Dear Colleagues,
I would like to discuss the following questions pertaining to the so-called fMRI data scrubbing (http://www.humanconnectome.org/hosted/docs/Power-et-al-NeuroImage.pdf):
1. How and when do you think it is better to do scrubbing (e.g., Is it better to remove spikes before/after bandpass filtering, before independent component analysis or prior to functional connectivity analysis [for example, with mancovan]?)
2. How many volumes associated with "bad event" do you usually remove? (e.g., only 1 volume, or also neighbouring volumes)
3. How do you threshold your scrubbing? (e.g. FD-threshold=0.5mm)
4. Do you interpolate between your time-points afterwards? If so - how? (Nearest Neighbour, Linear, Cubic Spline)
5. When do you exclude a subject from the analysis completely? (e.g., how many volumes one should miss in %s?)
Any comments and suggestions are very welcomed.
If you are doing any sort of band pass filtering I would suggest doing your "scrubbing" before-hand. Mathematically, allowing a spike or sharp discontinuity to pass through a band-pass filter (Butterworth included) will introduce Gibb's ringing artifact across time which will bias your correlation coefficients to the positive. This is quite easy to show in numerical simulation.
To do my interpolation I do a Loess smoothing (with the time point removed) and then a cubic interpolation for the missing data point.
Thank you very much for your comment Robert!
How many volumes associated with 1 spike do you usually remove? Is it 1 volume before and 1 after the event?
What I want to add is that any motion correction may change the real signal, including motion Scrubbing. You should ask your participants to try their best to minimize the motion during scanning. In data mining stage, you would rather use data with appropriate head motion, than tangle in the choice of different head motion correction methods.
Just my personal opinion.
1. I do "Motion scrubbing" or "Motion censoring" in a non-destructive manner. By that, I mean that I never delete the high motion volumes but I ignore them in my regression analyses (nuisance regression, connectivity calculations, or activation calculation following your tasks...).
2. Depending on your TR, I would either ignore the one TR or the one TR plus one before.
3. For Task-fMRI, I have seen thresholds ranging from 1mm to 0.5mm while for resting state my choice is usually 0.25 or 0.2mm
4. I believe that the function I use for my fMRI analysis only ignores the time without removing them so no interpolation seems necessary. I believe this is true because we use scrubbing based on volumes and not voxels.
5. I think it is getting more and more common to exclude a subject if more than 50% of the volumes are censored and/or less than 5 minutes of data is remaining (based on resting state fMRI). Now if you have only have a 5 minute run, I would probably look at the distribution of how many volumes are censored for each subject and pick a threshold that keeps most of the subject (I don't think I would go lower than 75% of the run remaining).
I definitely agree with Robert about needing to remove large spikes before temporal filtering. I have examined my own resting state with and without doing this step, and removing spikes before filtering makes a substantial difference (in my data, anyway). What I'm not entirely clear on is why you need to add back interpolated scans to replace the ones that were censored in order to bandpass properly? I assume also that you remove the interpolated scans after filtering is complete?
Hi Steffie,
I agree. On the one hand, interpolating between time-points after band-pass filtering might be more accurate, as in this case you are working with "cleaned" data... But on the other hand, as Robert pointed, if you introduce any "artificial drifts" or sharp discontinuity (caused either by interpolation or volume removal, especially when many "neighbouring" volumes are missing) into the filtered data, this may certainly influence correlation coefficients (I observed this when working with ICA).
I've eventually switched to voxel-specific head motion correction (Satterthwaite et al., 2013).
P.S.:
Here is a good report comparing different strategies for motion correction and scrubbing: http://cdn.f1000.com/posters/docs/250841110
P.P.S.:
Re adding back interpolated scans.. I might have phrased that wrong. I just wanted to say that some researchers remove 3 volumes in total: 'bad event' (1) + neighbouring volumes (2) [Of note, I used to remove only 1 volume].
FSL Motion Outliers
The tool fsl_motion_outliers is designed to detect timepoints in an fMRI dataset that have been corrupted by large motion. It creates a confound matrix that can be used in the GLM to completely remove the effects of these timepoints on the analysis, without any adverse affects in the statistics. This is intended to deal with the effects of intermediate to large motions, which corrupt images beyond anything that the linear motion parameter regression methods can fix.
Source: http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/FSLMotionOutliers
Has someone observed that, when "many neighbouring volumes are missing", said by Alexander, spurious artifact will show up if you use spline interpolation to "recover" the censored volumes? I met this cases several times. So, strongly recommend using linear interpolation; and, be ware of "many neighbouring volumes are missing" case by eyeballing all data by yourself.
Thanks for this post! I was wondering how conservative are most people with scrubbing? Do they mark only large striping, or do they mark any note of striping?
Are there pros or cons of being more or less conservative?
Some use FD > 0.5, some use FD > 0.25 to identify bad frames. Different "FD" algorithms differ in conservativeness. FD_power seems more often used. Some may also calculate meanFD (across all frames left), and exclude subjects with meanFD > 0.15 (can be very stringent). But excluding subjects with too many frames being "censored" (e.g., remaining data < 4.5 min) may be always conducted. People may also remove frames before and after the bad ones. To me, eyeball your time series (plot all regions mean time series in one figure) or plot an FC matrix to get an initial impression of your data could be necessary. Less conservative motion artifact removal may lead to globally high FC values in FC matrix, indicating very similar time series across different brain regions. More conservative could be a good idea but, when head motion is related to your researching problems (AD&D child tend to have more head motion), it's another issue. At last, when you do dynamic FC analysis, censoring data is not a good idea coz it will damage the temporal continuity. My two cents. Take with cautions.
You can use the Frame-Wise Displacement (FD) as defined by Power et al.2012 (doi: 10.1016/j.neuroimage.2011.10.018).
For resting-state data, a threshold of FD=0.5mm is commonly used and we remove 2 volumes before and 1 after the motion spike (=bad volume detected).
Thanks for your answers! Any recommends about the threshold to exclude subject's date based on meanFD in task fMRI?
- Badges
- Science topic