Title : Diagnostic accuracy of AI-based models in stroke detection from neuroimaging: A systematic review and meta-analysis
Abstract:
Introduction: Artificial intelligence (AI) is rapidly transforming diagnostic radiology, particularly in stroke imaging. Accurate and timely stroke detection is critical to reduce morbidity and mortality, and AI models- especially deep learning may enhance diagnostic performance from neuroimaging. This meta-analysis evaluates the pooled diagnostic accuracy of AI-based models in detecting stroke from CT and MRI, analysing sensitivity, specificity and AUC across various algorithm types and imaging modalities.
Methods: A pre-specified protocol was registered on the Open Science Framework (DOI: 10.17605/OSF.IO/FZNQS) and PRISMA guidelines were followed. A comprehensive search was conducted across PubMed, Embase, Scopus, Web of Science and IEEE Xplore to identify English language studies (2018-2025) reporting diagnostic accuracy of AI models for stroke detection on CT or MRI. Studies were included that used imaging to diagnose stroke, were published in peer-reviewed journals or as preprints and reported (or had enough data to calculate) sensitivity and specificity. Risk of bias was assessed using QUADAS-2 tool, which revealed low to moderate risk. A bivariate random effects model was employed to pool sensitivity and specificity. Subgroup analyses were performed for CT vs. MRI and model types (CNNs, 3D CNNs, Random Forest, commercial tools).
Results: A total of 25 studies were included, encompassing diverse AI algorithms and imaging modalities.
Pooled Sensitivity: 0.90 (range 0.80-0.96)
Pooled Specificity: 0.90 (range 0.71-0.98)
AUC values: Frequently >0.90, with highest performance observed for CNNs and 3D CNNs.
Modality-specific findings
CT -based models (especially CTA) demonstrated high sensitivity for large vessel occlusion and hemorrhage detection (0.90-0.95), with AUCs often exceeding 0.92.
MRI -based models (notably DWI and MRA) showed comparable sensitivity (0.87-0.92) and higher specificity (up to 0.98), supporting their utility in early ischemic detection.
Model-specific findings
Convolutional Neural Networks (CNNs) and 3D CNNs were the most widely used and yielded the highest diagnostic performance (AUCs: 0.90-0.95).
Random forest algorithms showed promise in predictive tasks (sensitivity up to 0.96), but were less commonly applied to imaging.
Heterogeneity was assessed using the I2 statistic, which showed moderate to substantial variability across studies (I2= 69.3%) attributed to differences in imaging modalities, AI models and study populations.
Conclusions: AI-based models, particularly CNNs applied to CT and MRI, demonstrate strong diagnostic accuracy for stroke detection, with pooled sensitivity and specificity around 90%. CT excels in acute stroke triage, while MRI offers superior specificity in early ischemia. These findings support further integration of validated AI tools in stroke workflows. A full meta-analysis with subgroup and SROC curve analysis is ongoing and will guide future clinical implementation.
