Keywords: Graph Deep Learning, Computed Tomography Angiography, Stroke
Abstract: Large vessel occlusions (LVO) typically lead to severe ischemia of brain parenchyma. Identifying such LVOs is thus a crucial objective in stroke diagnosis. As shortening the time to treatment is essential for a good outcome, fast automated detection can be a valuable tool in clinical routine. This can be achieved using deep learning approaches. In a CTA scan, an LVO can be detected as an unexpected interruption in the contrast-enhanced vessel tree. These cerebrovascular trees can be represented as graphs and analyzed using graph deep learning (GDL) methods. Representing the vasculature as a graph instead of a (very sparsely populated) Euclidean volume massively reduces the model input dimensionality, which promotes time and memory efficiency. In this study, we investigate the use of graph deep learning methods for classifying the presence of a large vessel occlusion compared to state-of-the-art image-based methods. Furthermore, the influence of vascular attributes and different graph topologies is investigated. The proposed model achieves performance comparable to the baseline with an accuracy of $0.95$ and an AUC of $0.89$. Compared to the image-based approach, the graph-based approach is ten times faster and requires 80\% less memory.