基于T_(1)WI增强不同机器学习模型鉴别胶质母细胞瘤与原发性中枢神经系统淋巴瘤

Construction of different machine learning models based on T_1WI-enhanced images for differentiating between glioblastoma and primary central nervous system lymphoma

目的:基于T_(1)WI增强图像采用六种不同机器学习分类算法构建预测胶质母细胞瘤(GBM)与原发性中枢神经系统淋巴瘤(PCNSL)的模型,比较不同机器学习模型的诊断效能。方法:回顾性分析中南大学湘雅医院经病理证实的GBM 57例和PCNSL 49例患者的临床及影像资料。应用ITK-SNAP软件在术前T1WI增强图像手动逐层勾画瘤体感兴趣区(ROI)。基于慧医汇影放射组学Radcloud平台进行ROI影像组学特征提取并采用方差阈值法(阈值>0.9)、单变量特征选择法(P<0.01)和最小绝对收缩选择算子(LASSO)进行特征降维,筛选出的特征采用支持向量机、极致梯度提升、逻辑回归(LR)、线性判别分析(LDA)、随机森林、K近邻等6种分类器构建影像组学预测模型。使用5折交叉验证方法进行验证,采用受试者工作特征曲线下面积(AUC)评估6种预测模型的诊断效能,模型之间AUC比较采用DeLong检验。结果:共提取1 688个影像组学特征,经过特征降维及筛选后保留显著特征(5折交叉验证、每组分别25、10、31、17、14个特征)构建预测模型,6种模型中LDA、LR模型诊断效能最佳,在5折交叉验证集中LDA、LR模型平均AUC分别为0.965、0.958,准确度为87.8%、89.6%,敏感度为86.0%、86.0%,特异度为89.4%、93.0%。6种模型AUC差异均无统计学意义(P>0.05)。结论:基于T_(1)WI增强图像影像组学特征构建机器学习模型可用于预测GBM与PCNSL且准确率较高,其中LDA、LR模型诊断效能最佳。

Objective:To construct models for predicting glioblastoma(GBM) and primary central nervous system lymphoma(PCNSL) using six different machine learning classification algorithms based on T_1WI-enhanced images,and to compare the diagnostic efficacy of different machine learning models.Methods:A retrospective analysis was conducted on the clinical and imaging data of 57 patients with pathologically confirmed GBM and 49 patients with PCNSL at Xiangya Hospital of Central South University.The ITK-SNAP software was used to manually outline the tumor region of interest(ROI) layer by layer on preoperative T_1WI-enhanced images.ROI imaging omics features were extracted based on the Radcloud platform,and the variance threshold method(threshold>0.9),univariate feature selection method(P<0.01),and least absolute shrinkage and selection operator(LASSO) were used for feature dimensionality reduction,and the screened features were used to construct an imageomics prediction model using six classifiers including support vector machine,extreme gradient boosting,Logistic regression(LR),linear discriminant analysis(LDA),random forest,and K-nearest neighbor.The 5-fold cross-validation method was used for validation,and the diagnostic performance of six predictive models was evaluated using the area under the curve(AUC),and the DeLong test was used for AUC comparisons between models.Results:A total of 1 688 imaging omics features were extracted,and the significant features(25,10,31,17,and 14 features per group in 5-fold cross-validation,respectively) were retained after feature dimensionality reduction and screening to construct the prediction models,and among the 6 models,the LDA and LR models had the best diagnostic efficacy,and the average AUC of the LDA and LR models in the 5-fold crossvalidation set was 0.965 and 0.958,respectively,with accuracy of 87.8% and 89.6%,sensitivity of 86.0% and 86.0%,and specificity of 89.4% and 93.0%.There was no statistically significant differences in AUC among the six models(P >0.05).Conclusion:Machine learning models based on T_1WI-enhanced image omics features can be used to predict GBM and PCNSL with high accuracy,among which LDA and LR models have the best diagnostic efficacy.