| Non-rigid image registration plays an important role in medical image analysis, disease diagnosis and statistical parametric mapping. In this thesis, we particularly focus on developing novel features for robust image registration and designing an efficient evaluation protocol to measure the robustness and discriminant power of the features.;First, in the spatial domain, a new irnage feature called the uniform spherical region descriptor (USRD) is proposed. The USRD feature is rotation and monotonic gray-level transformation invariant, and is also computationally efficient. Each voxel is represented by its own USRD feature signature. The USRD feature is integrated with the Markov random field labeling framework for image registration. Second, we propose the symmetric alpha stable ( SalphaS) filters to extract image features in the frequency domain. The SalphaS filters are proposed because the energy spectrums of brain MR images often exhibit non-Gaussian heavy-tail behaviors which cannot be satisfactorily modeled by the conventional Gabor filters. The conventional Gabor filter is a special case of the SalphaS filters. The maximum response orientation criterion is designed to make the S?S feature rotation invariant. The SalphaS feature is integrated with the subvolume deformation model in the registration process. Moreover, in this thesis, we propose the Fisher's separation criterion (FSC) protocol which can directly evaluate the discriminant power of various types of features.;Finally, a multi-layer framework is proposed to extract features from input images from different views. The proposed methods are evaluated by performing non-rigid image registration experiments. The proposed methods are also compared with several state-of-the-art registration approaches. It is demonstrated that the proposed methods consistently achieve the highest registration accuracies among all the compared methods, which is matched with the results obtained from the proposed FSC evaluation protocol. |
