Reconstructing cerebrovascular networks under local physiological constraints by integer programming.

Abstract

We introduce a probabilistic approach to vessel network extraction that enforces physiological constraints on the vessel structure. The method accounts for both image evidence and geometric relationships between vessels by solving an integer program, which is shown to yield the maximum a posteriori (MAP) estimate to a probabilistic model. Starting from an overconnected network, it is pruning vessel stumps and spurious connections by evaluating the local geometry and the global connectivity of the graph. We utilize a high-resolution micro computed tomography (μCT) dataset of a cerebrovascular corrosion cast to obtain a reference network and learn the prior distributions of our probabilistic model and we perform experiments on in-vivo magnetic resonance microangiography (μMRA) images of mouse brains. We finally discuss properties of the networks obtained under different tracking and pruning approaches.