Three-dimensional registration methods for multi-modal magnetic resonance neuroimages
In this thesis, image alignment techniques are developed and evaluated for applications in neuroimaging. In particular, the problem of combining cross-sequence MRI (Magnetic Resonance Imaging) intra-subject scans is considered. The challenge in this case is to find topographically uniform mappings in order to register (find a mapping between) low resolution echo-planar images and their high resolution structural counterparts. Such an approach enables us to effectually fuse, in a clinically useful way, information across scans. This dissertation devises an alternative framework by which this may be achieved, involving appropriate optimisation of the required mapping functions, which turn out to be non-linear and high-dimensional in nature. Novel ways to constrain and regularise these functions to enhance the computational speed of the process and the accuracy of the solution are also studied. The algorithms, whose characteristics are demonstrated for this specific application should be fully generalisable to other medical imaging modalities and potentially, other areas of image processing. To begin with, some existing registration methods are reviewed, followed by the introduction of an automated global 3-D registration method. Its performance is investigated on extracted cortical and ventricular surfaces by utilising the principles of the chamfer matching approach. Evaluations on synthetic and real data-sets, are performed to show that removal of global image differences is possible in principle, although the true accuracy of the method depends on the type of geometrical distortions present. These results also reveal that this class of algorithm is unable to solve more localised variations and higher order magnetic field distortions between the images. These facts motivate the development of a high-dimensional 3-D registration method capable of effecting a one-to-one correspondence by capturing the localised differences. This method was seen to account not only for topological differences but also for non-linear deformations in size and shape. Validation of the algorithm is carried out on geometrical objects, simulated data and real images to ensure that the important requirements for a topologically useful mapping; invertibility, smoothness of the deformation field and an almost perfect correspondence can be maintained between two image sequences.