Research On Fast MRI Reconstruction Based On Feature Fusion

As a non-invasive and radiation-free technology,compared with other imaging methods,MRI has better soft tissue contrast in visualizing anatomical structure and physiological function.One of the main disadvantages of MRI is that the acquisition process is too long.A long acquisition time may lead to discomfort of patients,which is easy to be affected by physiological movement and cause artifacts in the reconstructed image.In order to reduce the sampling time and accelerate MR imaging speed,a large number of traditional reconstruction algorithms based on undersampling k space have been proposed by many researchers.But these methods consume a long reconstruction time.In contrast,convolutional neural network（CNN）has strong ability of image feature extraction and nonlinear fitting,which has attracted the attention of researchers in the field of image processing.The MRI reconstruction methods based on deep learning have better reconstruction performance,but the existing methods are still insufficient in restoring the spatial structure information of MRI images.In order to solve the above problems,we proposed a new deep neural network based on feature fusion.Firstly,a multi-scale feature fusion network with gradient guidance is proposed.The network includes two branches: image reconstruction branch and gradient branch.The image reconstruction branch adopts the serial feature fusion blocks（FFB）to reconstruct the anti-aliased images in stages.And a three channel multi-scale residual block is proposed to adaptively extract the image features of different spatial scales.Meanwhile,the dilated convolution is used to further improve the feature extraction ability of the module without increasing the computational complexity.FFB groups four residual modules together and fuses the features from each residual modules to generate richer image features.Finally,the channel attention mechanism is introduced to further extract the features.Secondly,the gradient branch is introduced into the network.The gradient map of the image is exploited to guide the reconstruction of MRI image in the gradient branch.On the one hand,the gradient branch is used to restore the high-resolution gradient map,which provides additional prior information of image structure for the reconstruction process.On the other hand,the gradient loss function is introduced in the training process.Along with the image space loss function of the reconstruction branch,to help the network focus on the geometric structure information,so that the reconstructed result images contain rich structural information.Finally,extensive experiments demonstrate the effectiveness of the proposed method.Compared with other algorithms,comprehensive experiments on different acceleration rates and sampling patterns show that the proposed network can reconstruct high quality MR image with higher PSNR and SSIM.These experiments proved the robustness and the efficiency.