Fast Non-local Means Denoising for MR Image Sequences

Denoising algorithms are used for the enhancement of magnetic resonance (MR) images. MR images possess more structural details compared to normal images which have to be preserved for better diagnosis. Non-local means (NLM) filter is proved to be the best in preserving edges, however demands higher computations. From diagnostic perspective, the denoising algorithms should be computationally fast and accurate. The aim of this paper is to improve the accuracy and computational efficiency of unbiased NLM filter for MR image sequences. In this work we propose to do so by useful alternative of NLM technique in-conjunction with principal components for Rician noise. The variants of PCA based denoising techniques developed so far compute PCA locally for each image, thus decreasing computational efficiency. In this work we propose to compute PCA only once globally for each shot. A modified preprocessing step of shot boundary detection is employed to segregate 3D MR sequences in various shots based on its content similarity. Further denoising is carried out in non-local means framework with reduced dimensionality. We compare results with the existing NLM based MR denoising techniques and show that the proposed method is competitive in terms of attaining higher accuracy and computational efficiency. The performance of the proposed algorithm is evaluated with PSNR, SSIM and visual perception.