Research On Feature Points And Gradient Feature Methods Of Medical Image Registration Based On Entropic Graph Estimation

Medical image registration is a fundamental task in image fusion, and widely used for diagnosis disease, planning treatment, guiding surgery and studying disease progression. After image registration via maximization of mutual information was introduced by two independent groups:Collignon and Maes et al., and Viola and Wells, entropy-based similarity measure in medical image registration has been used widely. Variety of entropy-based similarity measures are needed to estimate the entropy from the samples of the images. Redmond and Yukich proved that when a graph is continuous and "quasi-additive", the graph can be used to estimate the entropy directly. Hero further proposed the theoretical framework of estimating the Renyi entropy by minimum spanning tree (MST).Based on these work, given on some limitations faced by most existing image registration approaches based on entropic graph estimation, this dissertation proposes a set of solutions based on feature points and gradient feature, and achieves good results. The main research works implemented by the dissertation include:Research on Feature Points Methods of Medical Image Registration Based on Entropic Graph Estimation(1) The conventional image registration approaches based on entropic graph estimation use a single type of feature points. But, only using a single type of feature points is unstable, which do not take into account the special nature of medical images. To solve these problems, the dissertation proposes a complementary scale space keypoints based medical image registration approach. In the algorithm, two complementary scale space keypoints (Harris-Laplace and Lapacian of Gaussian) are extracted. Harris-Laplace detects mainly corners and highly textured points, whereas the Lapacian of Gaussian responds to blobs. They have good noise resistance and robustness. Comparison results showed that the algorithm proposed achieves better robustness and high speed than the algorithm based on single type of feature points.(2) Based on the subsection (1), this dissertation proposes a medical image registration approach integrated with multi-feature points’intensity information. In the algorithm, three kinds of feature point from image:Harris-Laplace, Laplacian of Gaussian, and Grid are extracted from image. It takes full advantage of complementary scale space feature points. Meanwhile, grid points are added to cover the low contrast regions, which improve the uniformity of the distribution of feature points. Then genetic algorithm is used for selection of feature points, which can reduce the influence of registration robustness caused by the irrelevancy and redundancy between the feature points. Comparison results showed that the algorithm proposed can provide better robustness, higher accuracy and much faster than the algorithm based on the subsection(1).Research on Gradient Feature Methods of Medical Image Registration Based on Entropic Graph Estimation(3) The conventional image registration approaches based on entropic graph estimation only consider pixel intensity, ignoring the spatial information between pixels, and the registration results is so sensitive to sampling rate that the registration robustness is thus reduced. To solve the problem, this dissertation proposes a medical image registration approach based on gradient information. In the algorithm, the weighting function based on the edge gradient information between images is defined. Then, the weighting function is used to correct the joint Renyi entropy. The experimental results showed that the algorithm proposed can provide smoother registration function.(4) Because of different imaging techniques, the same tissue in multi-modality medical images has different intensities. However, the images fundamentally depict the same anatomical structures. Using this characteristic, this dissertation proposes a multi-modality medical image registration approach integrated with gradient orientation information. In the algorithm, the contrast reversals, translation and rotation invariant feature is constructed. Then, the high dimensional features which combine the pixel intensity and the new feature are applied into multi-modality medical image registration based on entropic spanning graph estimator. The experimental results showed that the algorithm combining the advantages of two features, achieves better overall performance in the aspect of registration robustness and accuracy.(5) To ensure the accuracy of medical image non-rigid registration, this dissertation presents an approach integrated with scale invariant feature transform (SIFT) features for medical image non-rigid registration. The algorithm proposed utilizes SIFT high dimensional features as well as intensity to integrate with spatial information. Meanwhile, a rigid deformation combined with free-form deformation (FFD) based on cubic B-splines is used to capture the global and local motion between images. Comparison results showed that the algorithm proposed can provide better accuracy and reduce effectively the difference between the two images.