Mixed-Model Noise Removal in 3D MRI via Rotation-and-Scale Invariant Non-Local Means

Mixed noise is a major issue influencing quantitative analysis in different forms of magnetic resonance image (MRI), such as T1 and diffusion image like DWI and DTI. Using different filters sequentially to remove mixed noise will severely deteriorate such medical images. We present a novel algorithm called rotation-and-scale invariant nonlocal means filter (RSNLM) to simultaneously remove mixed noise from different kinds of three-dimensional (3D) MRI images. First, we design a new similarity weights, including rank-ordered absolute difference (ROAD), coming from a trilateral filter (TriF) that is obtained to detect the mixed and high-level noise. Then, we present a shape view to consider the MRI data as a 3D operator, with which the similarity between the patches is calculated with the rigid transformation. The translation, rotation and scale have no influence on the similarity. Finally, the adaptive parameter estimation method of ROAD is illustrated, and the effective proof that validates the proposed algorithm is presented. Experiments using synthetic data with impulse noise, Rician noise, and the real MRI data confirm that the proposed method yields superior performance compared with current state-of-the-art methods.