GPU Parallel Computing And Its Application In Photoacoustic Image Reconstruction

With the development of the GPU computing technology,more and more programs with parallelism are implemented using the GPU parallel computing method to achieve higher performance and efficiency.Today,GPU parallel computing makes it possible for programs that we previously thought they are impossible to implement due to the long computation time.Compressed sensing photoacoustic computer tomography(CS-PACT)is a commonly used medical image reconstruction method that can produce high-quality images with sparse sampling.However,due to the reconstruction process requiring hundreds of iterations,the computational complexity is very high,making the image reconstruction very slow.Therefore,in order to improve the speed of image reconstruction,this paper applies GPU to photoacoustic image reconstruction,focusing on the GPU parallel computing framework and calculation method of CS-PACT algorithm,transplant the CPU algorithm to GPU.The main research contents of this article are as follows:(1)GPU parallel computing models and methods for small-size images.A GPU parallel computing architecture is proposed which based on iterative reconstruction algorithm.In general,we extracts five main types of parallelization operators form the algorithms,then design them in parallel,and optimizes them by the various optimization methods.Finally,two different human data(128 * 128,256 * 128)were used for qualitative and quantitative analysis to prove the accuracy of the GPU reconstruction method.For the B-scan photoacoustic image of 128 *128(pixels)human hand 1,we achieved an image reconstruction speed of 1.9-2.5s,which is 24-31 times faster than the CPU performance.For the B-scan photoacoustic image of256 *128(pixels)human hand 2,we achieved 5s-8s image reconstruction speed,which is 26-28 times faster than the CPU performance.(2)GPU parallel computing models and methods for large-size images.In the process of reconstructing a large image,the measurement matrix K causes a memory overflow,so we unable to reconstructed the image.To resolve the problem,a GPU parallel computing architecture based on CUDA stream is proposed.The large image is divided into multiple small images,and then the CUDA stream is used to implement the reconstruction of small images.In the reconstruction method for each small image,a specific kernel function is designed and analyzed.Finally,the phantom experiments of needle-inserting and rat experiment are conducted to verify the accuracy of the proposed GPU reconstruction method,and to evaluate the performance on acceleration effects compared with the CPU.For the needle-inserting image with 512 * 320(pixels),we have achieved 37 times acceleration compared with the CPU.Thereconstruction time of small images is measured under different numbers of CUDA streams,which proves that using CUDA streams can achieve computational acceleration.For rat experiment we have achieved 38-42 times acceleration compared with the CPU.