Researches And Improvement To Group-wise Registration And Atlas Construction Techniques For Medical Images

Medical image processing, as the joint field of image processing and related medical researches, has grown rapidly in the past several decades. More topics gradually occur with more advanced science and technologies. And medical image registration, especially group-wise registration, is among the most prevalent topics recently.Medical image registration is supposed to deform an image into the specified space without distortion to the content of the image. And group-wise registration is required to simultaneous registered several images. For medical images, those applications have brilliant prospects– they could track the symptom evolvement for an individual patient, and help insight the evolution mechanics of a certain disease, and even compare morphological diversity of human bodies from different populations and ethnics.This dissertation looks back the progress of medical image registration and its key points, and elaborates the registration model based on the definition of information entropy which is constructed over information theory. Then, the dissertation clarifies mechanism, merits and defects of the registration framework where stack entropy works as target function. In order to group-wisely register a large number of images, the dissertation later reports the hierarchical group-wise registration framework. The method firstly clusters images, which are properly pre-processed, into a pyramid, and then synthesize the final atlas gradually. Via this strategy, hierarchical group-wise registration method could decompose a large-scale group-wise registration into a series of small-scale problems, and decreases the computation complexity.In the following experiments and applications, hierarchical group-wise registration performs better than contemporary methods, including: feasibility to large image dataset, faster speed, more accurate results, and finer capture ability to morphological changes. The experiments have manifested that hierarchical image clustering is an efficient and effective group-wise registration strategy for medical images.