Undersampled MR Image Reconstruction Using An Enhanced Recursive Residual Network

Magnetic Resonance Imaging(MRI)is a non-invasive diagnostic technique providing multi-modalities of internal body tissues,which has been widely applied for numerous clinical applications.However,its long acquisition time causes discomfort to patients and thus hinders the time-critical applications.Moreover,subject movement also brings in artefacts to the image.Accordingly,many algorithms has been proposed to reconstruct the undersampled MRI,i.e.,Compressed Sensing(CS)is utilized to reconstruct the undersampled k-space data and Super Resolution(SR)is used to reconstruction the high resolution image from the low resolution measurement,so as to speed up the acquisition time.The conventional optimization-based methods has shown good performance to solve the image reconstruction problems.Nevertheless,it is very difficult for the conventional methods to achieve an acceptable image quality when using aggressive undersampling rate less than 20%.Motivated by the outstanding performance of deep learning in computer vision,we proposed an Enhanced Recursive Residual Network(ERRN)for the undersampled MR image reconstruction,including CS-MRI and SR-MRI.The proposed ERRN consisted by three subnets of embedding,inference and reconstruction,is based on the prevail recursive residual network,and enhanced by a set of user-design functional modules,i.e.,high-frequency feature guidance,error-correction unit and improved dense connections.First,the high-frequency feature guidance is designed to guide the network training and prediction of underlying anatomy by the image a prior,playing a complementary role to the residual learning,thus constraining the solution space.Second,the error-correction unit is used to interpret the data fidelity term of conventional optimization methods for the ERRN.According different undersampling schemes in applications,an error-correction unit of data consistency layer is adapted for CS-MRI to emphasize the data fidelity in k-space,and the back projection module is used for SR-MRI to enforce relationship between the low resolution measurement and high resolution reconstruction in spatial domain.Third,the ERRN is equipped with the improved dense connections,which strengthen feature propagation and encourage feature reuse in the deep network.Considering the recursive learning strategy where each residual block in ERRN shares the same parameters,our dense connections only assign a trainable scalar weight to the previous blocks and take their weighted average as the input of the current block,to reduce the parameter number and computation.The ERRN was evaluated using both real-valued and complex-valued MR data,and both healthy and pathological MR data,demonstrating that our ERRN could achieve the best performance at all sampling rates of difference undersampling strategies,compared to the optimization methods as well as the state-of-the-art CNN networks with similar depth but less parameters.Furthermore,we also introduce the local and global dense connections to construct an Enhanced Cascading Residual Network(ECRN),to overcome the limitation of training deeper network by the recursive learning and improve the network performance.