Rician Noise Removal in Diffusion Kurtosis Imaging

Nanjie Gong¹, Chun Sing Wong¹, Sau May Wong², Bart ter Haar Romeny²

In the past decade, many denoising algorithms have been proposed for MR images. Most often the Gaussian filter is used for this purpose. However it has already been proven that noise on the magnitude MR images is not Gaussian, but Rician distributed. Gaussian filters are used in general for the reason that the noise distribution of magnitude MR images with high signal-to-noise ratio (SNR) can be well approximated with a Gaussian distribution. Diffusion Kurtosis Imaging (DKI), an extension of Diffusion Tensor Imaging (DTI), generates images with a lower SNR compared to images obtained with DTI due to the need of larger and more b-values. Subsequently the noise distribution in obtained DKI MR images can no longer be treated as Gaussian distributed. In this study, two Rician filters, namely the non-local mean Rician filter (NLM) and the non-local maximum likelihood filter (NLML), and Gaussian filter are investigated on their denoising performance on DKI MR images. Results obtained from the simulations show more accurate derived parameters from DKI images denoised with the Rician filters than those from images denoised with the Gaussian filter. In addition, regarding the parameters� value derived from real human data, there are significant differences between those from data filtered with the Rician filters and those from data filtered with Gaussian filter.