Does only the time-series data of fMRI matter or does spatial distribution of the signal values also matter?

Question

I was denoising fMRI data using GLM and had a question regarding the “data” of fMRI in general.

(For reference you can look at the “GLM” part of https://fmriprep.org/en/stable/outputs.html or https://www.sciencedirect.com/science/article/pii/S1053811917310972)

Below is a time series of the before/after GLM for a given voxel:

As you can see, the raw values (actually we went through a preprocessing step prior to GLM denoising) had a value of around $600\pm 50$, while the residual (i.e. denoised data) had a value around $0\pm 50$. In other words, it seems that via GLM, the raw data was somewhat “centered”. At first I thought this was OK, as I thought that only the fluctuations of the data matter (i.e. I had this assumption that mean centering doesn’t remove any information in the data). However, I realized that when we did this, the spatial distribution of signal values before and after GLM changes, as the pictures below show (the picture below is the distribution of fMRI signal for a given time point before and after GLM denoising) :

Is this OK?

In other words,

• does only the “fluctuations” (across time) of the data matter in fMRI (i.e. is adding a constant to an fMRI time series data irrelevant?).
• if so, then since adding a constant to an voxel’s time series data is irrelevant, it stands to reason that if I add different constant values to different voxels (for example as part of a detrending step), the spatial distribution of values before and after would change. Is this OK? (In other words, does the distribution of fMRI signals for a given time point meaningless?)

Thank you :)

Answer 1

Usually, but not always, fMRI timeseries are used to compute functional connectivity by calculating Pearson correlations among voxels or ROIs.

The equation for the Pearson correlation can be written as:

$$\rho_{X,Y} = \frac{\mathbb{E}\left[ (X-\mu_X)(Y-\mu_Y)\right]}{\sigma_X \sigma_Y}$$

(screenshot)

As you can see, the numerator involves subtracting the mean of $X$ from $X$ and the mean of $Y$ from $Y$; when you're computing Pearson correlations on a large scale it's typically most computationally efficient to first $z$-score the data, so that the mean is 0 and SD is 1, which saves a couple computational steps for every pairwise comparison and turns it into just $\text{mean} \ (X \cdot Y$).

Changes in the spatial distribution caused by your denoising are different than the "centering" part of the transformation, though.