An Unsupervised Adaptive Non-local Means Filter and Its Application on Partially Parallel Magnetic Resonance Imaging
by
Weihong Guo
Department of Mathematics, the University of Alabama
Coauthors: Feng Huang

Partially Parallel Magnetic Resonance Imaging (PP-MRI) techniques have been widely used clinically to reduce scan time. However, it results in increased noise and artifact level. Clinically, a reliable technique is expected to efficiently reduce the noise and artifact level while preserving important salient information like edges and fine structures. Non-local means (NL-means) proposed by Buades et al. in 2005 is a robust and efficient image denoising technique. It assumes that the to-be-smoothed image has spatially homogeneous noise distribution, and hence a universal constant parameter h controlling the global amount of smoothing is used. Unfortunately, the assumption is not true for images reconstructed from PP-MRI, which have spatially variant noise/artifact levels. We provide an enhanced NL-means variant through three approaches: 1) utilizing local mutual information,local variance together with k-means segmentation to separate the image domain into three sets without any supervision: edges (E), regions with artifacts and high level noise (H) and others (O); 2) automatic and adaptive definition of h in each of these three sets; 3) applying the Gaussian adapted NL-means variant to real and imaginary part of the complex MRI images separately rather than the magnitude image which has Rician distribution.

With phantom data, the proposed method was compared with the plain NL-means, competitive state-of-the-art adaptive anisotropic diffusion and total variation minimization method. Quantitative and qualitative results show that the unsupervised adaptive NL-means generates the highest peak signal to noise ratio, the lowest relative error with respect to ground truth. It also does the best in sufficiently removing noise and artifacts while preserving the edges and fine structures. Finally, results on many human brain data sets acquired with different acquisition parameters demonstrates the robustness and applicability of the proposed method.

Date received: January 29, 2009