影像组学模型对高血压脑出血早期血肿扩大的预测作用研究

Role of radiomics model in prediction of hematoma enlargement in early stage of hypertensive intracerebral hemorrhage

目的 构建一个预测高血压脑出血早期血肿扩大的影像组学模型并探讨其预测价值。 方法 对北京潞河医院神经外科自2010年2月至2018年8月收治的发病6 h内的212例高血压脑出血患者于入院后0.5 h内行头颅CT检查,于入院后24 h内行头颅CT复查,依据血肿体积差异判断有无血肿扩大。在首次CT资料上勾画感兴趣区域,应用Matlab软件从中提取431个影像学特征,通过最小绝对收缩与选择算子(LASSO)回归模型筛选出预测效果最强的影像学特征,进一步用所选特征和支持向量机分类器(SVM)构建预测模型。使用受试者工作特征曲线(ROC)评价预测模型的预测效果。 结果 头颅CT复查发现血肿扩大发生率为18.9%(40/212)。通过LASSO回归模型筛选出18个影像学特征[图像灰度基本特征4个(标准差、峰度、能量、方差),图像形状和体积特征1个(表面和体积比),纹理类特征7个(长行程低灰度优势、惯性、90°相关性、短行程优势、全角相关性、长行程优势、逆差距),小波特征6个(自相关_3、相关信息测度2_3、长行程高灰度优势_4、短行程高灰度优势_4、短行程低灰度优势_7、总变异_3)],并结合SVM构建了预测模型。预测模型的ROC曲线下面积为0.928,敏感性和特异性分别为92.5%、83.5%。 结论 构建的影像组学模型有助于对高血压脑出血早期血肿扩大进行预测。

Objective To construct a radiomics model for predicting hematoma enlargement in early hypertensive intracerebral hemorrhage and explore its predictive value. Methods A retrospective collection of 212 patients with hypertensive intracerebral hemorrhage within 6 h of onset, admitted to our hospital from February 2010 to August 2018, was performed. CT examination was performed within half an hour of admission. CT re-examination was performed 24 h after admission to determine whether there was hematoma enlargement. The regions of interest were delineated on the first CT, and 431 image indicators were extracted from the Matlab software. The LASSO regression model was used to screen out the most predictive imaging features, and the selected features and support vector machine classifier (SVM) were used to build the prediction model. Receiver operating characteristic (ROC) curve was used to evaluate the predicted effect of the model. Results After 24 h of admission, the incidence of hematoma enlargement was 18.9% (40/212). Eighteen imaging ensemble features (including 4 first-order statistics features: standard deviation, kurtosis, uniformity, and variance;one shape- and size-based feature: surface to volume ratio;7 textual features: long run low grey level emphasis, inertia, correlation-angle 90, short run emphasis, correlation-all direction, long run emphasis, and inverse difference moment;6 wavelet features: autocorrelation-3, informational measure of correlation2-3, long run high gray level emphasis-4, short run high gray level emphasis-4, short run low gray level emphasis-7, and sum variance-3) were combined with SVM to establish a prediction model by LASSO regression model. The area under ROC curve was 0.928, enjoying sensitivity and specificity of 92.5% and 83.5%, respectively. Conclusion The constructed radiomics model is helpful in predicting the expansion of hypertensive cerebral hemorrhage.