Research On The Viscous Fluid Elastic Registration Algorithm Of Medical Images Based On GPU

With the development of medical imaging technology, medical image registration has become an indispensable part of the clinical medical diagnosis and treatment. It lets the same mode or different modal images to spatially consistent through some geometric transformation. Thus, it makes the useful information complementary between two images. It is beneficial to the clinical diagnosis and treatment of lesions and is an important research topic of medical image processing field. Due to the independent respiratory motion, the chest, lungs and other organs produce large scale deformation. The elastic registration based on physical model is suitable for elastic deformation, especially the viscous fluid model. It regards simulations of the registration process of images as the flowing process of fluid, and is more suitable for deformation of large scale. But the model of viscous fluid is sensitive for the change of gray, and its registration process is very time consuming.The viscous fluid model selects difference of gray of registration images as the similarity metric. It is more sensitive to the gray level transformation, easily drowned details. The similarity measure is a key step in medical image registration, and its selection directly influences the registration effect. This thesis uses mutual information as similarity metric, it is a good solution to the problem of gray sensitive, with high registration precision and good robustness. The algorithm of medical image registration based on viscous fluid finally obtains a problem of solving partial differential equation. It uses the method of successive over relaxation for solving partial differential equations in viscous fluid algorithm. The process of solving is very time consuming, this thesis uses the conjugate gradient method without the need to estimate the parameters for solving partial differential equations. The calculation of each iteration is mainly for the calculation of the vector, which is convenient for parallel operation, and the storage capacity is small and the stability is high. Experiments show that mutual information is more prominent in the details of the image, while the conjugate gradient method is effective in reducing the time required to solve the partial differential equations, while maintaining the accuracy of the registration and improve the registration efficiency.Medical image is the DICOM sequence image, which contains the information content is very rich. Although the improved viscosity fluid algorithm can reduce the time consuming of registration to some extent, but it can not meet the requirements of the clinical real time. In this paper, we study the improvement of viscous fluid and CUDA platform. Using GPU’s powerful computing power and CUDA parallel computing architecture, a new method based on GPU is proposed for viscous fluid medical image elastic registration. The algorithm is implemented by the mutual information and the partial differential equations in the CUDA platform, which makes the calculation of the two dimensional histogram and the velocity field of each moving direction. Experiments show that this method not only does not lose registration accuracy, but also reduces the time required for registration to a certain extent, and improves the efficiency of registration, and meet the clinical diagnosis..On the basis of the above work, the thesis completes the medical image registration system of viscous fluid based on GPU. The whole system has a friendly interface, the doctor can easily get the registration image he interested in. Its function is simple and easy to use, has high practical value.