I am reading the tutorial:
And I notice they use files with beta suffix in the name, for example:
However, I dont see any beta suffix in the output of fmriprep.
What are these files? How are they produced?
When and why are they used for preprocessing rather the the “regular” BOLD signal?
The beta maps / files are files containing the beta estimates / coefficients after fitting a GLM model on your fMRI data. These beta values are the estimated regression coefficients.
Read more here and here.
fMRIprep only outputs processed imaging files and do not perform the GLM analysis.
GLM fitting using nilearn reference: Nilearn: Statistical Analysis for NeuroImaging in Python — Machine learning for NeuroImaging
@makaros622 Thank you!
Can u please elaborate as to when it is advised to use the beta values, and when to use to beta activations?
@Steven Thank you very much!
What I still not sure is in what of the more advanced analysis (e.g. RSA, encoding model, ISC, MVPA, etc) I should use the beta series as the BOLD data and when I should use the actual preprocessed BOLD?
All of these techniques will start of using the preprocessed BOLD. The beta images are a derivative of the preprocessed BOLD and are often what statistical inferences are based off of.
@Steven Assuming that the preprocessed BOLD includes StandardScaling, can u pls explain why the beta images are “often what statistical inferences are based off of.”?
Why is it not good enough to run these methods based on the preprocessed BOLD?
BOLD images are timeseries. When we make our statistical inferences, we typically use data from several timepoints, as opposed to making judgments on a timepoint-by-timepoint basis. The beta images we get often represent conditions that last for long periods of time.
@Steven, Do you mean because of the temporal dynamic of the HRF function?
That’s certainly part of it! But also fMRI images are noisy, so getting beta images representing long time points should help account for the noise.
And applying various preprocessing tools, such as smoothing_fwhm , or
nilearn.signal.clean with detrend, standardize and high/low pass - shouldnt be good enough?
It is not enough, but should definitely still be used.
Smoothing improves signal-to-noise, as averaging signal from nearby areas should keep the main shape of the signal while averaging out high frequency noise (at the expense of some spatial precision).
Detrending/filtering just make sure that the average baseline of the signal isn’t moving too much over the course of the scan.
A big problem with fmri is that meaningful signal change is really really small. So we need many volumes to have any sort of confidence that an effect we are observing is due to brain activity and not random noise.
So do you think that using the beta time-series should improve mainly sensitivity or specificity?
Such modeling is important for both. Modeling your data with a GLM (which outputs the beta images) allows one to separate components of the BOLD image that are presumed to be related to the conditions from those that are related to noise. The BOLD image, even after minimal preprocessing, will still be a less manageable mixture of signal and noise.
@Steven Can you maybe direct me to a complete example of how I transform from a BIDS dataset to a beta-series per subject? (In my design, every subject has a single session and a single run).
I found some parts but nothing complete. And I want to make sure I do this part good to prevent any future mistakes.
Nilearn has a lot of examples, so you can find the one that most closely matches your needs and adapt it.https://nilearn.github.io/auto_examples/index.html#glm-first-level-analysis-examples
@Steven Sorry I wasn’t clear enough. In my data, there is a single condition - this is a naturalistic stimuli so each subject just hear 8 minutes long story.
In all of the example you attached, the GLM is estimated based on specific contrast (e.g. Go vs. No Go, Face vs. House).
The question is, is there an example on how to convert the BOLD of the 8 minutes to beta time series w/o contrast?
Ah, you may find this site helpful instead: Neuroimaging Analysis Methods For Naturalistic Data — Naturalistic Data Analysis . Alternatively you can look at GLMSingle, which is a promising way to analyze rare events such as what you may find in naturalistic data.
You can also try generating beta series with Usage — NiBetaSeries 0.6.0 documentation