Study Of MRI Reconstruction Based On Compressed Sensing And Deep Learning

Magnetic resonance imaging(MRI)is widely used in various clinical medical applications.It is non-invasive and provides quantitative measurements of tissue and excellent resolution to reveal different properties of the anatomy,that is very meaningful to early diagnosis and treatment of diseases.However,MRI is often limited by its relative slow acquisition speed and low reconstruction quality caused by sampling artefacts.What’s more,Dynamic Contrast-Enhanced MRI(DCE-MRI),as a special application of MRI,often suffers from motion artifacts.To solve these problems,we study the reconstruction algorithms about dynamic MRI and free-breathing DCE-MRI.Considering the demand of real-time imaging in clinical applications such as virtual surgery,we combine the deep learning with model-based MRI reconstruction method to further improve the Signal-Noise Ratio and reconstruction speed.Specially,the study includes the following three aspects:(1)To enable high quality reconstruction for free-breathing liver 4D DCEMRI,we presents a novel method called SMC-LS,which incorporates Sliding Motion Compensation into the standard L+S reconstruction to correct the respiratory motion(SMCL-LS).With sliding motion compensation,the reconstructed temporal frames are roughly registered before applying the standard L+S decomposition.The proposed method has been validated using in vivo free-breathing liver 4D MRI data and free-breathing liver 4D DCE-MRI phantom data.Results demonstrated that SMCLS reconstruction can effectively reduce motion blurring artifacts and preserve both spatial structures and temporal variations at a sub-second temporal frame rate for freebreathing whole-liver 4D DCE-MRI.(2)Many existing DCE-MRI reconstruction methods based on low-rank structure such as L+S often replace the rank function with the nuclear norm for the nonconvexity and discontinuous nature of the rank function,which is a suboptimal approximation.As a result,it limits the reconstruction quality of relative methods.We incorporate the Truncated Nuclear Norm Regularization(TNNR)which is mainly used in matrix complication into standard L+S method for multi-coil DCE-MRI reconstruction and propose Low-rank Plus Sparse matrix decomposition for DCE-MRI reconstruction based on TNNR(TNNR-LS).The proposed method has been validated on cardiac perfusion data,cardiac cine data and abdominal DCE-MRI data.Experimental results showed that TNNR-LS reconstruction achieved the best PSNR at different sampling ratio and different noise,compared to standard L+S method,which is based on nuclear norm.Furthermore,we also incorporate the TNNR into the SMC-LS proposed in section(1)and propose a novel MRI reconstruction,TNNR-SMCLS,to further improve the reconstruction quality of free-breathing DCE-MRI data.TNNR-SMCLS has been validated on free-breathing 4D dynamic MRI liver data and achieved a better result compared to SMC-LS.(3)The development of deep learning provides a novel way for real-time MRI reconstruction and fast imaging in clinical applications.With the aim of developing a fast algorithm for high-quality MRI reconstruction from undersampled k-space data,we propose a novel deep neural Network,which is inspired by Iterative Shrinkage Thresholding Algorithm with Data consistency(NISTAD).NISTAD consists of three consecutive blocks: an encoding block,which models the flow graph of ISTA,a classical iteration-based compressed sensing magnetic resonance imaging(CS-MRI)method;a decoding block,which recovers the image from sparse representation;a data consistency block,which adaptively enforces consistency with the acquired raw data according to learned noise level.The ISTA method is thereby mapped to an end-to-end deep neural network,which greatly reduces the reconstruction time and simplifies the tuning of hyper-parameters with a simpler network architecture with fewer parameters.NISTAD has been validated on retrospectively undersampled diencephalon standard challenge data using different acceleration factors,and compared with DAGAN,Cascade CNN and ISTA-Net,three state-of-the-art deep neural network-based methods.Experimental results demonstrated that NISTAD reconstruction achieved comparable image quality with DAGAN and Cascade CNN reconstruction in terms of PSNR metric,and subjective assessment.