Diagnostic performance of an algorithm for automated collateral scoring on computed tomography angiography

Area of Science:

• Neurology
• Medical Imaging
• Artificial Intelligence

Background:

• Collateral circulation is crucial for endovascular treatment outcomes in acute ischemic stroke.
• Accurate assessment of collateral status is vital for patient management.

Purpose of the Study:

• To evaluate the performance of a commercial AI algorithm for collateral scoring (CS) in acute ischemic stroke patients.
• To compare AI-driven CS with expert radiologist assessments.

Main Methods:

• Retrospective analysis of CT angiography (CTA) scans from the MR CLEAN Registry (n=1024).
• Comparison of visual CS (radiologists) with AI-quantified CS (0-100%).
• Assessment of Area Under the Curve (AUC) for discriminating good vs. poor collaterals and predicting functional independence.

Main Results:

• 59% agreement between visual and automated CS.
• AUC of 0.87 for AI in discriminating good vs. poor collaterals.
• AI-based CS showed similar predictive performance for functional independence (AUC 0.66) compared to visual CS (AUC 0.64).

Conclusions:

• Automated CS software performs comparably to expert radiologists in assessing collateral status.
• The AI tool effectively predicts functional independence in stroke patients.
• CTA acquisition timing did not impact the AI algorithm's performance.