Research On Cone-beam CT Reconstruction And Visualization Technology For CT Images Based On CUDA

3D cone-beam CT reconstruction and visualization technology are among major technologies in medical image processing,in that he former reconstructs a volume which contains a mass of anatomical information from 2-D projection images,and the latter render this volume for better understanding in visual sensation.Both act as powerful aided means in modern diagnosis such as lesion localization and are indispensable parts in real CT applications.As recent years has witnessed a constantly improvement in source-detector orbiting capabilities in real CT system,bottleneck in efficiency during real commodity CT applications has turned from hardware platform to the speed of CT reconstruction and visualization.On some occasions required for both real-time processing and high quality,for example,monitoring systems employed in radiotherapy or image-guided surgery,however,current approaches have still been far away from being able to keep up with the speed of data acquisition,unless proprietary,inflexible and expensive special hardware is used.Therefore,it is of great significance to seek effective methods for real-time 3D CT reconstruction and visualization on the general PC platforms.GPU(Graphic Processing Unit) seems to be that method.As its programmability and parallel computational capability evolved rapidly these years,it turns out that GPUs are in fact an excellent match for CT reconstruction and its visualization,as both are computational intensive and can be divided into parallel mode.Compared with the traditional method,GPU-based efforts have revealed itself in fast reconstruction and visualization.Even though,some restrictions such as complexity in GPU programming and lack of flexibility still exist and prohibit them from being adopted in real CT system.In the end of 2006,Geforce 8 GPU with unified rendering pipeline was cast by Nvidia and then CUDA(Compute Unified Device Architecture) was introduced based on it.This revolutionary programming model has made GPGPU(GPU used for General Purpose) more effective and simpler.In this paper,CUDA technology is first discussed and then be applied to Bi-cubic B-spline interpolation.After that, commonly used algorithm for CT reconstruction and volume rendering are reviewed, meanwhile the new parallel algorithm based on CUDA is proposed.Finally both parallel algorithms are realized in common PC.1.Bi-cubic B-spline interpolation based on CUDAThe computation of control-point grid and convolution are all done on GPU. First both parts are decomposed individually in SIMD mode under the principle of CUDA programming design,and then different memory pattern are used to test the performance of GPU under CUDA.In the end,the results are shown and the efficiency is compared with traditional method realized in CPU.The result proves that compared with Bi-cubic interpolation based on CPU,our CUDA-based effort has exceeded to a large extent in speed.Besides,this method is easy to extend and use in an expedient manner.2.Fast CT reconstruction based on CUDAFirst the mainstream algorithm for CT reconstruction-FDK algorithm is presented,and most time-consuming stage,the back projection is decomposed into parallel mode so that it can be realized in CUDA.This composition consist of three main parts:one is to load the projection consecutively preparing for consequent interpolation;another is to load the volume into GPU device memory and then be distributed into a thread grid to begin the back-projection part,so that each voxel can be computed by one thread and accumulated in every angle in next stage.Finally, after the intensity of voxel is computed and accumulated by every angle,the volume is reconstructed.The result of experiment shows that the amount of time needed for back projection is significantly reduced on the general PC platform for the modest data set(256*256*256),and real-time complied with high-quality effect has been acquired,thus proving it to be an effective way for real-time reconstruction.While real-time effect is nearly gained for more larger data set(512*512*512),it can also be made up to by using GTX 8800 GPU,which contains 128 streaming processors.In addition to that,the experiment also proves that computational complexity is not the bottleneck during GPU computing,so it can be extended to traditional iterative reconstruction algorithms.3.Parallel volume rendering based on CUDA interacting with openGLHigh-quality algorithm-ray casting algorithm is discussed and then be transformed into SIMD mode:the resolution of pixel-space act as an thread grid, every thread in this grid is used to compute the interpolation and composition along one ray,while the view point as well as the translation matrix are initialized by openGL and then be loaded to GPU,the result are bound from CUDA to openGL frame buffer object for interactive display.The whole process is realized in GPU. The rendering result shows that for the clinical data with resolution of 521*512*512, under CUDA flexible rendering effect can be easily done.Even for the most computational intensive rendering effect-phong model effect,this method can be displayed in more than 10fps.At the same time,as CUDA can interact with openGL, user can set some display parameters for better and clearer observation and in this way,a friendlier interface can be achieved.