TENT scaling

AFNI Message Board
1

Hi,

I've been reading Gang's recent paper on HRF modelling and it inspired me to investigate what effect changing HRF model might have on some task data I have.

It's a block task consisting of 60s rest, {20s stimulation, 30s rest}*5, 60 rest. MB-ME EPI images were acquired with TR=1.6s.

Before re-running 3dD, I've been graphing the output of the TENT functions. Since the 20 second block is not evenly divisible by the TR, I chose to set the duration of the TENT to 19.2 rather than have it bleed into the start of the following rest block.

3dDeconvolve -nodata 80 1.6 -num_stimts 1 -polort -1 -local_times -x1D stdout:  -stim_times 1  '1D: 60' 'TENT(0,19.2,12)'  | 1dplot -thick -one -stdin -xlabel Time -jpg tent_scale_chrysler.jpg
*+ WARNING: no -stim_label given for stim #1 ==> label = 'Stim#1'
++ 3dDeconvolve: AFNI version=AFNI_23.1.09 (Jun 25 2023) [64-bit]
++ Authored by: B. Douglas Ward, et al.
++ using TR=1.6 seconds for -stim_times and -nodata
++ using NT=80 time points for -nodata
++ -stim_times using TR=1.6 s for stimulus timing conversion
++ -stim_times using TR=1.6 s for any -iresp output datasets
++  [you can alter the -iresp TR via the -TR_times option]
++ -stim_times 1 using LOCAL times
++ Number of time points: 80 (no censoring)
 + Number of parameters:  12 [0 baseline ; 12 signal]
++ Wrote matrix values to file stdout:.xmat.1D
++ ----- Signal+Baseline matrix condition [X] (80x12):  1.58375  ++ VERY GOOD ++
++ ----- Signal-only matrix condition [X] (80x12):  1.58375  ++ VERY GOOD ++
++ 1dplot: AFNI version=AFNI_23.1.09 (Jun 25 2023) [64-bit]
++ Authored by: RWC et al.

With 12 knots, I get something like the Chrysler Building in the graph of the suite of regressors, and it does not scale uniformly to 1.

tent_scale_crysler
tent_scale_crysler1024×787 73.7 KB 

3dDeconvolve -nodata 80 1.6 -num_stimts 1 -polort -1 -local_times -x1D stdout:  -stim_times 1  '1D: 60' 'TENT(0,19.2,13)'  | 1dplot -thick -one -stdin -xlabel Time -jpg tent_scale_flat_top.jpg
*+ WARNING: no -stim_label given for stim #1 ==> label = 'Stim#1'
++ 3dDeconvolve: AFNI version=AFNI_23.1.09 (Jun 25 2023) [64-bit]
++ Authored by: B. Douglas Ward, et al.
++ using TR=1.6 seconds for -stim_times and -nodata
++ using NT=80 time points for -nodata
++ -stim_times using TR=1.6 s for stimulus timing conversion
++ -stim_times using TR=1.6 s for any -iresp output datasets
++  [you can alter the -iresp TR via the -TR_times option]
++ -stim_times 1 using LOCAL times
++ Number of time points: 80 (no censoring)
 + Number of parameters:  13 [0 baseline ; 13 signal]
++ Wrote matrix values to file stdout:.xmat.1D
++ ----- Signal+Baseline matrix condition [X] (80x13):  2.7679  ++ VERY GOOD ++
*+ WARNING: !! in Signal+Baseline matrix:
 * Largest singular value=1.41421
 * 1 singular value is less than cutoff=1.41421e-07
 * Implies strong collinearity in the matrix columns! 
++ Signal+Baseline matrix singular values:
             0      0.184592      0.366025      0.541196      0.707107
      0.860919             1       1.12197       1.22474       1.30656
       1.36603       1.40211       1.41421
++ ----- Signal-only matrix condition [X] (80x13):  2.7679  ++ VERY GOOD ++
*+ WARNING: !! in Signal-only matrix:
 * Largest singular value=1.41421
 * 1 singular value is less than cutoff=1.41421e-07
 * Implies strong collinearity in the matrix columns! 
++ Signal-only matrix singular values:
             0      0.184592      0.366025      0.541196      0.707107
      0.860919             1       1.12197       1.22474       1.30656
       1.36603       1.40211       1.41421
++ 1dplot: AFNI version=AFNI_23.1.09 (Jun 25 2023) [64-bit]
++ Authored by: RWC et al.

With 13 knots, i get a flat top in the graph of the suite of regressors, and it also does not scale to 1.

tent_scale_flat_top
tent_scale_flat_top1024×787 71.1 KB

I thought TENT functions scaled to 1. Is that not correct?

Why does changing the number of knots by 1 have such a drastic effect on the plateau, lack of it, in the graphs?

Thanks of your help.

2

Hello,

Indeed, making 19.2 s TR-locked across 12 intervals requires 13 tent knots, e.g. 'TENT(0,19.2,13)'.

But that onset time of 60 s is not on the TR-grid. So the set of locations where you sample the tents is a shifted version of the 0/1 sampling on the TR-grid. In the first example, with only 12 knots, the sampling shifted over time, making the magnitude shift over time. With 13 knots, it is a fixed TR-multiple shift, making the samples look even, but small. To get 0/1, the onset should be on the TR-grid as well, such as using '1D: 16'.

But note that on the plotted TR-grid, while each value might be low, the sum of the values at each (non-zero) time point should still be 1. For the plateaus, note that you are 1/2 TR shifted (60/1.6 = 37.5, an integer plus 0.5 TRs), so the sampling is 0, .5, .5, 0, rather than 0,1,0.

However 12 intervals does not seem like enough in any case. If the stimulus lasts 20 s, you might expect the BOLD response to last more like 32-35. So other choices might be TENT(0,32,21) or TENT(0,33.6,22), for example.

Does that seem reasonable?

  • rick
3

Thanks, Rick, that all makes sense.