Testing the distribution of nonstationary MRI data

An accepted model for MR image noise is a Gaussian distribution in the real and imaginary components of the complex valued image. We investigated a procedure for validating this model through repeated hypothesis testing. The procedure is relatively straight forward for the situation in which an image contains only noise. However there is an additional challenge when a signal component is added because the signal intensities tend to drift over time. We therefore extend the noise model by considering a time-varying mean, and then implement a procedure for modeling, estimating, and removing the drift component in order to test the underlying noise distribution. The results demonstrate consistency with the proposed noise model.