Research On MRI Image Enhancement Based On Deep Learning And Metamaterial Composite Technology

MRI has an important application in the field of medical imaging,that is,by scanning human body parts to obtain certain tissue images to help doctors diagnose the cause.The nuclear magnetic resonance instrument scans to make the hydrogen nucleus in the human body resonate to obtain the density distribution of hydrogen nucleus in human tissue.In different parts of the human body,because the water content of the diseased tissue is not the same as normal tissue,the status of the tissue is obtained by imaging means as a basis for diagnosis.Generally speaking,the quality of an MRI image depends to a large extent on its spatial resolution and signal-to-noise ratio.However,it is difficult and difficult to improve the signal-to-noise ratio and the spatial resolution of the image together.Therefore,how to quickly obtain high-quality MRI images through low-cost methods is of vital importance.This article will implement MRI image enhancement research based on deep learning and metamaterial composite technology.Based on deep learning technology and metamaterials,a hybrid metasurface with uniformly enhanced magnetic field and tunable based on metamaterials is prepared in this paper to enhance the receiving efficiency of the receiving coils of nuclear magnetic resonance imaging instruments.When the spatial resolution of the MRI image is constant,the SNR of the image is improved to obtain a high-quality high-resolution MRI image.Based on the high-resolution nuclear magnetic resonance images obtained from the application of metamaterials,this paper proposes a deep learning-based deep neural network for image super-resolution research.The optimization direction and innovations of the proposed deep neural network include:(1)delete the residual network Batch normalization layer.Because the batch normalization layer reduces the flexibility of network characteristic parameters,we delete it.At the same time,after deleting the batch normalization layer,the model will save a lot of GPU memory consumption during training.(2)Use L1 loss function to complete network training.The L2 loss function is generally the most widely used loss function in the field of image super-resolution,but research shows that it cannot achieve the best performance of the trained network at the network training level.Instead,the L1 loss function can solve this kind of problem to a certain extent Problem,therefore,this paper uses L1 loss function instead of L2 loss function to complete network training.(3)The SE module is added to adjust the weights of different feature channels,so that valid feature channels have a large weight,and invalid or small effect feature channels have a small weight.For convolution operations in the field of image super-resolution,a large part of the workload of network training is to fuse more features of the image itself in space,or to extract multi-scale spatial information.For feature fusion of feature channel dimensions,general convolution operations default to feature fusion of all channels of the input feature map with the same weight.But in fact we want to group and weight the feature channels to make the model more simplified and reduce the calculation amount of the model.Therefore,the SE module is used to make the model automatically learn the importance of different channel features.Finally,the advantages of the proposed algorithm and the existing image super-resolution algorithms are compared by using the peak signal-to-noise ratio and the structural similarity coefficient.Experiments show that compared with the existing Bicubic and SRResNet algorithms,the peak signal-to-noise ratio is improved to 34.11dB and the structural similarity coefficient is also improved to 0.8524.Therefore,the algorithm proposed in this paper can play a certain role in promoting MRI.