Magnetic Resonance Imaging Noise Filtering using Adaptive Polynomial-Fit Non-Local Means

Every image, whether it is a Magnetic Resonance (MR) image or a gray scale image, usually contains noise, which negatively affects image processing and analysis outcomes. For MR images, noise can be induced by environmental, equipment, and human factors. Rician noise obeys a Rician distribution. It degrades the quality of an image and makes it blurry. Rician noise is signal-dependent. Thus, it is a difficult task to separate signals from noise. In order to reduce Rician noise in MR images, noise-removing techniques are necessary to be applied before the image undergoes further processing. In this paper, a noiseremoving technique is developed by cascading a new noise estimation method known as Nonlinear Spatial Mean Absolute Deviation (NSMAD) with a new noise filter known as Adaptive Polynomial-Fit Non-Local Means (Adaptive_PFNLM) filter. The NSMAD method is used to estimate the level of noise standard deviation in MR images. Then, the value of noise standard deviation is passed to the Adaptive_PFNLM filter to remove noise. The NSMAD method is compared with three existing estimation methods, namely Brummer's method, Maximum Likelihood (ML) method, and Local Mean method. The Adaptive_PFNLM filter is also compared with three existing filters, namely Non-Local Means (NLM) filter, Linear Minimum Mean Square Error (LMMSE) filter, and Polynomial-Fit NonLocal Means (PFNLM) filter. The comparison is evaluated by using the mean absolute error (MAE), signal-to-noise ratio (SNR), mean square error (MSE), peak signal-to-noise ratio (PSNR), structure similarity (SSIM) and quality index (Q). The results indicate that NSMAD and Adaptive_PFNLM perform better than the existing noise estimation methods and noise filters.