Brain Image Segmentation Based on Hypergraph Modeling

In this paper a new framework for medical image segmentation is presented based on the hypergraph decomposition theory. Each frame of the clinical image atlas is first over-segmented into a series of patches which assigned some clustering attribute values. The patches that satisfy some conditions are chosen to be hypergraph vertices, and those clusters of vertices share some attributes form hyperedges of the hypergraph. The task of extracting objects from the scanned brain images atlas is thus converted to be a hypergraph partition problem. The distributed multilevel partition algorithm is then employed to split the hypergraph into clusters, each of the clusters is assigned a modularity attribute to indicate the compactness of the cluster. Experiment shows that these modularity attributes are generally of large values for those clusters formed by organs such as tumor, which demonstrates the effectiveness of our proposed scheme and algorithm.