Accuracy of artificial intelligence software for CT angiography in stroke

Area of Science:

• Neuroimaging
• Artificial Intelligence in Medicine
• Stroke Diagnostics

Background:

• CT angiography (CTA) is vital for acute ischemic stroke assessment.
• AI-powered software aims to automate the identification of arterial occlusion and collateral vessel scoring.
• Independent validation of AI tools is essential for clinical adoption.

Purpose of the Study:

• To assess the diagnostic accuracy of e-CTA software for identifying arterial abnormalities in acute ischemic stroke.
• To compare AI-driven e-CTA results against expert human interpretation.
• To evaluate software performance across different arterial territories.

Main Methods:

• A large sample of baseline CTA scans from 6 acute stroke studies was analyzed.
• e-CTA software results were compared to masked expert interpretations.
• Diagnostic accuracy was assessed for identifying any arterial abnormality and specifically anterior circulation occlusions.

Main Results:

• The study included 668 patients with acute stroke symptoms.
• e-CTA software achieved a diagnostic accuracy of 72% for detecting arterial abnormalities, with a sensitivity and specificity of 72%.
• Excluding non-anterior circulation occlusions improved accuracy to 76% in a sensitivity analysis.

Conclusions:

• The diagnostic accuracy of e-CTA software for acute arterial abnormalities in stroke patients ranges from 72% to 76% compared to expert readers.
• While promising, the current accuracy necessitates that users maintain CTA interpretation competence.
• Ensuring all potential thrombectomy candidates are identified requires skilled oversight of AI-assisted interpretations.