Model-based Graph Cut Method for Segmentation of the Left Ventricle.

Model-based medical image analysis allows high level information to guide image segmentation. However, most model-based methods rely on evolution methods which may become trapped in local minima. Graph cuts have been proposed for image segmentation problems where the cost of the cut corresponds to an energy function which is then globally minimized. However, it has been difficult to include high level information in the formulation of the graph cut. We have developed a method for integrating model-based a priori information into the graph cut formulation. A 4D model prior of the left ventricle is calculated from an average of historically analyzed cases. This is scaled and rotated to the given case and a 2D spatial prior is calculated for each image. The spatial prior is then combined with pixel intensity data and edge information in the graph cut optimization. Both epicardial and endocardial contours can be found using variations of this procedure. We report results on 11 normal volunteers and 6 patients with heart disease, compared with the results from two experienced observers. A modified Hausdorff distance measure showed good agreement between the model-based graph cut and the expert observers.