Nonlocal Maximum-Likelihood Algorithm Based On Discrete Cosine Transform And K-Means For Magnetic Resonance Images Denoising

Denoising of magnetic resonance(MR)images is an indispensable vital preprocessing step for improving clinical diagnostic accuracy,the quality of image manipulations such as registration and segmentation.The traditional denoising methods are applied to the magnetic resonance images denoising,with the increase of the noise level,it is obvious that the denoising effect needs to be further improved,and people struggle to visually accept the magnetic resonance images denoised by the traditional denoising methods.Researchers expect to have a more advanced denoising method,therefore,many researchers are devoted themselves to researching the denoising algorithms,they find that nonlocal means(NLM)denoising algorithm is applied to the magnetic resonance images denoising,regardless of the subjective or objective criteria,the denoising effect has a great improvement compared to the traditional denoising methods.In recent years,the improvement and optimization of nonlocal means denoising algorithms have been continuously developed,in the process of improvement,the noise in the magnetic resonance images was found to be a Rician noise.According to the special statistical characteristic of the Rician noise in the magnetic resonance images,the nonlocal maximum-likelihood denoising algorithm has been proposed for the magnetic resonance images denoising.The nonlocal maximum-likelihood denoising algorithm restores the true underlying intensity of the pixel being denoised by excluding dissimilar pixels,therefore,the nonlocal maximum-likelihood denoising algorithm has great potential.Our work is devoted ourselves to improving the nonlocal maximum-likelihood denoising algorithm.The main improvements include the following two aspects:(1)The measure of similarity is conducted in the discrete cosine transform(DCT)subspace of pixel neighborhood.However,in the classical nonlocal means denoising algorithm and nonlocal maximum-likelihood(NLML)denoising algorithm,the similarity measure is based on the intensity similarity of pixel neighborhood.Due to low data correlation and high energy compaction of discrete cosine transform,the accuracy of the selection of similarity samples is improved.(2)The samples are adaptively selected by using K-means clustering algorithm in the discrete cosine transform subspace of pixel neighborhood.However,a fixedsample size is used for the maximum-likelihood(ML)estimation in the classical nonlocal maximum-likelihood denoising algorithm.Quantitative and qualitative comparison experiments among the proposed filter,nonlocal maximum-likelihood filter,nonlocal means filter and variants of these filters are carried out on the magnetic resonance images.The comparison experiments indicate that the proposed filter achieves better capability of removing noise while preserving image details in MR images.Especially when the noise level in magnetic resonance images is higher,the denoising effect of the proposed filter is more significant.