Research On MRI Image Reconstruction Method Based On Convolutional Global Attentio

Magnetic resonance imaging(MRI)is an important medical imaging technique,it has been widely used in clinical diagnosis due to its high contrast,high soft tissue resolution,and lack of radiation compared to computed tomography(CT).The MRI system collects signals point by point in k-space(i.e.frequency domain)and outputs corresponding high-resolution images through Fourier inverse-transform.High-resolution MR images can provide detailed anatomical information and enable radiologists to make accurate diagnoses.Due to the hardware limitation,the imaging time of MRI is too long,which causes great inconvenience to patients.To shorten the scanning time,the k-space signals are usually under-sampled at a lower rate than that of the Nyquist-Shannon sampling theorem.The images to be reconstructed not only loss a lot of details and textures,but also contain too many artifacts.For this reason,some techniques are needed to reconstruct the under-sampled images.In the past few years,the image reconstruction methods based on convolutional neural network(CNN)has made rapid development and been used in MRI widely.However,in the forward propagation of CNN,the low-frequency information is usually preserved,while the high-frequency information is lost severely,resulting in over smooth reconstruction results and lack of textures and details;In addition,due to the inherent limitation of the receptive field,CNN shows obvious deficiency in capturing global information.To address the serious loss problem of high-frequency information,a frequencydependent feature extraction module(FDFEM)is proposed.FDFEM consists of several cascaded multi-feature extraction modules(MFEMs),which use a low-frequency path and a high-frequency path to separate the input features into different frequency bands,the model can effectively extract local features containing both low-frequency and highfrequency information.To address the limitation of convolutional operation in extracting global information,a novel CNN-based global attention module(GAM)is designed.GAM uses a convolutional global attention(CGA)module to obtain the global attention weights for the input features,and gets features with global correlations by making element-wise multiply with the input.With only two 1x1 convolutional operations,GAM makes each pixel of the input feature globally correlated with all other points,and the low computational complexity enables the input feature to have a global receptive field even in the shallowest layer of the network.We trained and tested the network under different acceleration rates of different datasets and found that,the proposed method can achieve good results in quantitative metrics;In the qualitative comparation,the images reconstructed by our network were more in line with human visual perception.