Cardiac development is a highly dynamic process since the heart is beating long before it has reached its final shape. Currently there is a need for novel image processing techniques for image component separation, denoising, and reconstruction of high frame-rate, high resolution, multi-dimensional cardiac datasets acquired during live (in-vivo) imaging of the embryonic heart. Such methods would help in quantifying blood flow and tissue deformation during cardiac development, thus facilitating a better understanding of the interplay between genetic and epigenetic factors that contribute to cardiac morphogenesis.

The first part of this thesis deals with improving the specificity in label-free, high speed brightfield (BF) microscopy images. We propose a motion-based separation algorithm to decompose a single-channel 3D+time volume into three channels showing cyclic heart-wall, static support structures, and transient blood cells. The technique is based on non-uniform temporal synchronization, selection, and combination of images from multiple cardiac cycles and z-sections to produce 3D+time image volumes of one full cardiac cycle that are highly suitable for 3D visualization and analysis of the individual structures' dynamics in the embryonic heart.

In the second part, we develop a computational noise reduction technique to enhance optical coherence tomography (OCT) datasets of cyclically moving structures. This allows imaging dynamic structures and fluid flow within scattering tissue like the embryonic mouse heart, while preserving temporal and spatial resolution.

In the final part, we discuss reconstruction of a 3D+time cardiac volume from multiple high-speed 2D+time sequences that are sequentially acquired along multiple heart beats. Current methods overcome the limitation of low frame-rates in direct 3D+time imaging either by synchronizing the 2D+time sequences using complex triggering hardware during acquisition or by image-based retrospective temporal registration of the 2D+time sequences which accumulates registration errors. Our technique uses image-based retrospective gating for temporal alignment but mitigates the cumulative registration errors by using a second set of 2D+time sequences acquired at an angle orthogonal to the first. By globally minimizing an objective criterion that depends on the similarity between the data present at the intersecting slices, the two sets are registered and fused to obtain a high-speed high-resolution multi-dimensional reconstruction of the heart.

The techniques presented in this thesis have been applied for reconstruction of high-speed (100-3000 frames per second), 3D+time cardiac datasets with micrometer resolution (0.8-4.0 mum) in live cardiac BF microscopy images of zebrafish embryos and cardiac OCT images of mouse and rat embryos. The tools developed improve the visualization of cardiac BF and OCT datasets over existing image processing techniques, and enable quantitative study of cardiac morphogenesis.