Medical Image Registration Based On Spatially Weighted Mutual Information

Medical image registration plays an increasingly important role in medical researchand clinical treatment, it can transform the images of different mode or at different timesinto the same coordinate system, so that we can integrate the information of differentimages. Researchers have done a lot of work about image registration over recent years,various methods have been proposed. Among them, registration strategy based on mutualinformation (MI) has been proved to be a promising method and has been widely used inmedical image registration.However, MI only uses the statistics of the image intensity but ignores spatialinformation, so we introduce a novel metric extension to mutual information calledspatially weighted mutual information (SWMI) used for image registration. SWMIincorporates an adaptable weighting function with spatial information of pixel into mutualinformation in the process of computing the image information entropy, which not onlyconsiders the information of the image intensity, but also considers the spatial locationinformation of all pixels. In order to improve the automation of registration method basedon SWMI, we propose Harris weight (HW) function and LBP weight function used forcalculating the SWMI. We use Harris corner detection operator to classify the referenceimage area and set different weights for the classification pixels to generate the HWfunction, and set different parameters for Harris corner detection to get different weightingfunctions used for registration. LBP weight function can achieve automatic registration,and we define two forms of LBP weight function. We use the LBP operator to analysis thereference image to generate LBP single-made (LBPS) weight function, and LBPdouble-mode (LBPD) weight function uses LBP features of two images together togenerate, which provides a new way for defining weighting function. The results of rigidregistration show that the SWMI is a viable similarity measure, SWMI based on LBPDweight function can improve the performance of registration than MI, and we realign theimages using HW weight function while misalignment can improve the successful rate ofregistration.In many applications, non-rigid registration method is often required to use for aligning the images which exist deformation, we studied a B-splines based non-rigidregistration method, and use the normalized SWMI as similarity measure. In order toimprove the speed and accuracy, we also use hierarchical B-splines transform. The resultsshow that SWMI can help improve the accuracy of registration, and the hierarchicalB-splines transform method can help improve the accuracy and speed of registration at thesame time.