PAVSAT: an automated blood vessel analysis tool using deep learning-based segmentation and image processing

Accurate blood vessel morphology analysis is essential for understanding vascular system development and pathology. Although advanced imaging technologies. have enabled the rapid acquisition of complex vascular images, quantitative analysis methods have not kept pace. Importantly, current methods cannot accurately represent vessels with complex curved structures and variable diameters, and the existing software lacks efficient automation for large datasets.

In this study, we developed PAVSAT (Python-based Auto Vessel Segmentation Analysis Tool), an automated computational system for quantifying vascular structures by integrating deep learning-based segmentation, image processing, and graph theory. The system employs a YOLOv8 architecture trained on confocal immunofluorescence microscopy images of CD31-stained vascular structures, followed by skeletonization to extract the vessel centerlines. A novel branch-point detection algorithm identifies bifurcations by analyzing local connectivity patterns and performs dense, segment-wise vessel diameter profiling through repeated perpendicular sampling along the locally estimated vessel orientation. To mitigate systematic detection errors at tile boundaries, we implemented a boundary-aware overlapping tiling scheme in which images are processed with spatial offsets and outputs merged to ensure that each sinusoid is analyzed under optimal central-tile conditions. Ablation analysis confirmed that this overlapping integration substantially reduced undetected vessels (by 40.5 and 81.0% in two independent samples) with only marginal increases in false-positive detections (2.2 and 6.3%, respectively), indicating a clearly favorable trade-off between sensitivity and specificity. A graph-based representation converts the vascular structures into nodes and edges for network analysis. Validation against expert manual measurements demonstrated detection rates of 91-98% and measurement accuracy of 89-93% within 10 pixels of the manual measurements, with no detectable systematic bias between automated and manual diameters (mean difference < 0.3 pixels, paired t-test and Wilcoxon p > 0.2 for both samples) and strong agreement (Pearson r > 0.94).

PAVSAT successfully identified complex branching patterns, measured the diameters along curved vessels, and generated quantitative data suitable for large-scale studies, thereby facilitating vascular biology, development, and pathology research.