Computational methods for live heart imaging with speed-constrained microscopes

Imaging methods to capture the beating and developing heart inside embryonic animal models such as the zebrafish are a key component for the study of fundamental biological processes such as cardiac birth defects or tissue regeneration. However, live heart imaging is subject to many challenges that limit the achievable spatial and temporal resolutions. While many commonly available microscopes offer the necessary optical sectioning capabilities to produce slices or volumes at the required scale, they often lack the necessary speed to seamlessly resolve the beating heart within a still frame, the entire heartbeat, or over a complete developmental sequence. In this thesis, I aim to assemble dynamic image series of the heart by using a wide range of microscopes, including, in particular, scanning microscopes---which are widely available but slow---and microscopes with a slow framerate. First, I considered the problem of reconstructing an image sequence covering one heartbeat from still images arbitrarily triggered throughout multiple cardiac cycles, when the underlying cardiac motion is temporally asymmetrical. I proposed to sort the images such as to maximize the overall similarity between contiguous frames, which I formulated as a TSP. The performance of the method increases with the number of images. I reconstructed image sequences of the heart at a virtual frame-rate of up to 300 frames per second. Second, to resolve the ill-posedness of the sorting problem when identical object poses appear multiple times in one cycle, I proposed to illuminate alternate frames with a temporally-asymmetric pattern. Sorting issues are mitigated at the cost of a custom setup and additional light exposure, without adversely affecting the reconstructed sequence. Third, I considered the problem of imaging the beating heart on conventional scanning confocal microscopes (LSM) that have a frame scanning rate comparable to the heartbeat rate. My approach takes a set of arbitrarily triggered images that may contain scanning aberrations and assembles it into a sequence that covers one heartbeat. Reconstructing scanning-aberrated frames offers a robust alternative for LSM. The quality of recovered cardiac heartbeat dynamics is comparable to that obtained by fast microscopes. Finally, I considered the problem of building smooth time-lapse sequences in the cardiac phase versus development or imaging depth spaces, based on a set of still images captured at weakly-constrained times within the heartbeat and over the course of development or at various depths. This method determines a smooth path in cardiac phase-development or depth space enforcing neighborhood similarity via a within-stage (or within-depth) sorting followed by a cyclical least-squares phase alignment approach that I formulated as a mixed integer linear problem. Constraining the paths of time-lapses by providing key-points while enforcing local similarity successfully stabilizes the the excursion error from wh