Dictionary Learning For MRI Reconstruction

Magnetic Resonance Imaging is achieved by the in vitro high-frequency magnetic filed,which generates signals by radiating energy from the body material to the surrounding environment.MRI has many advantages,for example,ionizing radiation has neither radioactive damage nor biological damage to brain tissue.MRI can directly make cross-sectional,sagittal,coronal and various slanted tomographic images,unlike CT,there is no artifact like hardening radiation in MRI.Also,MRI has more extensive pathological process of showing disease as well as clearer structure.However,due to the limitations of physiology and hardware conditions,the main problem of MRI is the long time needed for scanning,which will result in patients' uncomfortableness and more expense.For this,many effective methods have been proposed to accelerate the imaging speed,like parallel imaging and Compressed Sensing MRI(CS-MRI).CS-MRI can reconstruct high-resolution images with randomly under-sampled k-space data by developing the sparsity of the image in some domain.In this thesis,a dictionary has been applied for the sparse representation of the image,then high-quality MR images can be reconstructed with only a little k-space data based on the compressed sensing theory.Two kinds of dictionaries have been applied,the first is low-layer dictionary,including fixed dictionary and adaptive dictionary,the second is deep-layer dictionary.The fixed dictionary,such as TV,which ignores some non-local information,can't suppress the artifacts and noise resulting from undersampling efficiently.While the adaptive dictionary has relatively better sparse representation ability,which can remove the artifacts and noise efficiently as well as keep some features at the same time,but a decreasing sampling rate will cause a low image quality.Then we adopt a deep-layer dictionary,which can be learned through convolutional neural networks,due to it has more useful priori knowledge,a relatively high image quality of the reconstruction can be ensured.Also,it takes only about 1s to reconstruct a 2D MR image with the acceleration of GTX 1080 GPU.