Artificial intelligence in acute stroke imaging: A comprehensive survey from imaging assessment to outcome prediction

Acute strokes encompass two main types: acute ischemic stroke (AIS) and intracerebral hemorrhage (ICH). Each type involves specific imaging techniques and entails different clinical considerations. While deep learning has facilitated the analysis of stroke imaging autonomously, the technical differences between these two stroke types have not been examined within a shared analytical framework. This review addresses that gap through the examination of 84 peer-reviewed articles. A PRISMA-compliant search was conducted using multiple databases. Analysis of the ICH pipeline predominantly uses CT imaging and is primarily centered on a classification task. The main focus is on hemorrhage identification, subtype classification, hematoma expansion prediction, and forecasting associated clinical outcomes. The AIS pipeline is primarily based on quantification and therefore uses a variety of imaging techniques to optimize and automate the analysis of a specific set of clinical questions related to the stroke. Core tasks include segmentation of the infarct area, assessment scoring, large vessel occlusion, core-penumbra mismatch analysis, and prediction of outcomes. Convergence is also observed in outcome prediction approaches across both domains. Existing models demonstrate clinically acceptable performance across numerous tasks. Clinical deployment nonetheless remains limited by the lack of external validation, high costs of annotation, demands of real-time processing, and poor prospective evidence. This review offers practical guidance to develop fully operational AI systems that can be integrated into emergency stroke care.