To construct a radiomics model for predicting hematoma enlargement in early hypertensive intracerebral hemorrhage and explore its predictive value.

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.

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.

The constructed radiomics model is helpful in predicting the expansion of hypertensive cerebral hemorrhage.

Hypertensive cerebral hemorrhage; Hematoma enlargement; Radiomics; Prediction model