Magneto-Resonance (MR) images are believed to have Rician distributed noise. In this talk, we propose two variational models involving total variation (TV) regularization to denies images corrupted by Rician distributed noise. For the first model, we implement the L2 and Sobolev H1 gradient descent methods in our numerical simulations on synthetic 3D MR images of the brain. In addition, we show the existence of a minimizer and a maximum principle result. For the second model, we incorporate the image formation model in the data fidelity term together with the Rician noise assumption. We perform numerical experiments on High-Angular Resolution Diffusion Imaging (HARDI) data of the brain to show the validity of the proposed model.