摘要
目的 探讨基于MRI表现特征结合机器学习算法在预测弥漫性较低级别胶质瘤(1p/19q)缺失状态的价值.方法 选取经手术病理证实为Ⅱ～Ⅲ级胶质瘤 79 例[异柠檬酸脱氢酶(IDH)突变伴 1p/19q共缺失组 39 例,IDH突变伴1p/19q非共缺失组 40 例)],所有患者术前均行常规头颅MRI平扫及增强(T1WI、T2WI、SWI、FLAIR、DWI、CE-T1WI),由未知病理结果的神经影像医师提取影像特征:钙化或出血、T2-FLAIR错配征、瘤周水肿、强化程度、T2 异质性、皮质受累、边界规则、中线偏倚.利用卡方检验或Fisher精确检验评估两组胶质瘤影像学特征的统计学差异,并构建逻辑回归模型;另外,利用所提取的MRI特征构建出机器学习模型,使用受试者工作特征曲线(ROC)分析其预测 1p/19q缺失状态的诊断效能.结果 钙化或出血、T2-FLAIR错配征、T2 异质性三个影像特征在不同 1p/19q缺失状态中比较,差异有统计学意义(P<0.05),联合以上三个影像特征的逻辑回归模型的曲线下面积(AUC)可达 0.859;另外,利用MRI特征所构建的机器学习模型鲁棒性较佳,测试集AUC可高达 0.910.结论 术前MRI表现特征结合机器学习算法可用于无创的预测弥漫性较低级别胶质瘤 1p/19q缺失状态.
Abstract
Objective To explore the value of MRI image features combined with machine learning algorithm in predicting 1p/19q deletion in diffuse lower grade gliomas.Methods A retrospective collection of 79 cases of glioma confirmed by surgical pathology as grade Ⅱ～Ⅲ[39 cases of isocitrated dehydrogenase(IDH)mutation with 1p/19q co-deletion,40 cases of IDH mutation with 1p/19q non-co-deletion]was conducted and all patients underwent conventional head MRI scan and enhance-ment(T1WI,T2WI,SWI,FLAIR,DWI,CE-T1WI),and imaging features were extracted by neuroimaging radiologist with un-known pathological results as follows:calcification or hemorrhage,T2-FLAIR mismatch,peritumoral edema,degree of enhance-ment,T2 heterogeneity,cortical involvement,boundary rules,and midline bias.Chi-square test or Fisher's exact test was used to evaluate the statistical differences in imaging features between the two groups of gliomas and logistic regression analysis was performed on the image features.In addition,a machine learning model was constructed using the extracted MRI features,and the receiver operating characteristic(ROC)curve was used to analyze its diagnostic efficacy in predicting 1p/19q deletion.Re-sults The three imaging features of calcification or hemorrhage,T2-FLAIR mismatch sign,and T2 heterogeneity were statisti-cally different in different 1p/19q deletion states(P<0.05),combined with the above three imaging features,the area under the curve(AUC)of the logistic regression model could reach 0.859;in addition,the machine learning model constructed using MRI image features was more robust,and the test set AUC could be as high as 0.910.Conclusion The preoperative MRI features combined with machine learning algorithm can be used to predict 1p/19q deletion states in diffuse lower grade gliomas noninva-sively.
维普
万方数据