Interpolating values with 3dTproject -cenmode NTRP option

AFNI Message Board
1

Hi fellow AFNIers,

I have a question regarding the functioning of 3dTproject.
We processed some resting state fMRI time series and wanted the time series of all participants to be of equal length while still taking care of outlier volumes. To achieve this we used 3dTproject after afni_proc.py to interpolate censored values using the -cenmode NTRP option.
I must admit I did not come up with this idea originally but it has been frequently used in our research institute.

Here is a code snippet so you can see what’s going on:


3dTproject -polort 0 -input $output_dir/${session_id}.results/pb04.${session_id}.r01.volreg+tlrc. \
      -censor $output_dir/${session_id}.results/censor_${session_id}_combined_2.1D -cenmode NTRP \
      -dsort $output_dir/${session_id}.results/Local_FSWe_rall+tlrc \
      -ort $output_dir/${session_id}.results/X.nocensor.xmat.1D \
      -prefix $output_dir/${session_id}.results/errts.${session_id}.fanaticor-interp

Now in review of the paper we applied this approach in the question was raised what happened if the neighboring volumes of the interpolated volumes were outliers as well, since those neighboring volumes are used to interpolate the censored value.
Would someone happen to know how many neighboring volumes are used to interpolate the censored volume and what would happen in the case outlined above (i.e., if two or more consequent volumes in a time series are outliers)?

Thank you all for your help!

Best,
Jonas Steinhäuser

2

Would someone happen to know how many neighboring volumes are used to interpolate the censored volume and what would happen in the case outlined above (i.e., if two or more consequent volumes in a time series are outliers)?

The two closest neighboring volumes (in time) are use – one before and one after, and the interpolation used is linear. For example, if time indexes 1 and 5 are kept, and time indexes 2,3,4 are censored, then


  s(2) = 0.75*s(1)+0.25*s(5)
  s(3) = 0.50*s(1)+0.50*s(5)
  s(4) = 0.25*s(1)+0.75*s(5)

where s(t) is the signal at time index t.

If s(5) is still an outlier (that is, wasn’t censored but “should have been”), then the interpolated values will be outlier-ish themselves. However, you have to judge the likelihood of this happening enough to contaminate your results significantly. No signal processing method (even “machine learning”) can perfectly suppress noise and outliers.

(As an aside, when I did resting state analysis [in my pre-retirement life], I usually used a fairly strict motion threshold for part of the censoring decision.)

The NTRP option was put there at the request of a particular Spanish researcher. Personally, I don’t like it, as it gives the impression that you have a full set of “real” data, whereas some of your data times series is now fictional – that is, you don’t actually have the number of degrees of freedom that it seems like from the time series length. However, if you are going to do inter-subject time series correlations, then you have to do something to make that possible.

3

Hi Bob,

thanks for returning from your well-deserved retirement to answer this question and providing insight into the -cenmode NTRP interpolation option! I have always wondered what the reference to the Spanish Inquisition in the -help file of 3dTproject meant… now I not only understand the method better but I know one more bit of AFNI trivia.

I agree that it’s not ideal to have the impression that you’re dealing with a data set of 100% real data but as you mentioned, for some inter-subject analyses it’s a necessity to have equal lenght of time series. But we will keep this limitation in mind for our analyses and transparently communicate the choice and method of interpolation in our manuscript.

Jonas