机器学习结合影像组学特征鉴别胶质母细胞瘤与脑转移瘤  

Differential diagnosis of glioblastoma and brain metastasis via machine learning combined with radiomic features

作　　者：符礼孔 李泽 孟思 吕佳洋 周超 刘海洲 李志伟[1] FU Likong;LI Ze;MENG Si;LüJiayang;ZHOU Chao;LIU Haizhou;LI Zhiwei(Department of Radiology,Sanya Central Hospital(Hainan Third Peop le’s Hospital),Sanya,Hainan Province 572000,China;Department of Radiolog y,Tianjin Huanhu Hospital,Tianjin 300222,China)

机构地区：[1]三亚中心医院(海南省第三人民医院)放射科,海南三亚572000 [2]天津市环湖医院放射科,天津300222

出　　处：《实用放射学杂志》2023年第5期697-700,共4页Journal of Practical Radiology

摘　　要：目的基于术前影像组学特征建立机器学习模型鉴别胶质母细胞瘤与脑转移瘤。方法回顾性选取105例单发环形强化病灶患者,将患者随机分为训练组(n=58)和验证组(n=47)。所有患者在术前1周内接受MRI检查。使用A.K.软件从训练组的MRI增强扫描图像中提取影像组学特征,运用单因素分析与最小绝对收缩和选择算子(LASSO)算法筛选最佳的影像组学特征,构建机器学习模型并在验证组中检验模型的稳定性,通过绘制受试者工作特征(ROC)曲线评价机器学习模型鉴别胶质母细胞瘤和脑转移瘤的准确性。结果根据组内相关系数(ICC)筛选出698个高稳定性特征(ICC>0.8),在训练组依次行正态检验(Kolmogorov-Smirnov检验)和方差齐性检验(Bartlett检验),最终LASSO选择10个最佳特征用于构建机器学习模型,鉴别单发环形强化的胶质母细胞瘤和脑转移瘤。在训练组和验证组中,ROC曲线分析显示不同机器学习模型诊断准确性均>70%,支持向量机(SVM)模型在敏感性和特异性均优于其他模型。结论基于MR影像组学特征构建的SVM模型可以在术前显著提高鉴别诊断胶质母细胞瘤与脑转移瘤的准确性,具有较高的稳定性。Objective To based on preoperatively differential glioblastoma from brain metastasis via machine learning combined with radiomic features.Methods A total of 105 patients with single ring enhanced lesions were retrospectively selected,and then randomly divided into two groups,including training group(n=58)and validation group(n=47).All patients underwent MRI examination within 1 week before treatment.Radiomic features were extracted from MRI images in the training group by used the A.K.software.Univariate analysis and least absolute shrinkage and selection operator(LASSO)algorithm were used to selected optimal radiomic feature to build machine learning models,and then the stability of the model was tested in the validation group.The accuracy of machine learning models in differential glioblastoma from brain metastasis was evaluated by receiver operating characteristic(ROC)curve.Results A total of 698 highly stable features were selected according to the intraclass correlation coefficient(ICC)(ICC>0.8).In the training group,Kolmogorov-Smirnov test and Bartlett test were performed successively in the training group.Finally,10 best features were selected by LASSO to develop machine learning models in order to differential glioblastoma from brain metastasis.In the training and validation groups,ROC curve analysis showed that the diagnostic accuracy of different machine learning models was all greater than 70%,and support vector machine(SVM)model was superior to other models in sensitivity and specificity.Conclusion SVM model based on MR radiomic features can improve the accuracy of differential diagnosis between glioblastoma and brain metastasis with high stability.

关 键 词：机器学习 影像组学 胶质母细胞瘤 脑转移瘤 磁共振成像 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程] R739.41[自动化与计算机技术—控制科学与工程] R445.2[医药卫生—肿瘤]