Abstract:

Microscopic imaging brings together all aspects of science in an effort to understand the complexities of the very small world. Medicine and biology have benefited immensely by being able to observe the microscopic world, for example in our understanding of the differences between normal and diseased cells. Image analysis has an important role here since the data is gathered in the form of images that are typically acquired in the three spatial dimensions and multiple spectral channels, and could also incorporate temporal information through time-lapse imaging. This dissertation presents algorithms for an image analysis system for such multi-dimensional images with emphasis on microscopic imaging of genomic structures in biology. We describe image processing approaches that refine images to make the task of segmentation easier and more precise. Segmentation algorithms are described that use domain knowledge to identify objects in images. Our segmentation also uses deformable models to extract surfaces of three dimensional structures which are used to compute various morphological properties of these objects and test scientific hypotheses. We present a novel approach for structural matching and registration that improves the iterative closest point (ICP) algorithm by using an intensity similarity measure to obtain the matching score. Some well known statistical and information theoretic measures are compared for use as the similarity measure and the robustness of the algorithm is evaluated on real and simulated images. The algorithm is applied to register temporal sequences of living cells showing chromatin domains. We extend the registration algorithm to correct for non-rigid deformations in images after describing the difficulty of the problem for microscopic data. We outline the problems inherent to the current non-rigid registration algorithms and present an approach that uses landmark correspondences obtained either from the previous matching algorithm or by user input, to deform images with a particular form of radial basis functions known as thin plate splines. We also present the application of fractal analysis for understanding genomic organization within the cell nucleus. The fractal dimension is used as a metric to characterize stages of the cell cycle in which the genomic material of the cell shows different packing properties. The fractal dimension is calculated using alternating sequential filters (ASF) and experimental results are given to show that there are statistically significant differences between the fractal dimensions of images taken from different stages of the cell cycle.