Vessel-Graph Neural Network Detects Leptomeningeal Collaterals on DSA

A team of researchers led by Junyong Cao from the University of Zurich and TU Munich has developed the first AI framework capable of detecting individual leptomeningeal collateral (LMC) vessels directly from digital subtraction angiography (DSA) scans. LMCs are crucial prognostic markers in acute ischemic stroke, but until now automated methods relied on CT angiography (CTA), which cannot resolve individual collaterals. The new vessel-graph neural network approach treats collateral detection as a graph classification problem, where each vessel segment becomes a node in a vascular graph.

The architecture combines a topology-aware graph branch with a dense pixel branch, fusing both in a shared node-probability space. In rigorous five-fold cross-validation, the fused model achieved a PR-AUC of 0.434, significantly outperforming the graph-only (0.403) and pixel-only (0.362) baselines. The graph-pixel fusion allows the model to leverage both structural connectivity and fine-grained image features, enabling precise per-vessel analysis.

This breakthrough moves DSA-based collateral assessment from subjective manual grading (which suffers from poor inter-rater agreement) to objective, AI-driven evaluation. The ability to individualize LMCs opens the door to discovering novel biomarkers and collateral patterns that could improve stroke prognosis and treatment planning. The researchers note this is the first method to enable quantitative, individualized LMC detection on DSA.