A vectorial total variation model for denoising high angular resolution diffusion images corrupted by Rician noise

Abstract

The presence of noise in High Angular Resolution Diffusion Imaging (HARDI) data of the brain can limit the accuracy with which fiber pathways of the brain can be extracted. In this work, we present a variational model to denoise HARDI data corrupted by Rician noise. We formulate a minimization model composed of a data fidelity term incorporating the Rician noise assumption and a regularization term given by the vectorial total variation. Although the proposed minimization model is non-convex, we are able to establish existence of minimizers. Numerical experiments are performed on three types of data: 2D synthetic data, 3D diffusion-weighted Magnetic Resonance Imaging (DW-MRI) data of a hardware phantom containing synthetic fibers, and 3D real HARDI brain data. Experiments show that our model is effective for denoising HARDI-type data while preserving important aspects of the fiber pathways such as fractional anisotropy and the orientation distribution functions.