Research On CS-MR Image Reconstruction Based On Deep Learning

In clinical medical diagnosis,imaging technology as an auxiliary means can very well help doctors make judgments.Among them,magnetic resonance imaging(MRI)is widely used because of its high safety,no radiation damage,and the ability to obtain high-resolution threedimensional images and cross-sectional images of the human body.However,the long time of MRI has limited its further development.How to balance the sampling time and the amount of sampling data has always been a difficult problem.In recent years,the proposed compressed sensing theory has realized the reconstruction of the original signal with less sampled data,thereby speeding up the imaging speed of MRI and further promoting clinical trials.Its theoretical focus is the compressed sensing reconstruction algorithm,the quality of the compressed sensing reconstruction algorithm directly affects the quality of the reconstructed image.The traditional compressed sensing reconstruction technology has a complicated solution process and poor reconstruction quality.The reconstruction algorithm based on deep learning is the current mainstream method.This article aims to use deep learning to propose two innovative algorithms in Compressed Sensing MRI(CS-MRI)to achieve high-quality MRI reconstruction.The following studies have been completed jobs:(1)Aiming at the problem of incoherent MRI structure in the reconstruction process,a compressed sensing reconstruction algorithm based on multi-scale convolution U-NET is proposed.The network structure of this scheme consists of two parts: multi-scale convolution U-NET and data consistency layer.The two parts are alternately cascaded.A new variant of UNET is proposed here,namely multi-scale convolution U-NET.This variant is to design multiple convolution filters of different scales in U-NET to capture the structural information of MRI.The data consistency layer used in the model is to avoid data loss during the reconstruction process.At the same time,the model also adds a long skip link to combine the output of the last multi-scale convolution U-NET module with the zero input at the beginning.The filling image is combined to solve the problem of the disappearance of low-frequency information.The experimental results show that the reconstruction effect of the proposed method is better than other methods proposed in recent years compared in the experiment,and the reconstruction quality has been improved.(2)In order to solve the problem of the loss of high-frequency information in the reconstruction process,a deep frequency division network is proposed to perform compressed sensing reconstruction.The proposed method uses deep iterative reconstruction network to replace regular terms and corresponding parameters with stacked convolutional neural networks.Continuously cascade multiple deep iterative reconstruction network blocks into a deeper neural network.The data consistency layer is merged after each block to correct the kspace data of the intermediate result.The image content loss is calculated after each data consistency layer,and the frequency division loss is obtained by weighting the high-frequency loss and low-frequency loss after each deep iterative reconstruction of the network block.The experimental results were analyzed qualitatively and quantitatively,and analyzed from the visual and evaluation indicators.Compared with the comparison method,the proposed method has improved the reconstruction quality.This article hopes to study CS-MRI reconstruction based on deep learning methods,using the latest artificial intelligence technology to improve the quality of reconstructed images,assist medical diagnosis,and promote the development of smart medical construction.