Segmentation of 4D cardiac MR images using a probabilistic atlas and the EM algorithm

Abstract

In this paper an automatic atlas-based segmentation algorithm for 4D cardiac MR images is proposed. The algorithm is based on the 4D extension of the expectation maximisation (EM) algorithm. The EM algorithm uses a 4D probabilistic cardiac atlas to estimate the initial model parameters and to integrate a priori information into the classification process. The probabilistic cardiac atlas has been constructed from the manual segmentations of 3D cardiac image sequences of 14 healthy volunteers. It provides space and time-varying probability maps for the left and right ventricles, the myocardium, and background structures such as the liver, stomach, lungs and skin. In addition to using the probabilistic cardiac atlas as a priori information, the segmentation algorithm incorporates spatial and temporal contextual information by using 4D Markov Random Fields. After the classification, the largest connected component of each structure is extracted using a global connectivity filter which improves the results significantly, especially for the myocardium. Validation against manual segmentations and computation of the correlation between manual and automatic segmentation on 249 3D volumes were calculated. We used the 'leave one out' test where the image set to be segmented was not used in the construction of its corresponding atlas. Results show that the procedure can successfully segment the left ventricle (LV) (r=0.96), myocardium (r=0.92) and right ventricle (r=0.92). In addition, 4D images from 10 patients with hypertrophic cardiomyopathy were also manually and automatically segmented yielding a good correlation in the volumes of the LV (r=0.93) and myocardium (0.94) when the atlas constructed with volunteers is blurred.

Introduction

Advances in cardiac imaging techniques have made possible to obtain high resolution images in a complete cardiac cycle. An accurate identification of the borders of the structures to be analysed is needed in order to extract physiologically meaningful quantitative information from the images. Potential applications of cardiac segmentation include the calculation of volume and mass, blood ejection fraction, analysis of contraction and wall motion as well as the 3D visualisation of cardiac anatomy (Frangi et al., 2001). While good results for 3D wall motion and left ventricular (LV) function have been shown in imaging modalities like SPECT or 3D echocardiography (Thirion and Benayoun, 2000; Sanchez-Ortiz et al., 2000), the application of automatic segmentation of other important anatomical structures such as the right ventricle (RV) and myocardium are limited by the imaging acquisition and the low contrast and signal to noise ratio of the images.

Segmentation of such images needs to be automated in order to be clinically valuable and to avoid time-consuming and partly subjective manual delineations. Segmentation of cardiac images is not trivial since the images are noisy, the edges are blurred due to partial volume effects and they can have motion artifacts. There is also especially low tissue contrast between the myocardium and surrounding tissues. Several approaches have been proposed for the automatic segmentation of cardiac structures (for a review see (Suri, 2000; Hammoude, 1998)). Recently, a number of techniques which are based on the use of a model or an atlas have been proposed (Vincent et al., 2000; Lelieveldt et al., 1999; Kaus et al., 2003; Lorenzo-Valdes et al., 2002). In these approaches a statistical model or atlas is used to incorporate a priori information which enables the use of both intensity and spatial information during the segmentation process. In particular, some techniques based on active appearance models (AAM) have emerged showing improved reliability and consistency (Mitchell et al., 2001). Moreover, extensions to 3D (Mitchell et al., 2002) and time sequences (Bosch et al., 2002) have been recently proposed. However, the applicability is restricted to the magnetic resonance (MR) imaging sequence used for training since the intensity appearance and distribution is an explicit part of the statistical model, i.e. an active appearance model trained on spin-echo MR images is not necessarily useful for the segmentation of TrueFISP MR images of the heart.

This paper proposes an approach that works for different MRI sequences, since it uses probability distribution maps and the intensity models are created for the image to be segmented at the time of the segmentation. Our approach combines the expectation maximisation (EM) algorithm (Dempster et al., 1977) and a 4D probabilistic atlas of the heart for the automatic segmentation of 4D cardiac MR images. Methods based on the EM algorithm have been previously proposed for the classification of MR images of the brain (Wells III et al., 1996; Liang et al., 1994). A number of authors (Leemput et al., 1999; Zhang et al., 2001) have proposed to include contextual information into the EM algorithm by means of Markov Random Fields (MRF). In this work we use an extension of the EM algorithm to 4D (space and time) and MRFs to segment a complete 4D sequence of cardiac images. In addition the approach extracts the largest connected component for each structure as a global connectivity filter after the classification. Introducing contextual information in 4D into the EM algorithm and adding the connectivity filter improves significantly the consistency of the segmentation process. We also use a 4D probabilistic cardiac atlas to include spatially and temporally varying a priori information into the EM segmentation. The following section introduces the EM algorithm, MRFs and the probabilistic atlas. Section 3 describes the automatic segmentation technique for 4D cardiac images. Section 4 presents some results of the proposed segmentation approach applied to 4D cardiac MR images of 14 healthy volunteers and 10 patients. Finally, Section 5 discusses the results and proposes future research.