基于临床特征及平扫CT影像组学特征模型预测早期脑出血血肿扩大的价值

目的：探讨基于临床特征和平扫CT影像组学特征机器学习模型预测早期脑出血血肿扩大（HE）的价值。方法：收集2018年1月至2020年5月温州医科大学附属第二医院急性早期自发性脑出血患者261例，根据是否存在早期HE分为HE组和非HE组，并将所有样本按7:3随机分为训练集（182例）和验证集（79例）。采用3D Slicer软件对病灶感兴趣区进行勾画。提取影像组学特征并收集患者的临床特征（人口统计学特征、CT影像学特征）。使用最小绝对值收缩和选择算子（LASSO）对影像组学特征进行筛选，保留非零系数特征；采用单因素分析及多因素Logistic回归分析筛选独立危险因素。根据筛选出的特征分别构建临床、影像组学及混合逻辑回归（LR）模型以预测早期脑出血血肿是否扩大。采用受试者工作特征曲线（ROC）及曲线下面积（AUC）对模型的预测效能进行评估。结果：从CT影像中共提取396个影像组学特征，经LASSO算法降维后筛选出7 个具有鉴别意义的影像组学特征。收集临床特征共计10 个，经单因素分析及多因素Logistic回归分析后，发现漩涡征、黑洞征、形状不规则是HE的独立危险因素（P ＜0.05）。影像组学模型、临床模型及混合模型预测效能：训练集中AUC分别为0.924、0.836和0.968，特异度分别为91.4%、81.0%和95.2%，敏感度分别为81.8%、78.4%和84.4%；验证集中AUC分别为0.919、0.796 和0.929，特异度分别为81.8%、77.5%和88.1%，敏感度分别为76.1%、64.5%和80.4%。结论：基于临床及CT影像组学特征构建的LR模型对HE具有一定的预测效能。

The prediction value of early hematoma expansion after intracerebral hemorrhage based on clinical features and noncontrast CT radiomic feature models

Objective: To investigate the predicting value of early hematoma expansion after intracerebral hemorrhage based on clinical and non-contrast CT radiomic feature machine learning models. Methods: A total of 261 cases of acute early spontaneous intracerebral hemorrhage from the Second Affiliated Hospital of WenzhouMedical University from January 2018 to May 2020 were collected. Patients were assigned as hematoma expansion group and non-hematoma expansion group according to the presence of early hematoma expansion. All samples were divided into training set (182) and testing set (79) randomly according to ratio of 7:3. Regions of interest of lesions were delineated by 3D Sicer software. Radiomic features were extracted and clinical features (demographic and CT imaging features) of each patient were collected. Least absolute shrinkage and selection operator (LASSO) was used to select radiomic features. Univariate analysis and multivariate logistic regression analysis were used to select clinically predictive independent risk factors. Logistic regression models for predicting invasiveness of pulmonary adenocarcinoma were established based on clinical features, radiomic features and clinical features combined with radiomic features. Receiver operating characteristic curve andarea under curve were used to evaluate the predictive performance of models. Results: A total of 396 radiomic features were extracted from CT images, of which 7 radiomic features with discriminative significance were selected after dimensionality reduction by least absolute shrinkage and selection operator (LASSO). A total of 10 clinical features were collected, and swirl sign, black hole sign and irregular shape were found to be independent risk factors for predicting HE after univariate analysis and multivariate Logistic regression analysis (P<0.05).Predictive performance of the radiomic feature model, clinical feature model and combined model were as follows: in the training set, AUC was 0.924, 0.836 and 0.968, specificity was 91.4%, 81.0% and 95.2%, sensitivity was 81.8%, 78.4% and 84.4%; in the testing set, AUC was 0.919, 0.796 and 0.929, specificity was 81.8%, 77.5% and 88.1%, sensitivity was 76.1%, 64.5% and 80.4%. Conclusion: Logistic regression model based on clinical and CT radiomic features has certain predictive efficiency for early hematoma expansion after intracerebral hemorrhage.