Automated Collateral Classification on CT Angiography in Acute Ischemic Stroke: Performance Trends Across...

Area of Science:

• Neurology
• Medical Imaging
• Artificial Intelligence

Background:

• Collateral status is crucial for acute ischemic stroke (AIS) treatment decisions.
• Current visual collateral grading from CT angiography (CTA) is subjective and inconsistent.

Purpose of the Study:

• To develop and evaluate a convolutional neural network (CNN) for automated collateral classification from CTA images.
• To analyze performance trends across various CNN architectures for tailored clinical application.

Main Methods:

• Extracted binarized vascular morphological features from CTA images of 157 AIS patients.
• Developed and systematically analyzed 192 CNN architectures.
• Stratified patients into good (Menon score 3-5) and poor (Menon score 0-2) collateral groups.

Main Results:

• Achieved AUCs ranging from 0.753 to 0.773.
• Identified three model tendencies: sensitivity-oriented (87.18% sensitivity, 65.00% specificity), balanced (72.65% sensitivity, 77.50% specificity), and specificity-oriented (63.25% sensitivity, 85.00% specificity).
• Key determinants of performance directionality include kernel size, filters in the first layer, and number of convolutional layers.

Conclusions:

• Automated collateral classification from CTA using CNNs is feasible and overcomes subjectivity.
• Systematic analysis provides a framework for developing models optimized for specific clinical needs (sensitivity, specificity, or balance).
• Findings support potential integration into AI systems for endovascular thrombectomy patient triage.