Denoising of MR images using adaptive multiresolution subband mixing

Abstract

In this paper we are proposing an algorithm based on Blockwise Non-Local (NL-) means filter and Dual Tree Comlex Wavelet Transform (DTCWT) for 3D Magnetic Resonance (MR) image denoising. The idea of the proposed filtering is adaptive subband coefficient mixing. The image is filtered using blockwise NL-means filter with two different sets of filtering parameters. The first set parameters were chosen to preserve the features in the image i.e less smoothing and feature oriented and the second set will do more smoothing i.e noise oriented. Finally these two images are fused using DTCWT and adaptive subband coefficient mixing to remove the noise while preserving the sharp details in the image. The filter is designed for removing Gaussian and Rician noise from the image volume. Quantitative validation of the proposed method was carried out on Brainweb datasets by using several quality metrics. The results show that the proposed filter performed well than the recently proposed filters based on 3D Discrete Wavelet Transform, Rician NL-means filters. The proposed filter removes noise effectively while preserving fine structures such as edges, lines etc. even for very noisy cases.