Research On The Image Reconstruction Algorithms For Linear Scan CT

Owing to a relatively simple implementation and high scanning speed, the recently existedLinear scan CT (LCT) has great significance and promising prospect in applications as securityinspection and industrial non-destructive testing. However, because of the limit of the ray-beamflare angle and detector size, the projection data sets are not sufficient for exact reconstructionand it suffers from a limited-view problem. Therefore, research on how to reconstructhigh-performance images from limited-view projections for LCT is one of the hotspots anddifficulties in the field of CT and has important theoretical significance and practical value.Considering the different characteristics and practical demands of spatial and Fourierreconstruction methods, this paper focuses on the research of the reconstruction algorithms inspatial domain and frequency domain for the limited-view problem of LCT, respectively, as wellas the parallel acceleration of reconstruction algothrims. The main research work is included asfollows:1. A spatial iterative reconstruction algorithm named ADTVM (Alternating Direction TotalVariation Minimization) which is based on total-variation (TV) regularization is proposed.Utilizing the piecewise constant characteristic of most images, employing the total variationminimization as objective function, the main idea is to reformulate the reconstruction problem asa linear equality constrained problem where the objective function is separable, and thenminimize its augmented Lagrangian function by using alternating direction method (ADM) tosolve subproblems. Experimental results show that compared with the widely used ASD-POCS(Adaptive-Steepest-Descent-Projection Onto Convex Sets) algorithm, the proposed algorithmconverges much faster and could achieve higher image quality at same iterations.2. A Fourier-based iterative reconstruction algorithm named RecSPF (Reconstruction fromSparse Partial Fourier data) based on a sparse resampling method is proposed. Considering theproperties of Fourier samplings in LCT, a sparse resampling strategy is used to build the imagespectrum from projection spectrum. With the aid of filtering the sampling data and keeping thebalance between the Fourier data integrity and the precision of samples, the new method couldprovide improvement in interpolation accuracy. The step of recovering missing information fromimage spectrum is developed based on the RecPF (Reconstruction from Partial Fourier data)method. In this step, a TV minimization model is designed and solved by the augmentedLagrangian function method and alternating direction method. Experimental results show thatthe proposed approach has the benefits in both computational efficiency and reconstructionaccuracy. And a satisfactory performance could be obtained when dealing with the limited-view problem of LCT.3. An acceleration strategy based on Sparse Matrix Vector multiplication (SpMV) and OpenComputing Language for ADTVM algorithm is proposed. Iterative algorithm is a timeconsuming process. Firstly, by exploiting the matrix’s structure of projector and backprojector inLCT scan, a matrix approach with SpMV to accelerate the projector and backprojector process isproposed. Then, a fast parallel implementation of ADTVM algorithm is proposed based on theOpenCL and Graphics Processing Unit (GPU) with some optimization techniques, includinggrouped CSR, ELLR-T method and so on. The results of simulation experiments show that, thefully optimized algorithm can run with an80times speedup ratio with no precision reduction.