Accelerating Magnetic Resonance Image Reconstruction Based On Commodity Graphics Hardware

Magnetic resonance imaging(MRI)is a medical imaging examination technology without ionizing radiation,which is widely used in various scenes of clinical diagnosis.In recent years,there have been many advances and breakthroughs in the field of MRI.Various new imaging methods,such as positive contrast magnetic resonance imaging and cardiac movie imaging,have been proposed successively,further promoting the application of MRI in clinical medicine.However,these advanced imaging methods often require some complex iterative reconstruction algorithms to recover magnetic resonance images from the acquired signals.On the other hand,the widespread application of fast magnetic resonance imaging methods such as parallel imaging has made the data set acquired by the multi-channel coils method larger.As a result,huge data sets and complex reconstruction algorithms have resulted,making the entire reconstruction task computationally expensive and reconstruction calculation more time-consuming.Modern Graphic Processing Unit(GPU)has the characteristics of multi-core and multi-threading,and their computing power and bandwidth are significantly higher than the central processor used for processing general tasks.These characteristics of the GPU make it particularly suitable for computing tasks that can be expressed as data parallel.Based on these characteristics of GPU,this paper proposes a multi-level parallel magnetic resonance image reconstruction method to reduce the time required for image reconstruction,and to meet the increasingly urgent need of clinical magnetic resonance images in clinical medicine.This method is based on the Compute Unified Device Architecture(CUDA)provided by NVIDIA Corporation,which makes full use of the characteristics of large amounts of data that can be repeatedly calculated in magnetic resonance image reconstruction,and realizes multiple levels of parallelization on the GPU at the same time.This method first parallelizes the matrix vector operation in the magnetic resonance image reconstruction calculation based on the multithreading characteristics of the GPU;then based on the dynamic parallel technology,the reconstruction of multiple matrices or vectors in the spatial dimension achieves task-level parallelization;Further,based on multiple GPUs,higher-level operations such as 3D image reconstruction are parallelized.In this paper,the proposed multi-level parallel reconstruction method is experimentally verified on the Wave-CAIPI data set and the positive contrast magnetic resonance data set.Experimental results using 5 data sets of different sizes show that the method proposed in this paper can significantly reduce the time required for MRI image reconstruction without reducing the quality of the reconstructed image.Compared with the traditional method of image reconstruction by the Central Processing Unit(CPU),the proposed method implemented on the GPU is 6-15 times faster than that implemented on the CPU.