Performance Evaluation of Nonlinear Spatial Filters on MR Images

Abstract

Noise is one of the major factors which adversely affect the quality of Magnetic Resonance (MR) images. Performance of denoising filters is always image specific and their selection is a hard task. No reliable guidance for the selection of denoising technique is available so far. Hence, the performance of nonlinear spatial filters like Bilateral Filter (BF), Nonlocal Mean (NLM), Smallest Univalue Segment Assimilating Nucleus (SUSAN), Total Variation (TV), Kuwahara, Anisotropic Diffusion (AD), Wavelet thresholding and Linear Minimum Mean Square Error (LMMSE) filters is evaluated on MR images, in this article. The comparison is performed both on Shepp-Logan phantom images based on Peak Signal to Noise Ratio (PSNR) and on real-time MR images based on a Denoising Quality Matric (DQM). From the comparison, it has been noticed that at low noise levels, BF and NLM exhibit better performance than other filters. However, PSNR exhibited by BF and NLM drastically falls at high noise levels. The performance of TV filter is observed to be consistently good at low and high noise levels.