An Improved Non-local Means Denoising Technique for Brain MRI

Abstract

Noise artifacts introduced in Magnetic Resonance Imaging (MRI) due to imperfection in radio-frequency coils typically deteriorate the performance of automated analysis of a brain MRI. Thus, it is necessary to eliminate the noise for effective analysis of MRI. The denoising techniques that work on pixels vicinity (e.g., Averaging filter and Gaussian filter) suffer in denoising due to loss of salient information in brain MRI. As a result, these techniques gain poor Peak Signal-to-Noise Ratio (PSNR) in denoising of an MRI. However, the methods that denoise an image based on non-local correlation of patches show significant improvement in the performance. In this work, we proposed an improved denoising method for brain MRI based on non-local correlation of patches. The input image is decomposed into two components: periodic component and smooth component using Fast Fourier Transform (FFT). These two components of the image are denoised separately using non-local based averaging. The filtered image is obtained by reconstructing these two components. Quantitative and qualitative results on a T1-w simulated brain MRI dataset and a local brain MRI dataset indicate the effectiveness of the proposed method.