Study On Helical CT Image Reconstruction Based On AVX Instruction Set

At present, the degree of concern of Cone-Beam Computed Tomography isbecoming higher and higher in the field of medical imaging, because of the cone-beamCT has many unique advantages in the field of three-dimensional image reconstruction.Katsevich algorithm is the most representative algorithm in the exact reconstructionalgorithms of CBCT, it can do exact reconstruction for the reconstruction image, but thereconstruction formula of this algorithm is very complicated, at the same time thisalgorithm needs a very large of data, thus causes the slow speed imaging, and it is abottleneck in the practical promotion and application for this algorithm.This thesis mainly optimizes the software of Katsevich algorithm, and themodeling language is C++, which is based on the instruction sets of AVX, because of itincreased a lot of new functions on the basic of C++.Content of this paper mainly includes the following several aspects：1. This paper analysis the parallel processing of data in various steps of Katsevichalgorithm, and a design scheme about parallel processing of data is proposed. Inparticular, the Hilbert transform module of this algorithm using FFT and IFFT programsin order to parallel processing of data, and the method of combining the split radix andmixed radix is proposed in the design of FFT, which can use parallel data to achieve the2048-point FFT operation, so that the Hilbert transform can be sufficiently play theparallel of instruction sets of AVX.2. According to the result of the previous analysis, the modules of the Katsevichreconstruction algorithm can be designed and implemented to use the instruction sets ofAVX. Especially in the design and implementation of the Hilbert transform module, theoptimization of the rotation factor can reduce the memory usage of the split radix for16-point FFT.3. Each module which is based on instruction sets of AVX is simulated: first, theresult of each module is verificated by the result of each module which is based on C++;then, the image is simulated, and this image is reconstructed by the algorithm which isbased on the AVX instruction sets; in the last part of the paper, the speed of theKatsevich reconstruction algorithm which is based on the AVX instruction sets issimulated. Our algorithm performs reconstruction of a512×512×265volume from360views of the size512×512about16.4s by using AVX instruction sets，and this speedincreases three times than the speed of the algorithm which is based on C++.