Model-based strategies for automated segmentation of cardiac magnetic resonance images

Abstract

Segmentation of the left and right ventricles is vital to clinical magnetic resonance imaging studies of cardiac function. A single cardiac examination results in a large amount of image data. Manual analysis by experts is time consuming and also susceptible to intra- and inter-observer variability. This leads to the urgent requirement for efficient image segmentation algorithms to automatically extract clinically relevant parameters. Present segmentation techniques typically require at least some user interaction or editing, and do not deal well with the right ventricle. This thesis presents mathematical model based methods to automatically localize and segment the left and right ventricular endocardium and epicardium in 3D cardiac magnetic resonance data without any user interaction. An efficient initialization algorithm was developed which used a novel temporal Fourier analysis to determine the size, orientation and position of the heart. Quantitative validation on a large dataset containing 330 patients showed that the initialized contours had only ~ 5 pixels (modified Hausdorff distance) error on average in the middle short-axis slices. A model-based graph cuts algorithm was investigated and achieved good results on the midventricular slices, but was not found to be robust on other slices. Instead, automated segmentation of both the left and right ventricular contours was performed using a new framework, called SMPL (Simple Multi-Property Labelled) atlas based registration. This framework was able to integrate boundary, intensity and anatomical information. A comparison of similarity measures showed the sum of squared difference was most appropriate in this context. The method improved the average contour errors of the middle short-axis slices to ~ 1 pixel. The detected contours were then used to update the 3D model using a new feature-based 3D registration method. These techniques were iteratively applied to both short-axis and long-axis slices, resulting in a 3D segmentation of the patient’s heart. This automated model-based method showed a good agreement with expert observers, giving average errors of ~ 1–4 pixels on all slices.