Improved NORDIC denoising for submillimetre BOLD fMRI using patch formation via non-local pixels similarity - pixel-matching (PM) NORDIC

Submillimetre BOLD fMRI enables studying brain function at the mesoscopic level but is limited by low SNR. The NORDIC PCA algorithm reduces thermal noise levels in fMRI in a patch-wise manner via singular value thresholding (SVT). However, the NORDIC patch formation uses adjacent pixels that often contain signals from multiple tissues, which can degrade the denoising performance. We propose an alternative patch formation using the similarity between non-local pixels, dubbed pixel-matching (PM) NORDIC. PM-NORDIC outperforms standard NORDIC in terms of temporal SNR and spatial smoothness estimates.