Abstract
OBJECTIVES: To harness the U(2)-Net deep learning framework for automated quantification of collateral circulation in acute ischemic stroke (AIS) via computed tomography angiography (CTA) images, comparing its performance against traditional visual collateral scores (vCS). METHODS: A cohort of 118 confirmed AIS cases was assembled and stratified into 94 development and 24 test cases. CTA images underwent preprocessing and annotation. The U(2)-Net was trained to segment collateral vessels, yielding a quantitative collateral score (qCS) based on vessel volume ratios between affected and healthy hemispheres. Performance was assessed via Dice Similarity Coefficient (DSC), Spearman correlation, Intraclass Correlation Coefficient (ICC), and accuracy, with comparisons to vCS (Tan and Menon score) and ground truth. RESULT: The U(2)-Net demonstrated robust segmentation capabilities, achieving a mean DSC of 0.75 in the test set. The qCS showed a strong correlation with vCS with ρ ranging from 0.78 to 0.92. When compared to the more refined six-class Menon score, the qCS exhibited stronger consistency (development set: ICC = 0.83, test set: ICC = 0.93) than when compared to the four-class Tan score (development set: ICC = 0.76, test set: ICC = 0.79). In terms of classification accuracy, the AI model achieved 0.83 and 0.71 against ground truth and vCS, respectively, for four-class classification. This accuracy escalated to 0.88 and 0.83 for binary classification, emphasizing its proficiency in differentiating collateral status. CONCLUSION: Our U(2)-Net AI model offers a reliable, objective tool for quantifying collateral circulation in AIS. The qCS aligns well with vCS and demonstrates the feasibility of automated collateral assessment, which may enhance diagnostic accuracy and therapeutic decision-making.