Automatic segmentation of cardiac MRI

Abstract

A new method is proposed for automatic segmentation of cardiac MRI. The novelty is that we analyze motion to extract the information that other algorithms rely on users to input manually, such as an initial contour or seed points. The motion map, once computed, restricts subsequent processing to be constrained within a region of interest that envelops the heart. Focusing the computation in this manner resolves some of the ambiguity that causes the image segmentation problem to be an ill-posed one. The segmentation is then performed using a new contextual dependency network (CDN) that incorporates context via hierarchical processing. Within this framework, voxels are first classified as independent events with an EM approach to simultaneously correct for field inhomogeneity. The algorithm next considers neighborhood interactions using a Markov random field, followed by region-level properties, and finally, relationships between regions. © Springer-Verlag Berlin Heidelberg 2003.