Fast 3D Cone-beam CT Image Reconstruction On GPU

Computed tomography (CT) has been used widely in medical diaglosis and non-destructive evaluation(NDE). Compared with two-dimension(2D) CT, three-dimension(3D) cone-beam CT, has much shorter scan time, higher axis resolution and better efficiency of using rays. The three-dimension CT reconstruction can be realized by algebraic or analytic methods. There are two types of analytic algorithms: exact and approximate algorithm. Although the images quality reconstructed by the approximate algorithms is not as good as that reconstructed by exact algorithm, the efficiency is higher than that of the exact algorithms, and scan locus could be incomplete. Feldkamp, Davis and Kress (FDK) proposed the practical approximate cone-beam reconstruction algorithm based on filtered backprojection(FBP) for the 3D CT with circular ray-source trajectory, which is extended to helical trajector. Recently, FDK algorithm is primary reconstruction algorithm in practical cone-beam CT. Algebraic reconstruction methods, such as the simultaneous algebraic reconstruction technique(SART), in certain scenarious, have many advantages over the FBP approaches and have attracted increasing attention in medical diagnosis and NDE.The three-dimension CT reconstruction is a time consuming process due to the vast complexity of data. In many applications, the reconstruction time is unsatisfied for practical requiring. The reconstruction speed of CT image restricts the application in engineering. Now, many methods have been proposed to accelerate the image reconstruction. These methods comprise two types, one is new fast reconstruction algorithm, the other is fast reconstruction hardware. Although the reconstruction time is shorted by using an application special integrated circuit(ASIC) or by using a field programmable logic array(FPGA), the high cost of these hardware limits their accessibility to many institutes. But the development of Graphic Process Unit(GPU) has changed the picture. The CT reconstruction acceleration based on GPU has become popular because of its low cost, good practicability and high computational speed.In this paper, we analyzed the characteration of GPU when it is used to accelerate the image reconstruction, and a new accelerated 3D reconstruction method for FDK algorithm with circular ray-source trajectory using multitexture of GPU(Graphic Process Unit) was developed. Using multitexture hardware, the reconstruction speeds up the backprojection, reduce the temporary memory for saving the result of backprojection, and decrease the floating-point sum passes. Using the color channels of vertexes to realize FDK depth weighting computation, and extending the texture units for backprojection, the reconstruction is accelerated. Therefore, the backprojection of a 256~3 volume can be completed with 16bit floating-point textures in less than 10 seconds. The experiment results show that the GPU accelerated 3D reconstruction method get a fast performance and good quality.By camparision with circular ray-source trajectory, we analyzed the reconstruction algorithm on GPU with helical ray-source trajectory. Based on the methods to speed up the reconstruction with FDK algorithm, we extended it to the condition of helical trajector. Therefore, the backprojection of a 256~3 volume with 512 projections(two circles) can be completed in less than 16 seconds. The experiment results show that the GPU accelerated 3D reconstruction method for helical cone-beam CT also get a fast performance and good quality.