A new paper in Neuroimage suggests that methods for removing head motion and physiological noise from fMRI data might be inadvertently excluding real signal as well.
The authors, Molly G. Bright and Kevin Murphy of Cardiff, studied the technique called nuisance regression. It’s a popular approach for removing fMRI noise. Noise reduction is important because factors such as head movement, the heart beat, and breathing, can contaminate the fMRI signal and lead to biased results. Nuisance regression works by estimating parameters, e.g. head position at each point in time, and statistically controlling for these using regression.
In theory, nuisance regression should leave the ‘real’ fMRI signal, the brain activity, untouched, and just get rid of the unwanted noise. However, Bright and Murphy show that this isn’t the case.
Here’s a comparison showing the remarkably similar spatial structure of the “cleaned” fMRI data (after noise is removed), in green, and of the “noise” that was removed, in red. All of the classic networks are present, such as motor cortex, visual cortex and default mode:
Does this mean that these “brain networks” are actually – at least partly – artifacts of head motion and physiology? If so, it would be huge. It would force neuroscientists to rethink the whole of brain organization.
This worrying possibility was raised by a comment on PubPeer. Luckily, I don’t think that’s correct. Bright and Murphy go on to show that the same spatial network structure emerges, even if you use randomized noise parameters. In fact, even if you borrow the parameters from another person‘s scan, and regress those out, the same networks are seen in the noise variance.
What this suggests is that (almost) any way of partitioning fMRI data into “clean” and “noise” will leave the same spatial structure in both. This is less scary than the PubPeer interpretation. Phew.
But the paper does have important implications. For one, it looks like there’s a limit to the useful number of nuisance regressors. For motion regression, for instance, various researchers have suggested using 3, 6, 12, and 24 parameters to remove motion artifacts. However, Bright and Murphy show that once you go above 6 parameters, the additional regressors don’t account for any more variance than randomized parameters do. 6 parameters seems to capture most of the ‘real’ motion in this dataset.
Using excessive true noise regressors related to head motion may be detrimental, removing amounts of variance similar to that of simulated noise regressors while affecting degrees of freedom such that interesting “signal” is removed from the resting state data. This balance between de-noising and signal loss can be adjusted by considering datasets with greater numbers of time-points… It remains an open challenge to identify when nuisance regressors are no longer beneficial and become an added confound themselves in resting state fMRI.
Edit: Although note that this is only true of the dataset Bright and Murphy examined, which includes 160 volumes (time-points). The authors go on to say that in longer data-sets, “the problems associated with excessive regression could be attenuated” – in other words, the more data you have, the more motion regressors it makes sense to use.
Bright MG, & Murphy K (2015). Is fMRI “noise” really noise? Resting state nuisance regressors remove variance with network structure. NeuroImage PMID: 25862264