Graph-Based Classification of Penetrating Cortical Arteries and Veins in Whole-Mouse Vasculature Datasets

Abstract

Advances in whole-brain imaging have produced large-scale vascular datasets, yet most lack artery-vein annotations. However, this is essential for quantitative analysis of whole brain vascular structure, blood flow simulation and translational cerebrovascular research. For unlabelled datasets, manual labelling is infeasible at network scale, and surface-vessel tracing is hampered by reconstruction artefacts, limiting the insights that can be gained from these datasets. Here, we focus on penetrating vessel trees, i.e. descending arteries (A) and ascending veins (V). A supervised Random Forest classifier is trained on graph-derived morphological and topological tree features, including persistent homology. Training uses datasets from two imaging modalities: Light sheet (LS) microscopy with arterial staining and two-photon (TP) micrscopy with manual labelling in smaller samples. Classification reached 96% balanced accuracy on independent TP dataset and 93% on LS datasets. Topological features alone achieved competitive accuracy without relying on diameter measurements, demonstrating cross-modality portability. Arteries exhibit larger diameters and longer inter-bifurcation segments, likely facilitating their functional task of directed blood delivery. The veins show greater branching complexity, allowing low-pressure drainage. These signatures are stable across modalities, supporting their physological relevance. The framework enables A-V label transfer to unlabelled whole-brain TP graph datasets, validated through the V:A ratio as a function of middle cerebral artery (MCA) distance. Operating on graph representations derived from any volumetric imaging modality and requiring no retraining, the approach is applicable to LS, TP, and likely other imaging platforms, providing a scalable foundation for A-V quantification, vascular topological analysis, and haemodynamic modelling.