Orientation histograms as shape priors for left ventricle segmentation using graph cuts

Abstract

Poor contrast in magnetic resonance images makes cardiac left ventricle (LV) segmentation a very challenging task. We propose a novel graph cut framework using shape priors for segmentation of the LV from dynamic cardiac perfusion images. The shape prior information is obtained from a single image clearly showing the LV. The shape penalty is assigned based on the orientation angles between a pixel and all edge points of the prior shape. We observe that the orientation angles have distinctly different distributions for points inside and outside the LV. To account for shape change due to deformations, pixels near the boundary of the prior shape are allowed to change their labels by appropriate formulation of the penalty and smoothness terms. Experimental results on real patient datasets show our method's superior performance compared to two similar methods. © 2011 Springer-Verlag.