Denoising in magnetic resonance imaging: theory, algorithms and applications

In the context of medical image processing, denoising is widely considered as one of most fundamental postprocessing tasks. In this field, the non-local means (NLM) filter demonstrated to be a robust and performing approach respect to the previous state-of-art denoising methods. As the filtering strength must be tuned to obtain an optimized and customized restoring process, the estimation of image noise variance is an important issue. Althought in clinical practice noise estimation is performed on background (no signal area) of magnitude MR images, in case of parallel MR imaging (pMRI) techniques noise estimation from the image background produces biased results due to spatially varying noise distribution of the pMRI images. A novel NLM approach based on local noise estimation is introduced (hereafter indicated as SVN-NLM). As second task, since the susceptibility-weighted imaging (SWI) suffers from reduced SNR due to the high resolution required to obtain a proper contrast generation, a novel pipeline (Multicomponent-Imaginary-Real-SWI, hereafter MIR-SWI) to obtain susceptibility-weighted images with higher SNR and improved conspicuity is proposed. In this context, the application of a denoising filter is non-trivial as the distributions of magnitude and phase noise may introduce biases during image restoration. Taking advantage of the potential multispectral nature of MR images, the multicomponent approach of the MIR-SWI approach performs better than a component-by-component image restoration method. Finally, a new strategy to address the computational demand of the NLM filter is investigated. Due to high computational complexity of the NLM denoising filter, in literature several 2D NLM implementations on Graphic Processor Unit (GPU) architectures were proposed. Here a fully 3D NLM implementation on a multi-GPU architecture is presented and its high scalability is suggested