Fast MR Imaging With Feature Refinement

Since its inception in the early 1970’s,Magnetic Resonance Imaging(MRI)has revolutionized radiology and medicine.It has become the most important imaging modality for clinical diagnosis and life science research,due to its excellent visualization of both anatomical structure and physiological function from the variety of contrast mechanism and no any ionizing radiation.However,the main limitation of MRI,affecting both clinical application and image quality,is the long scan time.Therefore,achieving fast imaging of MRI has significant scientific research and application value.At present,undersampling k-space to accelerating scan is an important strategy for fast MR imaging.In order to overcome the issue of image quality degradation caused by ill-condition due to undersampling,it is necessary to incorporate the prior information.Compressed Sensing(CS)based fast image is the hot topic in the field of prior-based fast MR imaging in past decade.Compressed Sensing indicates that the sparse signals can be reconstructed exactly from very few incoherent measurements by a nonlinear procedure.The sparsity of the image to be reconstructed is the fundamental premise of the successful application of CS theory and the prior to be used in MR imaging.Representing the image as sparse as possible can reduce the reconstruction error and improve the reconstruction quality.However,it is difficult to select an optimal sparsifying transform to maximize the sparse representation of the MR image.In practical applications,image cannot be completely represented sparsely,which leads to the loss of fine details.This thesis aims to overcome the issue of detail loss in the image reconstructed from partial k-space data,and ensure high reconstruction quality.The main works and contributions are as follows:1.A feature refinement strategy was proposed.Based on the CS theory and the composition of the residual image,an image feature refinement module was designed and successfully embedded in the framework of CS parallel imaging in MRI.Specifically,a powerful feature descriptor was developed to highlight the location of edge boundaries and enhance the image structural information extraction.The developed descriptor consists of two parts: texture estimator and structure estimator.Experiments show that with the texture and structure parts,the descriptor represents the useful details more effectively,so that the reconstruction method with the proposed feature refinement module has superior performance its original version,and the reconstructed image preserves more details.2.A general framework on combining the MR imaging model with deep learning was proposed.Based on the traditional imaging model and the iterative optimization algorithm,deep learning is introduced.And the iterative algorithm is unrolled to deep networks to learn the uncertainties of the imaging model and optimization algorithm from the training data.With gradually relaxing the constraints in the imaging model and further breaking the fixed structures of variables in the iterative algorithm,the learning ability of the deep network is fully utilized to learn more prior information,thereby improving the reconstruction quality and optimizing the detail information in the reconstructions.3.The feature refinement method and the model-based deep learning method have been successfully applied to 3D high-resolution vessel wall imaging.Compared with the traditional compressed sensing reconstruction,the feature refinement strategy can effectively improve the sharpness and contrast of the intracranial artery vessel wall.And the deep learning method can provide good depiction of the vessel wall while avoiding complicated parameter tuning of CS.