Research On Low-rank Matrix Decomposition MRI Denoising Algorithm Based On Image Prior

Magnetic Resonance Imaging(MRI)is an effective and useful tool for medical diagnosis.However,the noise generated during image acquisition or transmission damages the MRI quality and seriously reducing the accuracy of these diagnoses.The noisy,low-quality MRI image affects the accuracy of automated computer analyses,such as classification,segmentation,and registration.The texture and detail structure in MRI contain important medical information,which should be preserved as much as possible while removing noise.Therefore,research into MRI denoising has become significant for obtaining high-quality MRI output.Image denoising as an inverse problem is ill-posed.In order to get the true solution,it is necessary to limit the size of solution space.A common way to make the denoising problem well-conditioned is to introduce image prior in the model and obtain the corresponding regularization term to restrict the solution space.Denoising methods based on low rank matrix has become a research hotspot in recent years due to its good denoising performance.However,with the increase of noise level,the noise greatly influences the low rank of the image,and results in the denoising performance's insufficiency of the model.The research content of this paper is to introduce the prior information of MRI into the low rank matrix factorization denoising model to improve the denoising performance,so as to achieve the purpose of removing the noise in MRI and retaining the details of the MRI to the greatest extent.The main research contents of this paper are as follows:(1)Image patches contain abundant structural information,which can provide sufficient prior for denoising.This paper combines the prior of noise-free MR image patches with the prior of non-local self-similarity of MR image patches.We characterize the prior of noise-free MR image patches by using the Gauss mixture model through learning.Then,based on the non-local self-similarity,image patches are clustered under the guidance of the learned Gaussian mixture model,so as to maintain the low-rank of image patch matrix and improve the denoising effect of the low rank matrix decomposition model.The experimental results show that the Gaussian mixture model with noise-free MR image patches prior can effectively guide the clustering of noisy image patches and help to improve the denoising performance of low rank matrix decomposition algorithm.(2)In order to reduce the ringing artifacts caused by image patch aggregation in image reconstruction,the gradient sparsity of MRI is fitted by the Hyper Laplacian distribution,and a Hyper Laplacian regularization term based on gradient sparsity of MRI is added to the low rank matrix decomposition denoising model.The experimental results show that the algorithm with Hyper Laplacian terms in the model can effectively suppress ringing effect,and have an improvement of about 2.8% in peak signal-to-noise ratio.At the same time,several MRI denoising algorithms are compared with our method.The corresponding experimental results show that the low rank decomposition denoising algorithm based on MRI prior in this paper can retain the information of MRI itself to a large extent while removing noise effectively,and have a great improvement both in peak signal-to-noise ratio and structural similar image metric.