Separation of Cortical Arteries and Veins Using Intrinsic Optical Signals Extracted by Canonical Correlation Analysis

This paper presents an artery-vein separation method in cerebral cortical image with optical imaging of intrinsic signals. The method utilizes three distinct intrinsic signal sources including low frequency oscillation, respiration and heartbeat, which are extracted from the recorded optical signals by temporal canonical correlation analysis, to reflect the artery-vein difference in temporal domain. Each signal source constructs a correlation-coefficient map to reveal the spatial structure of a specific type of vessel. Low frequency oscillation and heartbeat sources reveal the arterial structure while respiration source reveals the venous structure. Based on the three feature maps, classification of vessel types is achieved by SVM on segmented vessel network. With hand-labeled arteries and veins as the reference standard, the algorithm gives 95.7% true positive rates (TPR) and 7.5% false positive rates (FPR) for the arteries, as well as 92.5% TPR and 4.1% FPR for the veins when tested on ten sets of image sequence. Comparison with previously reported methods demonstrates that this method improves the artery-vein separation performance.