Title : Revolutionizing acute stroke care: AI-assisted imaging in CT and MRI
Abstract:
Background: Artificial intelligence (AI)-derived software technologies have rapidly evolved to assist in the review and interpretation of neuroimaging, particularly computed tomography (CT) and magnetic resonance imaging (MRI), for patients with suspected stroke. These tools aim to enhance the detection of large and medium vessel occlusions (LVO, MeVO), optimize triage, and improve treatment decisions such as mechanical thrombectomy.
Objectives: This review aims to evaluate the clinical performance, diagnostic accuracy, and cost-effectiveness of AI-assisted imaging software for acute stroke detection, with a focus on CT angiography and MRI in both research and clinical settings. Additionally, it highlights gaps in current evidence and future directions.
Methods: A systematic review and meta-analysis of studies published up to 2025 were conducted, covering AI applications in stroke detection using CT angiography and MRI modalities. Twenty-five databases and multiple clinical trials were reviewed, with a particular emphasis on sensitivity, specificity, and diagnostic accuracy of major AI platforms including RapidAI, Viz.ai, Brainomix, and Avicenna CINA. Economic models assessed the cost-effectiveness of AI integration in stroke care pathways. Bias and reporting quality were evaluated using QUADAS-2 and MI-CLAIM checklists.
Results: AI technologies demonstrated high sensitivity (83.8% to 98.1%) and specificity (79.4% to 98.2%) in detecting large vessel occlusions, with RapidAI showing superior performance in identifying medium vessel occlusions (93% sensitivity) compared to Viz.ai (70%). MRI-based AI tools also demonstrated high accuracy in detecting ischemic lesions with pooled sensitivity and specificity around 93%. Cost-effectiveness analyses suggest that AI-assisted imaging review may be cost-effective if it improves the sensitivity of occlusion detection, thereby enabling timely mechanical thrombectomy. However, current evidence is limited by heterogeneity in study design and lack of large-scale, prospective clinical validation.
Conclusions: AI-driven imaging tools hold significant promise for improving stroke diagnosis, enabling rapid detection of both large and medium vessel occlusions, and supporting clinical decision-making. Despite encouraging diagnostic performance, further multicenter studies are needed to validate clinical effectiveness, standardize performance metrics, and clarify the cost-benefit impact of these technologies in routine clinical practice.
Future Directions: Ongoing research should focus on large prospective trials evaluating AI integration into clinical workflows, expansion of AI capabilities for hemorrhagic stroke detection, and addressing ethical considerations surrounding AI use in neuroimaging.
Keywords: Artificial Intelligence; Stroke; Neuroimaging
