Cerebrovascular network registration via an efficient attributed graph matching technique.

Registration of vascular networks is an indispensable element of prognostic and diagnostic studies that require structural analysis and comparison over time, among different samples, and to a gold standard. However, vascular networks manifest low spatial texture and sparse structural content so that even small variations in their location can make the intensity-based registration inefficient and prone to errors. Motivated by geometrical graph-based models developed in our prior work, we use the shape information in the graph topology sense to enhance the registration performance. An efficient feature-based registration is presented that seeks correspondence of the bifurcations and branches in a graph matching scheme. Since the graph matching is originally posed a NP-hard quadratic assignment problem (QAP) in the literature, we have designed a node signature that incorporates edge correspondences indirectly. This allows removing the quadratic term in the QAP to recast the problem as a linear assignment problem (LAP) to relieve the computational burden. The LAP is efficiently solvable and is scalable to data with graph representation of larger size. The performance is tested and validated using clinical 3-D angiography images of the human cerebrovasculature as well as synthetic datasets. This method proves to be robust in the face of different structural and algorithm's parameters. Quality of inter-subject and multimodal matching of clinical data has also been confirmed.