Apparent diffusion coefficient map based radiomics model in identifying the ischemic penumbra in acute ischemic stroke

BACKGROUND Saving the ischemic penumbra (IP) is key in treating acute ischemic stroke (AIS). We aim to investigate the value of the apparent diffusion coefficient (ADC) map based radiomics model in the identification of IP in AIS. METHODS This study retrospectively analyzed the data of 241 patients with AIS involving the anterior cerebral circulation who were treated in our hospital within 24 h of stroke onset from January 2014 to October 2019. With the perfusion-weighted imaging (PWI)/diffusion-weighted imaging (DWI) mismatch model as the gold standard to determine whether IP exists, we divided patients into PWI/DWI mismatch (84 cases) and non-PWI/DWI mismatch (157 cases). Following the DWI high signal area, the region of interest (ROI) was drawn to the maximum level of the lesions on the ADC map, and a total of 896 features were extracted. Maximum correlation and minimum redundancy (mRMR) algorithm were applied to select the optimized features subsets, and then the least absolute shrinkage and selection operator (LASSO) were furtherly applied to select the best features to construct radiomics signature in predicting PWI/DWI mismatch. The performance of the model was evaluated using a receiver operating characteristic (ROC) curve. One hundred times internal cross-validation was applied to evaluate the stability of the model. The clinical value of the model was evaluated using decision curve analysis (DCA). RESULTS Twenty-one features were finally selected to set up the radiomics model. In the training set, the area under the ROC curve (AUC) was 0.92, and the sensitivity, specificity, and accuracy were 0.93, 0.75, 0.82, respectively. In the validation set, the AUC was 0.90, and the sensitivity, specificity, and accuracy were 0.88, 0.74, 0.80, respectively. The average AUC of internal cross-validation for 100 times in the training set were 0.88 and 0.83 in the validation set. DCA shows that within the threshold range of 0.08 to 1.0, the model gains more net benefit. CONCLUSIONS The radiomics model based on the ADC map can effectively determine the presence of IP in patients with AIS.