Stroke is a critical cerebrovascular event and a leading global cause of disability and death. While deep learning models have significantly improved the detection of acute ischemic stroke, current architectures often face challenges with model generalization and precise lesion localization. This systematic review aimed to evaluate deep learning models for stroke detection, highlighting their strengths and limitations. A comprehensive search across Google Scholar, IEEE Xplore, PubMed, and Web of Science identified 880 articles. Following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines, 45 duplicates were removed, leaving 835 unique records for screening. After a thorough full-text review of 110 potential studies, 100 met the inclusion criteria. The analysis revealed five main research trends: automated diagnosis, Alberta Stroke Program Early CT Score (ASPECTS) scoring, detection of large vessel occlusion, outcome prediction, and lesion segmentation. The results underscore the need for external validation and better localization of small infarcts. Therefore, creating specialized models for small infarcts and the curation of large, accurately annotated datasets are recommended to further advance the field.
Adebowale et al. (2026) studied this question.