A Radiomics-Based Machine Learning Approach to Assess Collateral Circulation in Ischemic Stroke on Non-contrast Computed Tomography

Assessment of collateral circulation in ischemic stroke, which can identify patients for the most appropriate treatment strategies, is currently conducted with visual inspection by a radiologist. Yet numerous studies have shown that visual inspection suffers from inter and intra-rater variability. We present an automatic evaluation of collaterals using radiomic features and machine learning based on the ASPECTS scoring terminology with non-contrast computed tomography (NCCT). The method includes ASPECTS regions identification, extraction of radiomic features, and classification of collateral scores with support vector machines (SVM). Experiments are performed on a dataset of 64 ischemic stroke patients to classify collateral circulation as good, intermediate, or poor and yield an overall area under the curve (AUC) of 0.86 with an average sensitivity of 80.33% and specificity of 79.33%. Thus, we show the feasibility of using automatic evaluation of collateral circulation using NCCT when compared to the ASPECTS score by radiologists using 4D CT angiography as ground truth.