GPU Acceleration Of3D Cone-Beam CT Reconstruction Based On FDK Algorithm

As the development of CT technology and improvement of CT applicationrequirements, three-dimensional CT imaging systems are increasingly common used,speeding up processing three-dimensional image reconstruction is one of the researchfocuses. The data of a3D scan of cone-beam CT is as large as hundreds of GB or TB,the reconstruction algorithm is complex in calculation，large in computing capacity andmuch in time-consuming, as a result it can’t meet the real-time requirement in practicalapplication. So it’s important to have a study of a quicker3D image reconstructionmethod for cone-beam CT.As FDK algorithm of3D image reconstruction has more remarkable advantagesin comparison, the FDK reconstruction algorithm and acceleration both were studied inthe paper. CUDA technology and the hardware characteristic of GPU have providedthe conditions for the FDK algorithm parallel acceleration. The paper firstly achievedthe FDK algorithm speed-up with CUDA technology in GPU. The result shows that theimage of2563volume in256rotation angles can be completed with single precisionfloating point, the reconstruction of GPU costs about2s while the reconstruction ofCPU costs2489s, which means the speed is more than1000times as fast, the image of10243volume in512rotation angles can be completed with single precision floatingpoint in about100s. While the time of the reconstruction of CPU is so long that it hasno practical value, acceleration of GPU has been fully embodied. In order to improvethe speed of CT image reconstruction in massive3D data further, the GPU cluster wasused to speed up FDK algorithm. Building a GPU cluster contains two GPUworkstations in the paper. CSocket mixed with CUDA programming, the GPU clusterfor amount of date used in FDK algorithm can be speeded up based on the Windowsoperating platform. In order to solve the limit of32-bit programming memoryallocation, the reconstruction for a size of10243was divided into8subtasks, in whicheach subtask only deals with the single-precision floating-point data format size of1024x1024x128. Using GPU cluster to reconstruction, the reconstruction resultswere sent back to the host, and then full reconstruction image was assembled. Theresult shows that the image of10243volume in512rotation angles can be completedwith single precision floating point in58s, compared to single GPU of the sameperformance, the speed is close to two times as fast. Parallelism of the FDK was analyzed in the paper, the FDK acceleratedreconstruction algorithm based on CUDA used single GPU and GPU cluster wereachieved. It costs more than1minute to collect1024×1024×128projection data, thereconstruction time had better be in1minute, the experiments results show that theFDK accelerated reconstruction algorithm based on GPU cluster satisfies the actualdemand, and is of great values of engineering application in industrial CT.