Studies On Magnetic Resonance Image Denoising And Related Algorithms

Magnetic Resonance Imaging (MRI) is one of the most commonly used imaging technique for clinical diagnosis and medical studies. It has the advantages of high tissue contrast, arbitrary imaging direction and no ionizing radiation. Due to the weakness of MR signal, improving image quality and accelerating the imaging process have long been hot topics in the field of MRI. Compared to other solutions, image denoising can effectively enhance the image quality without sacrificing the imaging speed and increasing costs. In this dissertation, several issues related to the MR image denoising are discussed. The details are as follow:1) Several MR image denoising algorithms were implemented and compared. Non-Local Means (NLM) was proved to be best for MR image denoising and thus we based our further work on NLM. Vector NLM, which was introduced to denoise Diffusion Tensor Imaging (DTI) images, was also implemented and modified. We also studied the existing NLM implementations and proposed a new way to implement NLM with a ten-fold acceleration. This greatly facilitates the studies and applications of NLM.2) Diffusion Kurtosis Imaging (DKI) image denosing. DKI is a new model of diffusion imaging with significant clinical value. Image denoising is crucial to DKI since DKI results are sensitive to image noise and DKI images normally suffers from low Signal-to-Noise Ratio (SNR). We proposed an efficient 4D NLM algorithm with a simplified weighted average range along dimension of different Diffusion Weighted images. While applying Vector NLM to DKI, we proposed several schemes to combine DKI images to make use of image redundancies in DKI images, grouping DKI images either with the same directions of diffusion gradient or with the same b-values in vectors. We also proposed a new DKI denoise algorithm by combining 4D NLM and Vector NLM. In order to evalutate denoise performance of different DKI images denoising algorithms, we synthesized a simulated dataset based on human brain DKI images, which is used to compare different algorithms quantitatively.3) Further improvement of NLM filter. First, we proposed a method to use information in residual image of denoising algorithm to improve denoising effect. The proposed method uses the image structure information in original image to extract remaining image signal from the residual image. The extracted signal is then added back to denoised image to compensate the detail loss due to denoising. Compared to the other methods, the proposed method can extract signals more effectively from residual images with ultra low SNR and thus can essentially improve the denoising results.4) Removing of spike noise in MR images. The malfunctions of MRI equipment may trigger spike noise in data, which will cause artifacts in MR images. We proposed a denoising method for spike noise based on compressed sensing (CS) technique. It uses the sparsity of MR images in transform domains to calculate the signal in K space which is corrupted by spike noise. The calculation process uses optimization algorithm to solve minimization problem of total variation and first order norm of wavelet coefficients. Compared to conventional interpolation method, the proposed method achieves better restoration of the data, especially when the spike noise occurs in low frequency region of K space.