Mutual Information Based Non-rigid Image Registration Using Adaptive Grid Generation: GPU Implementation And Application To Breast MRI

In this dissertation a new approach for non-rigid image registration using mutual information is introduced. A fast method for non-rigid registration is developed by adjusting divergence and curl of an intermediate vector field from which the deformation field is computed using finite difference method. The similarity measure mutual information is employed in the gradient-based cost minimization (or mutual information maximization) of the registration. The huge amount of data associated with MRI is handled by fully automated algorithm optimized with a multi-resolution topology preserving regridding scheme. The adaptive grid system naturally distributes more grids to deprived areas. The positive monitor function disallows grid folding and provides a mean to control the ratio of the areas between the original and transformed domain. The flexibility of the adaptive grid allocation could dramatically reduce processing time with quality preserved. Mutual information facilitates robust registration between different image modalities. Different types of joint histogram estimation are compared and integrated with the system. The whole system is also implemented on GPU which allows efficient parallel computation of vast v amount of 3D data in SIMD manner during different procedures. The GPU implementation offers up to 221 times speed up in the gradient normalization routine and around 40 times speed up in the overall calculation. This scheme is applied on 3D dynamic contrast-enhanced breast MRI, which requires the registration algorithm to be non-rigid, contrast-enhanced features preserving and within clinical visit time limit. Experiments show promising results and great potential for future extension.