Compressed Sensing MRI Reconstruction Based On Deep Convolutional Neural Network

Magnetic Resonance Imaging(MRI)is a medical imaging technology widely used in med-ical diagnosis,MRI has many advantages such as no radiation,providing more pathological and physiological information.However,owing to physical and physiological limitations,higher imaging quality(resolution,contrast and signal-to-noise ratio)and faster imaging speed have been the focus of research in the field of MRI.Compressed Sensing theory(CS)reconstruct MRI images from partially sampled data(CS-MRI)is an effective method to accelerate the speed of MRI imaging and improve the temporal and spatial resolution of imaging.However,traditional CS algorithms usually only have limited model capacity or spend a lot of time in the test phase.In recent years,deep learning technol-ogy has shown its potential in the CS field since it can effectively encode complex patterns in massive data.However,the current deep learning-based CS-MRI method only uses a single model to learn the residuals of the zero-reconstructed MRI image artifacts or only introduces the K-space correction term,ignoring the supervised information of the high-level semantic tasks.Besides,deep neural networks lose weak structural information with important diagnos-tic value when producing features layer by layer of input data.Therefore,this paper constructs an new end-to-end deep learning CS-MRI model by eflfectively combining segmentation priors into reconstruction networks,constructing dense connections between different reconstruction modules and introducing atrous spatial pyramid pooling.The main innovations of this paper are as follows:(1)By integrating the features from the segmentation task to the reconstruction model pro-viding the reconstruction process with more supervised information.The reconstruction model can understand the content of the input image and simplify the function mapping of the recon-struction process.(2)Dense connections are introduced between different modules of the reconstruction mod-el,achieving information complementation between deep and shallow modules,maintaining the independence and integrity in the information transmission process to avoid losing weak struc-tural information with important diagnostic value.(3)The Atrous Spatial Pyramid Pooling(ASPP)operation is introduced to stimulate the deep model to learn more meaningful features in both global and local perspectives,further improving the feature extraction ability for the reconstruction neural network and increasing the model receptive field.(4)By inputing the pre-trained brain tissue segmentation model with different CS-MRI reconstruction algorithms,our proposed model generates more discriminative textures for seg-mentation tasks and improves its analysis accuracy.By conducting a considerable amount of experiments on the brain tissue segmentation dataset MRBrainS 13,our MRI-RecNet is superior to the traditional optimization-based and existing deep learning-based algorithms in terms of subj ective and obj ective evaluation indexes.Besides,it verifies that segmentation-aware,dense connection,and ASPP operations can stim-ulate the reconstruction model to extract more recognizable features from limited K-space data and obtaining better reconstruction results.