基于增强T1加权图像构建不同模型鉴别胶质母细胞瘤与单发脑转移瘤

Application of enhanced T1-weighted images to distinguish glioblastoma and solitary brain metastases

关键词:  机器学习  影像组学  胶质母细胞瘤  脑转移瘤  列线图

Key words:  machine learning  radiomic  glioblastoma  brain metastases  nomogram

摘要:

目的 探讨基于增强T1加权图像构建的不同模型鉴别胶质母细胞瘤（GBM）与单发脑转移瘤（SBM）的应用效果。方法 227例患者（GBM 120例和SBM 107例）随机分为训练集（n=159）与验证集（n=68）。将增强T1加权图像导入3D-Slicer软件并手动勾画感兴趣区（ROI），然后进行影像组学特征提取并使用t检验、递归消除特征法及最小绝对收缩选择算子筛选特征。基于最佳特征，分别运用逻辑回归（LR）、支持向量机（SVM）及随机森林（RF）算法建立模型（影像组学模型、临床模型、临床-影像组学模型），通过绘制受试者工作特征曲线（ROC）及曲线下面积（AUC）delong检验评价各模型鉴别GBM与SBM诊断效能。结果 GBM与SBM病灶是否出血坏死及病灶分布情况（幕上或幕下）的比较差异有统计学意义（P<0.05）。根据组内相关系数（>0.75）初步获取965个稳定性影像特征，最终筛选得到5个最优影像特征构建模型，影像组学模型和临床-影像组学模型AUC均大于0.8，其中LR模型鉴别效能最优（P<0.05）。决策曲线分析显示在整个阈值概率范围内，影像组学模型和临床-影像组学模型有相当的总体净效益。结论 基于增强后T1加权图像构建的影像组学模型和临床-影像组学模型对GBM与SBM具有较高的鉴别诊断效能。

Abstract:

Objective To investigate the application effect of different models for distinguishing glioblastoma (GBM) from solitary brain metastases (SBM) based on enhanced T1-weighted images. Methods A total of 227 patients were included (107 cases of solitary brain metastasis) and were randomly divided into a training set (n=159) and a validation set (n=68) in a 7:3 ratio. The enhanced T1-weighted images were imported into the 3D-Slicer software, and the regions of interest (ROI) were manually delineated. Then, radiomics features were extracted and feature selection was performed using t-tests, recursive feature elimination, and least absolute shrinkage and selection operator. Based on the optimal features, logistic regression (LR), support vector machines (SVM), and random forest (RF) algorithms were used to establish models. The diagnostic performance of the models in distinguishing between glioblastoma and solitary brain metastasis was evaluated using receiver operating characteristic (ROC) curves and the AUC Delong test, and nomogram were drawn. Results There were significant differences in the presence of hemorrhage or necrosis and location above or below the tentorium (P<0.05). Based on the intraclass correlation coefficient (ICC) (ICC > 0.75) within the group, 966 stable imaging features were initially obtained, and 5 optimal imaging features were finally selected to construct the models. The area under the ROC curve (AUC) of the radiomics model and the clinical-radiomics model were both greater than 0.8. The logistic regression model had the best diagnostic performance, and a nomogram was drawn to visualize the model. Decision curve analysis showed that the radiomics model and the clinical-radiomics model had comparable overall net benefits across the entire range of threshold probabilities. Conclusion The radiomics model based on enhanced T1-weighted images are high diagnostic efficiency in distinguishing GBM from SBM.