Deep Learning Based Fast MRI Reconstruction

Magnetic resonance imaging(MRI)is non-invasive and radiation-free.It is wildly used in clinic diagnosis for in vivo detection.Patients must be still during imaging process.Following the Nyquist's criterion,fully sampling will take a long time.To reduce the sampling time,subsampling is used,which leads to aliasing artifact.MRI reconstruction aims to reconstruct clear image from sub-sampled K-space data,which is important in clinic diagnosis.Current methods fail in extremely low sampling rate,and some methods have weakness in time-consuming.Based on deep learning technology,this paper made the following contributions.First,we proposed a novel convolution neural network(CNN)architecture called Cascade Dilated Dense Network(CDDN).Existing network architecture is too simple to fit the mapping from aliasing images to clear images.The aliased images are reconstructed step by step with cascade dense de-aliasing modules.Dilation convolution is introduced to expend the receptive field without additional parameters.The complex aliasing artifact can be effectively removed with the combination of dense connection and dilated convolution.Second,we proposed Two-step Data Consistency(TDC).In general,there are two approaches to exploit the raw phase-encoding data from k-space in deep learning methods: loss functions based on frequency domain and simplify data replacement.The later method can ensure the data consistency but only corrects the data on sampled location.We convert the complex-valued result from replacement to real-valued images and applied another replacement,so that the data on other locations can be corrects.TDC can ensure the consistency between raw k-space data and reconstructed data as well as reconstructing natural result.Finally,extensive experiments demonstrate that the proposed method can reconstruct accuracy result in various sampling rate,even aggressive one(5%).We take experiments on both cardiac and knee MR dataset,and our method achieve the state-of-art result in measure criterions like PSNR and SSIM.These experiments proved the robustness and the efficiency.Our work has important significance in both deep learning and medical image process field.