Research On Magnetic Resonance Images Denoising Based On Structural Similarity And Low Rank Sparsity

Magnetic resonance imaging(MRI)have great application value in the medical field,but it is easily polluted by noise in imaging process,which affects the subsequent processing and analysis of MR images by doctors.Image denoising is the most important way to reduce the negative effect of noise on MR images.However,due to the complexity of noise in MR images,it is still a difficult task to denoise MR images,such as easily losing details and introducing artifacts.For MR images,we study the image denoising and propose denoising algorithms of MR images based on structural similarity and low rank sparsity.MR images have structural similarity,which means that there are many similar patches in MR images,such as repeated image edges or tissue textures.Low rank sparsity means that the high-rank noisy image information is converted into sparse low-rank image information by low-rank matrix approximation,and noise is eliminated in this process.The main contribution of this thesis is summarized as follows:(1)We proposed a denoising algorithm for MR image based on adaptive structureinformation,which can be used to remove Gaussian noise.The proposed algorithmextracts the structural information of MR images according to the structural similarity,and uses the structural information to guide the weight calculation,so as to avoid thenegative impacts of different pixels on the denoising process.The experimental resultsshow that the proposed algorithm reduces the introduction of artifacts,can betterpreserves image details,and achieves higher PSNR and SSIM values.(2)We proposed an adaptive denoising algorithm for MR images based on non-localstructure similar information and low rank sparsity,which can be used to remove thestationary and spatially varying Rician or non central Chi distributed noise.In thisalgorithm,structural similarity is used to extract structure similar 3D non-local patchesfrom MR images,and then uses our improved weighted nuclear norm minimization toperform low-rank sparseness on these similar 3D patches to achieve adaptiveprocessing of noise.The experimental results show that the proposed algorithm hasachieved better results in visual quality and quantitative indicators such as PSNR.(3)We proposed a denoising algorithm for diffusion weighted MR images based onnon-local and multi-directional structure similar information,which can be used toremove non-Gaussian noise,such as Rician noise.This proposed algorithm fuses thenon-local and multi-directional structure similar information of diffusion weighted MRimages to extract more structure similar 3D patches.Then,these image patches withsimilar structure are low-rank sparse represented by the improved weighted nuclearnorm minimization which proposed in Chapter 3,so as to remove noise.In addition,the proposed algorithm also uses phase correction technology for pre-processing,thereby reducing bias in denoising process.The experimental results show that theproposed algorithm has better denoising performance than the compared algorithms,especially in evaluation indicators such as PSNR and visual quality.For MR images denoising,this thesis proposes corresponding algorithms.The proposed algorithms can improve the quality of MR images,and further improve the ability of doctors to analyze and process MR images,which is helpful in the process of medical diagnosis.