Image sequence analysis for dynamic and time-lapse microscopy

Medical and biological imaging are moving towards smaller and less-invasive devices for live-tissue and live-cell diagnosis. For example, recent advances in microscopy have led to miniature devices that can be brought to the patient for imaging living tissue in context. With these advances has come the ability to generate increasingly large and complex image data sets. This work presents a set of tools for processing and interpreting these large image sets in order to better characterize cells and tissues and improve diagnosis.;The first topic of this dissertation is dynamic exploration of living tissue using miniature laser scanning microscopy, which has applications in early detection of cancer and visualization of tumor margins. I present an algorithm for real-time image mosaicing that combines multiple overlapping images to provide macroscopic views while retaining micro-architectural detail. This mosaicing algorithm is combined with a probabilistic model to deal with cumulative image registration errors as well as non-rigid scene deformation. I also present the design of a hand-held robotic scanner for creating large and robust image mosaics of accessible tissue surfaces.;The second topic of this dissertation is time-lapse imaging of human embryogenesis, which has applications in assessing embryo quality for in vitro fertilization procedures. I present the design of a multi-channel microscope array with dark-field illumination that fits inside a standard incubator. This microscope array is used to non-invasively observe human embryo growth over several days and identify image phenotypes that predict embryo viability. To enable automatic extraction of these parameters, I present an algorithm based on particle filters that estimates probable models of 3D cell divisions from 2D image sequences.