Affiliation: College of Arts and Sciences, Department of Computer Science
Non-rigid image registration is fundamentally important in analyzing large-scale population of medical images, e.g., T1-weighted brain MRI data. Conventional pairwise registration methods involve only two images, as the moving subject image is deformed towards the space of the template for the maximization of their in-between similarity. The population information, however, is mostly ignored, with individual images in the population registered independently with the arbitrarily selected template. By contrast, this dissertation investigates the contributions of the entire population to image registration. First, the population can provide guidance to the pairwise registration between a certain subject and the template. If the subject and an intermediate image in the same population are similar in appearances, the subject shares a similar deformation field with the intermediate image. Thus, the guidance from the intermediate image can be beneficial to the subject, in that the pre-estimated deformation field of the intermediate image initiates the estimation of the subject deformation field when the two images are registered with the identical template. Second, all images in the population can be registered towards the common space of the population using the groupwise technique. Groupwise registration differs from the traditional design of pairwise registration in that no template is pre-determined. Instead, all images agglomerate to the common space of the population simultaneously. Moreover, the common space is revealed spontaneously during image registration, without introducing any bias towards the subsequent analyses and applications. This dissertation shows that population information can contribute to both pairwise registration and groupwise registration. In particular, by utilizing the guidance from the intermediate images in the population, the pairwise registration is more robust and accurate compared to the direct pairwise registration between the subject and the template. Also, for groupwise registration, all images in the population can be aligned more accurately in the common space, although the complexity of groupwise registration increases substantially.