Phase-correcting denoising for diffusion magnetic resonance imaging

Diffusion magnetic resonance imaging (MRI) is a low signal-to-noise ratio (SNR) acquisition technique when compared to anatomical MRI. Multiple acquisitions have to be averaged to overcome this SNR problem. However, subject motion and/or local pulsations during diffusion sensitizing gradients create varying phase offsets and k-space shifts between repeated acquisitions, prohibiting direct complex averaging due to local signal cancellations in the resultant images. When multiple acquisitions are magnitude averaged, these phase issues are avoided at the expense of noise accumulation. This thesis proposes a reconstruction routine to overcome the local signal cancellations, while increasing the SNR. First, a global phase correction algorithm is employed, followed by a partial Fourier reconstruction algorithm. Then, a novel phase-correcting non-local means (PC-NLM) filtering is proposed to denoise the images without losing structural details. The proposed PC-NLM takes advantage of the shared structure of the multiple acquisitions as they should only differ in terms of phase issues and noise. The proposed PC-NLM technique is rst employed on diffusion-weighted imaging (DWI) of the spinal cord, and is then modi ed to capture the joint information from different diffusion sensitizing directions in diffusion-tensor imaging (DTI). The results are demonstrated with extensive simulations and in vivo DWI and DTI of the spinal cord. These results show that the proposed PC-NLM provides high image quality without any local signal cancellations, while preserving the integrity of quantitative measures such as apparent diffusion coefficients (ADC) and fractional anisotropy (FA) maps. This reconstruction routine can be especially beneficial for the imaging of small body parts that require high resolution.