fMRI Denoising Techniques

An educational talk by Lars Kasper, presented as part of the ISMRM 2020 Weekend Course: fMRI Across Spatial & Temporal Scales. Noise as "signal fluctuation of no interest" is the eminent adversary when studying brain function. First, it incurs sensitivity loss for our small effects of interest by drowning them in un(cor)related fluctuations. Second, noise may correlate with effects, reducing specificity or increasing false positives by conflating them with fluctuations of non-neuronal origin. Here, we revisit how both the MR scanner and the subject generate noise in fMRI time series through different pathways, namely as thermal noise, encoding noise (magnetic field), and physiological noise of different origin. Specifically, the impact of motion, cardiac and respiratory physiology is scrutinized. We structure different approaches to noise mitigation following the recycling waste hierarchy, which also applies to sustainable science: avoid, reduce, reuse. Both data-driven (GSR, ICA, FIX, AROMA) as well as model-based (RETROICOR, HRV, RVT) denoising techniques are introduced.