Artificial Intelligence for Ischemic Stroke Detection in Non-contrast CT: A Systematic Review and Meta-analysis

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Ischemic stroke (IS) diagnosis relies on non-contrast CT (NCCT).
• Artificial intelligence (AI) models are emerging tools for medical image analysis.
• Assessing AI's diagnostic accuracy for IS in NCCT is crucial.

Purpose of the Study:

• To systematically review and meta-analyze the diagnostic accuracy of AI models for IS detection in NCCT.
• To compare the diagnostic performance of AI models against clinicians.
• To evaluate the impact of AI assistance on clinician performance.

Main Methods:

• Systematic literature search conducted until February 2025 across major databases (PubMed, Web of Science, Cochrane, IEEE Xplore, Embase).
• Inclusion of studies using AI on human NCCT images for IS detection/classification.
• Risk of bias assessed using PROBAST; meta-analysis employed pooled sensitivities, specificities, and HSROC curves.

Main Results:

• AI demonstrated pooled sensitivity of 91.2% and specificity of 96.0% in internal validation (74 trials from 32 studies).
• External validation showed lower AI sensitivity (59.8%) but high specificity (97.3%).
• Clinicians' performance improved with AI assistance (sensitivity 83.7%, specificity 86.7%).
• Higher AI sensitivity linked to data augmentation and transfer learning techniques.
• Significant heterogeneity and high risk of bias (58% of studies) were noted.

Conclusions:

• AI exhibits acceptable diagnostic performance for IS detection in NCCT internally.
• Significant heterogeneity impacts meta-analysis findings.
• Limited external validation and generalizability hinder AI's immediate real-world clinical application.