FUZZY KERNEL REGRESSION FOR REGISTRATION AND OTHER IMAGE WARPING APPLICATIONS

In this dissertation a new approach for non-rigid medical im- age registration is presented. It relies onto a probabilistic framework based on the novel concept of Fuzzy Kernel Regression. The theoric framework, after a formal introduction is applied to develop several complete registration systems, two of them are interactive and one is fully automatic. They all use the composition of local deforma- tions to achieve the final alignment. Automatic one is based onto the maximization of mutual information to produce local affine aligments which are merged into the global transformation. Mutual Information maximization procedure uses gradient descent method. Due to the huge amount of data associated to medical images, a multi-resolution topology is embodied, reducing processing time. The distance based interpolation scheme injected facilitates the similairity measure op- timization by attenuating the presence of local maxima in the func- tional. System blocks are implemented on GPGPUs allowing efficient parallel computation of large 3d datasets using SIMT execution. Due to the flexibility of Mutual Information, it can be applied to multi- modality image scans (MRI, CT, PET, etc.). Both quantitative and qualitative experiments show promising results and great potential for future extension. Finally the framework flexibility is shown by means of its succesful application to the image retargeting issue, methods and results are presented.