Markov dependence tree-based segmentation of deep brain structures

Abstract

We propose a new framework for multi-object segmentation of deep brain structures, which have significant shape variations and relatively small sizes in medical brain images. In the images, the structure boundaries may be blurry or even missing, and the surrounding background is a clutter and full of irrelevant edges. We suggest a template-based framework, which fuses the information of edge features, region statistics and inter-structure constraints to detect and locate all the targeted brain structures such that manual initialization is unnecessary. The multi-object template is organized in the form of a hierarchical Markov dependence tree. It makes the matching of multiple objects efficient. Our approach needs only one example as training data and alleviates the demand of a large training set. The obtained segmentation results on real data are encouraging and the proposed method enjoys several important advantages over existing methods. © 2008 Springer Berlin Heidelberg.