Use this URL to cite or link to this record in EThOS:
Title: Locally adaptive registration of serially acquired 3D magnetic resonance images using supercomputers
Author: Coley, M. D.
Awarding Body: University of Cambridge
Current Institution: University of Cambridge
Date of Award: 2003
Availability of Full Text:
Full text unavailable from EThOS.
Please contact the current institution’s library for further details.
Magnetic Resonance Imaging (MRI) is a non-invasive, non-ionising imaging modality which allows three-dimensional images to be acquired. It is particularly suited to repeat studies of the same subject. To aid comparison of the resulting images and allow any anatomical differences to be more easily visualised, it may be necessary to register the images to a common frame of reference. This ensures the same anatomy can be found at the same location in each image. For repeat scans of subjects with little or no expected anatomical change on well-maintained scanners, a rigid-body transformation model can give acceptable registration results. However, there are often geometrical differences between the serial images because of distortion and anatomical change. Various spatial transformation models exist which can be used during the registration process and the particular model used depends on the expected deformation necessary to register one image to another. This assumption is investigated in this thesis and it is shown that registration accuracy can be improved through the use of a new registration algorithm which allows spatially separate parts of an image to undergo their own independent rigid-body registration. This is done by subdividing one image into many overlapping cubic sections before using a separate process to register each sub-cube. As each process is independent of the others, this task is suited to running on a parallel computer. Following introductory chapters explaining the basics of MRI, image registration and parallel computing, this new registration algorithm is developed to be run in parallel on a supercomputer. The algorithm is validated through the use of computer-generated phantoms and a highly-structured physical phantom. The effect of sub-cube size and noise on the ensuing registration results is also investigated. For a cubic region in the centre of the physical phantom, the error in correspondence for an array of 343 regularly spaced spheres reduces from 0.562 mm using a rigid-body transformation model to 0.058 mm using the new locally adaptive registration algorithm. The registration errors associated with the new algorithm are significantly smaller (p < 0.0001, n = 343) from those using a rigid-body transformation model. Comparisons are also made with other public-domain rigid and non-rigid registration algorithms, with the non-rigid algorithms having less registration error.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available