Minimum Spanning Tree Fusing Multi-Feature Point Information for Medical Image Registration

Medical image registration is a fundamental task in image process, and widely used for diagnosing disease, panning treatment, guiding surgery and studying disease progression. For medical image registration of high robustness, high accuracy and speed requirements, this paper proposes a minimum spanning tree (MST) algorithm of fusing multi-feature point information for medical image registration. This algorithm extracts three kinds of feature points from image: Harris-Laplace points, Laplacian of Gaussian points, and grid points. Then genetic algorithm is used for point selection; and by choosing appropriate cost function, the redundancy can be reduced in a great measure. The selected points are then used in building a vertices set of undirected complete graph by location-mapped method. Finally, MST is constructed by modified Kruskal algorithm, which estimates Rényi entropy directly. The new algorithm has solved the low robustness brought by the instability of extraction of feature points and the speed bottleneck problem when using MST to estimate the Rényi entropy. Experimental results show that in the images with noise, non-uniform intensity and large scope of the initial misalignment case, the proposed algorithm achieves better robustness and higher speed while maintaining good registration accuracy, compared with the conventional area-based and feature-based registration methods.