Performance of an artificial intelligence tool for multi-step acute stroke imaging: A multicenter diagnostic study.

Area of Science:

• Neurology
• Medical Imaging
• Artificial Intelligence

Background:

• Stroke remains a leading cause of death and disability globally.
• Current stroke imaging workflows can be time-consuming, delaying critical treatment decisions.
• Artificial intelligence (AI) offers potential solutions to enhance stroke diagnosis efficiency.

Purpose of the Study:

• To evaluate the combined performance of an AI-based device for stroke imaging.
• To assess the AI tool's accuracy in detecting intracerebral hemorrhage (ICH) and large vessel occlusion (LVO).
• To determine the AI tool's capability in calculating Alberta Stroke Program Early CT Scores (ASPECTS).

Main Methods:

• Retrospective collection of non-contrast CT (NCCT) and computed tomography angiography (CTA) scans from stroke patients.
• Processing of imaging data using the CINA-HEAD AI tool (Avicenna.AI) for ICH, LVO detection, and ASPECTS calculation.
• Comparison of AI results against ground truth evaluations by four expert neuroradiologists.

Main Results:

• The AI tool achieved 94.6% accuracy for ICH detection on NCCT.
• An accuracy of 86.4% was recorded for LVO identification on CTA.
• Region-based ASPECTS analysis yielded 88.6% accuracy, with 80.4% for dichotomized ASPECTS classification (≥ 6).

Conclusions:

• The AI-based stroke imaging tool demonstrates reliable, stepwise performance.
• The AI tool effectively assists in the diagnostic cascade for stroke.
• This technology shows potential for streamlining acute stroke triage and decision-making processes.