Study On The Techniques For 3D Reconstruction And Visualization Of Medical Images Based On GPU

3-D reconstruction and visualization of medical images are important application of computerized technology in biomedical engineering. On the basis of knowledge concerned about digital image processing, computer graphics, and general understanding in medicine, it uses sequence of cross-sectional slices (such as CT, MRI etc) as its data sets, provides third dimension image for modern medical diagnosis. To ensure high performance and speed, however, current approaches for 3-D reconstruction usually can only be implemented on professional graphic accelerator cards or large-scale graphic workstations, both of which are very expansive. It is therefore of great significance to seek effective methods for real-time 3-D reconstruction and visualization technology on the general PC platforms.As programmability and parallel computational capability of Graphic Processing Unit(GPU) evolved rapidly these years, researches based on General Purpose computation on GPU(GPGPU) have also grew very fast, among which GPU-based algorithms for 3-D visualization made it possible for real-time rendering on common PC. By taking advantage of hardware characteristics of GPU, 3-D reconstruction and image displaying using this approach can be rapidly implemented by programming through high-level shading languages at lower expanse compared with the traditional method, but some limitation such as complex programming model and instability still prohibits it from being widely adopted in real diagnosis system.In the end of year 2006, the new GeForce 8800 GPUs with unified shading architecture was released by NVIDIA. This series fully supports DirectX 10 and first introduced CUDA(Compute Unified Device Architecture) concept, which, as a result made the realization of GPGPU technology more effective and simpler. In this paper, GPGPU technology on the GeForce 8 Series architecture is explored, surface rendering and volume rendering approaches for 3-D reconstruction of medical image is studied, then two new reconstruction algorithm based on GPU are presented and finally implemented on the common PC platforms.1. Fast surface rendering based on geometry shaderThe whole process of surface rendering is implemented on the Shader Model 4.0 rendering pipeline. It can be generally divided into two parts: the first is Marching Cubes algorithm for reconstruction based on geometry shader. It uses 3-D texture as its carrier, extract isosurfaces in every cube and export triangles during each computation on geometry shader. By this way, batch processing capability of geometry shader can be utilized to a large extent and thus computation can be accelerated. The second part focus on constructing Phone light models on fragment shader for follow-up rendering and displaying of triangle generated .Modeling and displaying are executed on GPU, thus leaving less dependency on CPU. On the whole, this algorithm is implemented by programming using high level shading language: GLSL. The result shows that, for the large-scale volume data sets, The Large image dataset can be visualized quickly with this modified method on the common PC platform.2. Volume rendering parallel algorithm based on CUDAFirst, redesign the traditional ray casting algorithm in accordance with SIMD programming model. Then under the principle of CUDA design model, thread allocation and data fetching are optimized. Two-slice shared memories are allocated to accelerate data accessing in sampling, while optimization of multi-thread allocation is to better developing parallel computational capability of GPU. In this modified algorithm, Rays are assumed to be emitted from projection plane and each thread is responsible for computation alone a certain ray direction. Time-consuming parts of this algorithm: trilinear interpolation and image synthesis are parallel computed, as a consequence shortening the whole processing time. The whole process of reconstruction is programmed by CUDA on the G80 platform without intervention of CPU. The result shows that superb performance can be attained using the volume consisting of 512×512×400 voxels.