A New Rician Noise Bias Correction

M. Nezamzadeh, I. Cameron Physics, Carleton University, Ottawa, ON, Canada, Radiology, The Ottawa Hospital, Ottawa, ON, Canada Introduction: It is well known that pixel intensities of magnitude MR images are biased by Rician noise [1-4]. This is particularly noticeable for SNR ≤ 2. Rician noise bias corrections from the literature [2,3] work well for SNR > 2 but not for SNR < 2. Furthermore, the distribution of these “corrected” data points is far from Gaussian. This can have significant implications if the data is further processed using least squares procedures, which assume a Gaussian distribution of data points. In this work a new and improved procedure for Rician noise bias correction for small SNR is presented. Theory: For SNR ≥ 3, the Rician PDF reduces to a Gaussian with a mean of 2 2 A M σ + = , where A is the “actual” signal strength and σ is the standard deviation. This led Gudbjartsson and Patz [2] to propose that A could be estimated using