Variational methods for shape and image registrations.

Shape registration;Finite element method;Image registration;Image segmentation

Rachid Fahmi, 1971-

University:University of Louisville

Department:Electrical and Computer Engineering

关键词: Shape registration;    Finite element method;    Image registration;    Image segmentation;   

Others  :  https://ir.library.louisville.edu/cgi/viewcontent.cgi?article=1417&context=etd

美国|英语

来源: The Universite of Louisville's Institutional Repository

【 摘 要 】

Estimating and analysis of deformation, either rigid or non-rigid, is an active area of research in various medical imaging and computer vision applications. Its importance stems from the inherent inter- and intra-variability in biological and biomedical object shapes and from the dynamic nature of the scenes usually dealt with in computer vision research. For instance, quantifying the growth of a tumor, recognizing a person's face, tracking a facial expression, or retrieving an object inside a data base require the estimation of some sort of motion or deformation undergone by the object of interest. To solve these problems, and other similar problems, registration comes into play. This is the process of bringing into correspondences two or more data sets. Depending on the application at hand, these data sets can be for instance gray scale/color images or objects' outlines. In the latter case, one talks about shape registration while in the former case, one talks about image/volume registration. In some situations, the combinations of different types of data can be used complementarily to establish point correspondences. One of most important image analysis tools that greatly benefits from the process of registration, and which will be addressed in this dissertation, is the image segmentation. This process consists of localizing objects in images. Several challenges are encountered in image segmentation, including noise, gray scale inhomogeneities, and occlusions. To cope with such issues, the shape information is often incorporated as a statistical model into the segmentation process. Building such statistical models requires a good and accurate shape alignment approach. In addition, segmenting anatomical structures can be accurately solved through the registration of the input data set with a predefined anatomical atlas. Variational approaches for shape/image registration and segmentation have received huge interest in the past few years. Unlike traditional discrete approaches, the variational methods are based on continuous modelling of the input data through the use of Partial Differential Equations (PDE). This brings into benefit the extensive literature on theory and numerical methods proposed to solve PDEs. This dissertation addresses the registration problem from a variational point of view, with more focus on shape registration. First, a novel variational framework for global-to-local shape registration is proposed. The input shapes are implicitly represented through their signed distance maps. A new Sumof- Squared-Differences (SSD) criterion which measures the disparity between the implicit representations of the input shapes, is introduced to recover the global alignment parameters. This new