Collaborative non-local means denoising of magnetic resonance images

Noise artifacts in magnetic resonance (MR) images increase the complexity of image processing workflows and decrease the reliability of inferences drawn from the images. To reduce noise, the non-local means (NLM) filter has been shown to yield state-of-the-art denoising performance. However, NLM relies heavily on the existence of recurring structural patterns and this condition might not always be satisfied especially within a single image, where complex patterns might not recur. In this paper, we propose to leverage common structures from multiple images to collaboratively denoise an image. The assumption is that, although the human brain is structurally complex, common structures can be found with greater probability from multiple scans than from a single scan. More specifically, to denoise an image, multiple images from different individuals are spatially aligned to the image and NLM-like block matching is performed on these aligned images with the image as the reference. Experiments on synthetic and real data indicate that the proposed approach - collaborative non-local means (CNLM) - outperforms the classic NLM and yields results with markedly improved structural details.