Radiomics-based prediction of symptomatic intracerebral hemorrhage before thrombolysis therapy in unenhanced CT imaging

Symptomatic intracerebral hemorrhage (sICH) is rare but the most devastating complication of thrombolysis therapy for acute ischemic stroke, thus early prediction of sICH is critical. This study aims to predict the probability of sICH before thrombolysis therapy via radiomics analysis based on unenhanced CT images. 300 patients were retrospectively enrolled from three different centers included 1548 slices CT images. 538 radiomics features were extracted from each slice image. Random forest (RF) classifier in combination with minimum redundancy and maximum relevance (mRMR) feature selection were adopted to construct the prediction model. The five significant radiomics features (GLCM-LL-Correlation, GLCM-HH-Homogeneity, Histogram-Kurtosis, shape-PAratio, and NGTDM-HH-Contrast) selected by mRMR are incorporated to construct the final model, resulted in the area under the receiver-operating characteristic curve (AUC-ROC) of 0.74 and 0.71 in training and independent validation cohorts, respectively. Overall, the radiomics analysis on multi-center pre-therapy unenhanced CT images demonstrated potential to the early prediction of sICH in acute ischemic stroke.