Research On Mri Reconstruction Algorithm Based On Deep Learning And Compressed Sensing

Magnetic Resonance Imaging is an important auxiliary diagnostic technique in clinical medicine.With ionizing radiation damage to the technology,the structure of soft tissue can clear display,to obtain the advantages of the composite image is widely used in clinic,and the acquisition process need a longer time,there will be the introduction of artifacts,and reduce the scanning time,thought that less data,makes the reconstruction effect is poorer.Therefore,it is of practical significance and research value to quickly reconstruct high quality MRI images with a small amount of data.In view of the above urgent research problems,this paper mainly studies the application of compressed sensing and convolutional neural network in the field of MRI reconstruction.The main work is as follows:(1)A joint sparse model based on ADMM expanded network was proposed.In the model,Shearlet transform and wavelet transform were used for joint sparse.Shearlet transform to MRI image multi-scale decomposition to more direction for image detail part of approximation,the deficiency of the Shearlet transform is the dot of one dimensional characteristics cannot be effective,says that with the to point to a good approximation of wavelet transform,and considering the introduction of impulse noise,the norm of the fidelity term,when solving reconstruction model,the traditional thoughts and deep learning ideas,the combination of solving iterative data flow graph form neural network for training.Through experiments,the results show that the method proposed in this chapter can reconstruct a clear MRI image in both qualitative and quantitative evaluation criteria.(2)A multi-scale MRI reconstruction model based on U-net was proposed.In the model of plural MRI image data processing,the amplitude and phase data form two channels of data input to improve U-net model,which is based on structure and expansion Inception convolution based U-net under the single dimension convolution of sampling module,obtain multi-scale features,under the sampling part use deconvolution to low resolution characteristic figure obtained by sampling part to improve the resolution.The feature graph in the bottom sample and the feature graph obtained from the top sample are spliced by jump connection to obtain more detailed features of the image.The experimental results show that the proposed method is effective in image reconstruction even at low sampling rate.(3)A multiscale complex MRI reconstruction model based on generative countermeasures network was proposed.The model includes two parts,namely the generator and discriminator.In order to improve the perception quality of the reconstructed mri image,the generator is trained by the loss function of the coupling of MSE loss of pixel similarity and VGG loss of perception similarity.The effectiveness of the proposed algorithm is verified by experiments.