2D-3D Medical Image Registration Technology Research Based On CUDA

The goal of image registration is to obtain a transformation, so that the points in one image can be related to their corresponding points in the other. Registration process is essentially a multi-parameter optimization problem. It is the basis of image analysis and the prerequisite for image fusion. Automatic Fast-rigid medical image registration can place a multiple sets of medical images into the same coordinate space through the rigid transformation, which is a common approach of medical image diagnosis. Medical image registration has a wide range of applications in surgical navigation, radiation therapy, assessment of therapeutic effects.2D-3D medical image registration is becoming increasingly important because of its function in computer-guided diagnosis, surgical operation navigation and so on. In these usages image registration is an important part. Registration between 2D image and 3D volume data provides real-time information for diagnosis and treatment, so it can aid surgery navigation. We usually get 3D volume data before surgery and obtain 2D image during the surgery. Although 2D data is deficient in spatial information, it can be got quickly and decrease exposure time in X-ray. So we retrieve some spatial information through 2D-3D medical registration process.At present, researchers pay more attention on rigid registration. Because of this kind of registration need more time, and generally difficult to meet the clinical application of real-time requirements; this paper based on the study of graphics hardware--GPU programming techniques and the currently popular parallel programming language--CUDA is to accelerate the process of registration and hope the process under this technology can meet the clinical applications.Digitally Reconstructed Radiography plays a key part in 2D-3D medical image registration based on gray. It obtain 2D virtual X-ray image (DRR image) from 3D data such as CT, using ray-casting algorithm. The quality and speed of DRR image directly influence the precision and speed of registration. This paper discussed the technology of producing DRR image and realized the process of producing DRR image from CT volume data. This technology based-on CUDA greatly accelerated the DRR image generation and created the necessary preparations for the following-up 2D-3D registration studies.2D-2D,3D-3D registration is the most basic registration operation. In order to verify the feasibility of acceleration based on CUDA technology, we carried out relevant research of these two operations in this paper and obtained good results. Then, on this basis, we carried out 2D-3D acceleration registration based on this technology.Take 3D CT volume data for example, during the 2D-3D medical registration process, we build a reasonable coordinate system to simulate the 2D clinical X-ray image generation environment. Here the sum of the correspondence absolute difference of pixel gray value plays the role of similarity measure. As for the optimal method, we use Powell method. In order to improve the registration efficiency, we also adopt several strategies to accelerate practical registration process, such as data rescaling and grey rescaling approach.In order to estimate system's accuracy, DRRs taken from known viewpoints have been used as reference images to see whether the result is approximately the same or not. The result indicates that over a range of viewpoints, the translational and rotational accuracy are under 1mm and 1°respectively, registration accuracy can meet clinical demand pretty well. In order to reflect the time advantage of the technology, we transform registration platform from the GPU to the CPU, with quasi-operational processes, interpolation method, similarity measure function, and optimization methods remain unchanged. The result indicates that the technology has fairly obvious time advantages, with the average acceleration speed up to tens or even hundreds of times.