Spatiotemporal strategies for shape and motion recovery from cardiac image sequences

Coronary heart diseases often manifest as abnormalities of the ventricular geometry and wall kinematics. Accurate and robust estimation of the cardiac shape and motion, from cine tomographic medical image sequences, is thus of significant clinical values and offers great technical challenges. We present several novel strategies for the image-based analysis of myocardial function. The core ideas come form the intension of integrating various data-derived shape and motion cues with model-driven prior constraints, which offers the possibilities for robust and simultaneous shape and motion recovery. In this thesis, we present segmentation and motion estimation strategies that integrated various image-driven information and motion constrains together by using different model-based approaches. Three strategies have been used to solve both segmentation and motion estimation. A velocity-constrained front propagation approach is first presented for myocardium boundary segmentation. Embedded within the level set paradigm, we take advantage of the often complimentary intensity and velocity information provided by the MR phase contrast images. Then, a modified, physically-motivated active contour strategy is proposed for the joint estimation of heart boundary geometry and kinematics by integrating data and model constraints on boundary tissue shape and motion coherence. This effort is further extended to an integrated active region model (ARM) for the simultaneous shape and motion recovery of the cardiac volume, including the left ventricle and the whole heart. The ARM framework is built upon an elastic solid that evolves to reach the equilibrium between the internal elastic stress and the external data-derived and model-driven forces, and the main novelty of the technique is that the external driving forces are individually constructed for each nodal point through the integration of the data-driven edginess measures, the prior spatial distribution of the myocardial tissues, the temporal coherence of the image-derived salient features, the imaging/image-derived Eulerian velocity information, and the cyclic motion model of the myocardial behavior. The finite element method provides the representation and computation platform for the effort, and iterative procedures are used to solve the given equations. We demonstrate the robustness and accuracy of these strategies with very promising application results using canine magnetic resonance gradient echo, phase contract, and tagging image sequences. Some of the segmentation and motion analysis results are further validated against histo-chemical staining of the post mortem myocardial tissues, the current gold standard