Jelle Veraart1,2, Dmitry S. Novikov2, Jan Sijbers1, and Els Fieremans2
1iMinds Vision Lab, University of Antwerp, Antwerp, Belgium, 2Center for Biomedical Imaging, New York University School of Medicine, New York, NY, United States
We here adopt
the idea of noise removal by means of transforming redundant data into the
Principal Component Analysis (PCA) domain and preserving only the components
that contribute to the signal to denoise
diffusion MRI (dMRI) data. We objectify the threshold on the PCA eigenvalues for
denoising by exploiting the fact that the noise-only eigenvalues are
expected to obey the universal Marchenko-Pastur (MP) distribution. By doing so,
we design a selective denoising technique that reduces signal fluctuations
solely rooting in thermal noise, not in fine anatomical details.
