| This thesis addresses two basic topics in cardiac image analysis: segmentation and motion tracking. The segmentation of the heart boundary has wide and important applications in clinical practice and is a crucial step in shape-based tracking. The proposed new segmentation method is targeted to echocardiographic images with minimal human intervention. The resulting "combinative multi-scale level set" approach uses shape knowledge learnt at a coarse scale of a pyramid of images to compensate for poor features in ultrasound images at a finer scale. The algorithm based on a level-set implementation offers good results and outperforms other solely edge-based or region-based approach for ultrasound images with reasonable quality.; As the second part of this thesis, we have developed a new framework for 4D dynamic cardiac motion analysis. The core algorithm, named Generalized Robust Point Marching (G-RPM), extends the Robust Point matching (RPM) to include shape information and improves its capacity to handle noise/outliers in both matching point-sets. The algorithm is capable of accurately estimating left ventricular (LV) motion without the need of a prior, and often time-consuming, segmentation of the myocardium (but the segmentation is still an inevitable measure for other analysis purposes such as LV volume, ejection fraction and etc.). The shape information is directly derived from the gray level images by using area-based operators, instead of utilizing geometry-based curvature operators.; For non-rigid motion modeling, an Extended Free Form Deformation (EFFD) has been developed to capture cardiac motion patterns. Unlike Free Form Deformation (FFD), the EFFD uses arbitrary-shaped lattices which extend the set of possible deformations and thus provide the flexibility to be tailored for specific motion analysis. In particular, in the estimation of LV deformation, the lattice of the EFFD is based on the anatomical structure of LV which ensures a more accurate estimate of LV deformation can be obtained. In comparison with the Finite Element Method (FEM), EFFD doesn't require a close boundary condition and the computation is more efficient. In addition, the algorithm can be easily manipulated to accommodate different feature information for different imaging modalities, for example, the speckle tracking information in the mid wall for ultrasound images. This is very important to achieve a modality-independent, sustainable system development strategy. The methods are applied to both 4D Magnetic Resonance (MR) cardiac images and echocardiographic images for regional motion and strain measurement. The cardiac motion analysis system developed in this work is of paramount importance to studies of cardiac regional function through medical imaging. |
