Automated ASPECTS Classification in Acute Ischemic Stroke using EfficientNetV2

Abstract

Etiology of ischemic stroke can be attributed to large vessel occlusions (LVOs), which lead to insufficient supply of oxygen to brain. Early detection and evaluation of infarct core volume plays a crucial role in the optimal treatment for brain ischemia. In this work, we leverage transfer learning for automated computing of Alberta Stroke Program Early CT Score (ASPECTS) using Non-Contrast CT (NCCT) scans. We compare the performance of different state-of-the art ImageNet pre-trained networks for multi-class classification of NCCT scans based on ASPECTS score value ranging from 0-10. Additionally, based on ASPECTS and reperfusion therapy, we group the NCCT scans into two categories: 0-6 in Group0 and 7-10 in Group1, to perform binary classification using pre-trained networks. Our experiments validate the choice of pre-trained EfficientNetV2 with the highest Area Under Curve (AUC) of 0.995 in multi-class classification and 0.977 in binary class classification.