The Research Of The Medical Image Registration Based On New Criterion With Improved Optimal Strategy

In clinical diagnosis and treatment, in order to improve medical diagnosis and treatment effect, multiple lesion imaging of the focus is often required. This technique may obtain complementary, effective and comprehensive information. So, information of multiple images should be put togather and reflect a wide range of information on one image, that is called medical image data fusion. As the basis of medical image fusion, medical image registration has important clinical value, not only for diagnosis and treatment, but also can be used for tracking pathological changes, evaluation of treatment and other aspects. Medical image registration has become a hotspot in medical image processing, it aims to find an optimal transformation between two images, which makes one image comes to the unanimity with the other one both on the spatial location and the anatomical structure. Registration of medical images has a certain degree of complexity and difficulty, although many algorithms have been proposed, almost every method is designed for a particular problem with some limitations. So, these methods can not achieve the desired results both on the speed and accuracy.This article compares the characteristics-based and pixel-based methods for medical image registration, and then the mutual information based registration method is investigated in depth. After that, we analyzed the impacts of technologies on registration, such as differences between two images caused by the overlapping area, image gray level, interpolation and etc. Since mutual information is only involved with the image intensity, without taking into account the spatial relationship between pixels, so, a new similarity function, mini-mutual distance——CI_NMI is introduced. This function include the variance of each image. To improve the performance of medical image registration, a new method based on wavelet transformation are proposed, both with other different similarity functions. Decomposed wavelet sub-bands of the original images are used to improve the speed of image registration. Coarse-to-fine multi-resolution searches have been performed. Registrstion at higher levels can be carried out with the result at the pervious level, serving as the inatial condition. CI_NMI, which takes geometry into account is used as similarity measure at lower levels applied with Powell optimization method. Different methods are used at different levels. Compared with the traditional methods, it has advantages such as higher precision and better reliability according to the needs for time.A new partical swarm optimization (PSO) with dynamically variance inertia weight is proposed followed with the analysis of classical PSO algorithm. The weight is formulated as a function of these two factors, speed and aggregation factor, according to their impact on the search performance of the swarm. Additionally, a great lot of experiments are performed among these improvements. To improving the cease speed, parameter 'a' is added to make the iteration cease when it is approximately equal to one. Experimental results indicate that the new algorithm can have a high registration precision and rapid registration speed.