Feature optimization via simulated search for model-based heart segmentation

Medical images contain numerous different structures and organs. When segmenting a complex organ such as the heart, it is a typical problem of deformable models and related techniques that the image features defining the wanted object boundaries are not discriminative enough to distinguish between borders of different sub-organs. As a result, the model is often locally attracted by some wrong structures. We present a method to automatically learn from a representative set of 3D images which features are most appropriate at each position of the surface of the deformable model. The basic idea is to simulate the boundary detection for the given 3D images and to select those features that minimize the distance between the detected position and the desired object boundary. In addition, we present examples for heart segmentation showing that attraction by false edges is almost completely eliminated.