Retina features based on vessel graph substructures

We represent the retina vessel pattern as a spatial relational graph, and match features using error-correcting graph matching. We study the distinctiveness of the nodes (branching and crossing points) compared with that of the edges and other substructures (nodes of degree k, paths of length k). On a training set from the VARIA database, we show that as well as nodes, three other types of graph sub structure completely or almost completely separate genuine from imposter comparisons. We show that combining nodes and edges can improve the separation distance. We identify two retina graph statistics, the edge-to-node ratio and the variance of the degree distribution, that have low correlation with node match score.