End to end stroke triage using cerebrovascular morphology and machine learning

Area of Science:

• Artificial intelligence in medical imaging
• Machine learning for neurological diagnostics
• Cerebrovascular disease research

Background:

• Accurate triage of acute ischemic stroke (AIS) is crucial for timely revascularization and better patient outcomes.
• Individual cerebrovascular anatomy significantly impacts response to reperfusion therapies.
• Current methods often lack automation and comprehensive analysis of vascular features.

Purpose of the Study:

• To develop an end-to-end machine learning approach for automatic stroke triage.
• To automatically extract cerebrovascular features for occlusion detection and collateral circulation assessment.
• To predict 90-day functional outcomes using extracted vascular and clinical data.

Main Methods:

• Utilized a convolutional neural network (CNN) for cerebrovasculature segmentation from non-invasive angiography.
• Developed algorithms for automatic detection of occlusion presence, site, and collateral circulation grading.
• Integrated cerebrovascular features with clinical and imaging data for outcome prediction.

Main Results:

• CNN segmentation achieved 94% accuracy (Dice similarity coefficient).
• Automatic stroke detection showed 92% sensitivity and 94% specificity.
• Occlusion site detection and collateral grading reached 96% and 87.2% accuracy, respectively.
• Outcome prediction accuracy improved from 0.63 to 0.83 with the inclusion of automated cerebrovascular features.

Conclusions:

• The presented AI model offers fast, automated, and comprehensive stroke diagnosis.
• It aids in collateral assessment and enhances prognostication for treatment decisions.
• Leveraging cerebrovascular morphology improves the accuracy of stroke triage and outcome prediction.