In this thesis, we develop a new image registration framework with the following two major contributions. First, ordinal features are introduced and extended to represent images in the registration tasks. The ordinal features are extracted by passing through the images through the generalized ordinal filter bank, which effectively encodes the spatial information between neighboring voxels and specific micro-structural information in the images. The robustness of the ordinal features against noise is also investigated in this thesis. The ordinal features are then integrated with the intensity to form a two-element attribute vector. Second, we propose a new similarity measure function based on the generalized survival exponential entropy and mutual information (GSEE-MI). GSEE is estimated...[ Read more ]

In this thesis, we develop a new image registration framework with the following two major contributions. First, ordinal features are introduced and extended to represent images in the registration tasks. The ordinal features are extracted by passing through the images through the generalized ordinal filter bank, which effectively encodes the spatial information between neighboring voxels and specific micro-structural information in the images. The robustness of the ordinal features against noise is also investigated in this thesis. The ordinal features are then integrated with the intensity to form a two-element attribute vector. Second, we propose a new similarity measure function based on the generalized survival exponential entropy and mutual information (GSEE-MI). GSEE is estimated from the cumulative distribution function instead of the density function. It is observed that the interpolation artifact can be reduced and GSEE is more robust than the conventional MI. Also, the multi-dimensional GSEE-MI is proposed and defined so that the extracted features can be cooperate with the voxel intensity as the second layer. The proposed framework can be applied in both multi-modal and mono-modal image registrations.

The proposed framework is evaluated on four real MR-CT data sets for rigid multi-modal image registration. The experimental results show that the proposed method is more robust than the conventional mutual information based method and methods based on Gabor wavelets and gradient magnitudes. The accuracy of our method is comparable with these methods. For non-rigid mono-modal registration, the proposed framework is also evaluated by using both synthetic and real 3D datasets, and compared with two approaches: FFD using SSD alone and Demons. The experimental results show that the proposed method gives the highest accuracy in both normal and noisy environments among the compared methods.