Recognition of anatomically relevant objects with binary partition trees

Abstract

In this paper we demonstrate the application of a binary partition tree to the watershed segmentation with graph merging. An adjacency graph is used to represent the regions found in a watershed transform, merging of these regions is required to combine these regions for further processing. Each node in the binary partition tree represents a larger region that results from the merging of two small regions. Starting from the root node, image areas of child nodes can successively be investigated whether they belong to a certain class of objects. In our application we are e.g. interested in finding anatomical objects such as skull, lung, or heart in an X-ray image. The outlined classification strategy considers only a few, relevant region combinations and thus permits the introduction of sophisticated classification rules without compromising overall computation time. The use of rules improves the recognition rate over simpler linear or box-type classifiers.