Phase-correcting non-local means filtering for diffusion-weighted imaging of the spinal cord

Abstract

Purpose: DWI suffers from low SNR when compared to anatomical MRI. To maintain reasonable SNR at relatively high spatial resolution, multiple acquisitions must be averaged. However, subject motion or involuntary physiological motion during diffusion-sensitizing gradients cause phase offsets among acquisitions. When the motion is localized to a small region, these phase offsets become particularly problematic. Complex averaging of acquisitions lead to cancellations from these phase offsets, whereas magnitude averaging results in noise amplification. Here, we propose an improved reconstruction for multi-acquisition DWI that effectively corrects for phase offsets while reducing noise. Theory and Methods: Each acquisition is processed with a refocusing reconstruction for global phase correction and a partial k-space reconstruction via projection-onto-convex-sets (POCS). The proposed reconstruction then embodies a new phase-correcting non-local means (PC-NLM) filter. PC-NLM is performed on the complex-valued outputs of the POCS algorithm aggregated across acquisitions. The PC-NLM filter leverages the shared structure among multiple acquisitions to simultaneously alleviate nuisance factors including phase offsets and noise. Results: Extensive simulations and in vivo DWI experiments of the cervical spinal cord are presented. The results demonstrate that the proposed reconstruction improves image quality by mitigating signal loss because of phase offsets and reducing noise. Importantly, these improvements are achieved while preserving the accuracy of apparent diffusion coefficient maps. Conclusion: An improved reconstruction incorporating a PC-NLM filter for multi-acquisition DWI is presented. This reconstruction can be particularly beneficial for high-resolution or high-b-value DWI acquisitions that suffer from low SNR and phase offsets from local motion.