【摘要】 目的 探讨多模态颅脑MRI影像组学特征在预测胶质母细胞瘤（GBM）患者生存期风险分层中的作用。方法 回顾性分析BRATS2018数据库中经手术病理证实的GBM患者163例，按入组时间分为训练组114例，测试组49例。提取所有患者术前MRI图像影像组学特征，评估传统影像视觉特征，然后对训练组数据使用最大相关-最小冗余算法和梯度提升决策树算法进行降维后建立影像组学标签模型，计算患者的影像组学标签分数，最终结合影像视觉特征和临床因素，使用多元逻辑斯回归构建总生存期联合预测模型并绘制列线图。基于测试组数据使用ROC曲线评估模型的诊断效能，并用决策曲线分析验证。结果 从每例患者的4个MRI序列图像、5个感兴趣区共提取纹理特征7 920个，经降维后筛选出26个最优价值特征构建影像组学标签。使用多元逻辑斯回归构建包含了深部白质、年龄和影像组学标签的联合诊断模型，并生成列线图，该模型在训练组和测试组中预测长短生存期的准确率分别为0.848和0.800。列线图、联合影像、影像组学标签、深部白质受累和年龄在所有患者中的诊断准确率分别为0.941、0.908、0.873、0.663和0.655。基于模型区分的高危组与低危组中的GBM患者数差异有统计学意义（P<0.05）。结论 影像组学标签、深部白质和年龄是GBM患者的独立预测因子，基于三者的联合模型而绘制的列线图可用于预测GBM患者总生存期，有助于进行生存风险分层。更多还原

【Abstract】 Objective To explore the application of multimodal brain MRI imaging features in predicting the survival risk of patients with glioblastoma(GBM). Methods Clinical and imaging data of 163 cases of glioblastoma confirmed by surgery and pathology in the BRATS2018 database were retrospectively analyzed. Patients were divided into training group(n=114) and test group(n=49) based on the time of entering the study. The preoperative MRI radiomics features were extracted;the maximum correlation-minimum redundancy(m RMR) and the gradient boosting decision tree(GBDT) algorithm were used to reduce the dimensionality and machine learning was utilized to establish an radiomics signature, which was then used to calculate the rad-score of each patient. A prediction model for overall survival(OS) was constructed by using multivariate logistic regression of image visual features and clinical factors and finally a nomogram was generated. The ROC curve was used to evaluate the diagnostic performance of the model, and it was verified by decision curve analysis(DCA). Results A total of7 920 texture features were extracted from the four MRI sequence images from five regions of interest for each patient. After dimensionality reduction, 26 features with the best value were selected to construct radiomics signature. Multivariate logistic regression was used to construct a joint diagnosis model including deep white matter, age, and radiomics signature and then a nomogram was generated. The accuracy of the model to predict long and short-term OS was 0.848 and 0.800 in the training group and the test group, respectively. The diagnostic accuracy of nomogram, combined imaging, radiomics signature, deep white matter, and age in all patients were 0.941, 0.908, 0.873, 0.663, and 0.655, respectively. There was a statistically significant difference in the number of GBM patients between the high-risk group and the low-risk group based on the model(P<0.05).Conclusion Radiomics signature, deep white matter involvement, and age are independent predictors of OS of glioblastoma patients. The nomogram based on the combined model can be used to predict the OS of glioblastoma patients and stratify the patients accordingly.更多还原