Combining edge, region, and shape information to segment the left ventricle in cardiac MR images

Abstract

This paper describes a segmentation technique to automatically extract the myocardium in 4D cardiac MR images for quantitative cardiac analysis and the diagnosis of patients. An approximate outline of the left ventricle is obtained either from automatic localization based on the maximum discrimination method or from copying a template shape during propagation. The histogram of the image is analyzed and divided into peaks using the EM algorithm to produce a region-based segmentation. This result and the image gradient are combined to obtain candidate boundaries for the left ventricle by deforming the contour using a graph search active contour approach. The final boundary is chosen using a minimum cut graph algorithm, spline fitting, or point pattern matching to maintain the shape of the template. We have experimented with the proposed method on a large number of patients and present some quantitative and qualitative results.