Research On The Key Problems Of MRI Image Reconstruction Based On Deep Learning

Magnetic resonance imaging(MRI)has become one of the most commonly used imaging techniques because it has the advantages of no ionizing radiation damage,high contrast scanning of soft tissues,and multi-directional image acquisition.However,the long time of data acquisition in the process of magnetic resonance imaging is likely to increase the cost,aggravate the condition of patients and increase the possibility of motion artifacts.The most direct way to speed up MRI imaging is to reduce the scanning time,that is,to shorten the time of data acquisition.However,less data is used for reconstruction,and the image will be aliased.As an efficient feature learning method,deep learning is widely used in the field of image processing.Generative adversarial network(GAN)have attracted attention for their ability to generate sharper and clearer images than other models without inferring during learning.Based on this,this paper uses deep learning to quickly reconstruct undersampled data to obtain high quality images.By improving the Deep De-Aliasing Generative Adversarial Networks(DAGAN),the relationship between reconstruction time and reconstruction quality is balanced.First,the paper propose a new MRI reconstruction algorithm based on attention mechanism and spectrum normalization.In view of the problems such as insufficient extraction of detail information and unstable network training in DAGAN,attention mechanism is added to the generator module to carry out adaptive feature refinement of the input feature map,so that the generated image can obtain more details to better conform to the real image features.In the discriminator module,spectral normalization is used to constrain convolution in order to stabilize the training process,and its variation is limited within a certain range.The network was trained using 100 3D magnetic resonance images(each with multiple slices)and tested using 50 images.The results show that all the improved parts are effective.The indexes of PSNR,SSIM and NMSE of reconstructed images under different undersampling rates are better than the existing algorithms,and the model has good generalization.Second,the paper propose another MRI reconstruction algorithm based on improved residual network.More image information can be obtained by deepening the network directly.Meanwhile,the improved residual network architecture is designed to avoid the disappearance of gradient and the deterioration of network performance.The convolutional layer,batch normalization layer and Leaky Re LU activation function are added to the shortcut connection mode,so as to unify the number of channels input to the deeper network while further learning the details.The experimental results show that the improved method is better than the original algorithm in the reconstruction of low undersampling rate data,and it can recover more complete image information and eliminate most image artifacts when tested on different types of data sets.The above two methods are effective in improving the original DAGAN algorithm from different perspectives.Under the existing experimental conditions,the reconstruction time of the first method is about 20 ms,and the reconstruction time of the second method is about 1s,which has certain application value.