUC)评估模型的预测效能。结果共提取影像组学特征1218个,筛选出3个具有鉴别意义的影像组学特征,分别为原始一阶特征的第十百分位特征、指数灰度依赖矩阵的依赖方差特征、平方根灰度依赖矩阵的非均匀依赖性方差。使用影像组学特征构建的随机森林模型,在测试组中预测PCNSL与GBM的AUC为0.874,灵敏度、特异度分别为0.878、0.684,5折交叉验证的AUC分别为0.870、0.881、0.871、0.855、0.898,平均为0.870。结论基于影像组学特征构建的机器学习模型鉴别PCNSL与GBM的准确率较高。Objective To develop and validate a machine learning model based on radiomics features extracted from contrast enhancement-T1 weighted imaging(CE-T1WI)of MR for differentiating between primary central nervous system lymphoma(PCNSL)and glioblastoma(GBM).Methods The clinical and imaging data of 120 patients with pathologically confirmed PCNSL or GBM admitted to the Neurosurgery Center,Beijing Tiantan Hospital,Capital Medical University from January 2020 to December 2021 were collected retrospectively.The patients were randomly divided into the training group(n=84)and test group(n=36).The tumor boundary was delineated on the preoperative CE-T1WI MR images with the 3D-Slicer software and the radiomics features were extracted using the Pyradiomics package written in Python.The t-test and LASSO regression were used to screen the best radiomics features to distinguish PCNSL and GBM in the training group and the selected features were used to construct the diagnostic model by random forest classifier.The accuracy of the model′s prediction was evaluated by drawing the receiver operating characteristic(ROC)curve and measured using the area under the ROC curve(AUC).Results A total of 1218 imaging features were extracted by Pyradiomics package of Python software,and 3 radiomics features were screened by t-test and LASSO regression.The area under the receiver operating characteristic curve(area under the curve,AUC)of the diagnostic model constructed by random forest classifier was 0.874,and the sensitivity and specificity were 0.878 and 0.684 respectively.The AUC results of 5-fold cross-validation were 0.870,0.881,0.871,0.855 and 0.898(average value:0.870).Conclusion The machine learning model constructed based on radiomics features has a high accuracy in distinguishing PCNSL and GBM.

关 键 词：中枢神经系统淋巴瘤 胶质母细胞瘤 影像组学 机器学习 

分 类 号：R739.4[医药卫生—肿瘤]