Low pattern similarity between TR0 and TR1 when using TENT

Nevi · February 16, 2024, 8:29am

AFNI version info (afni -ver): Version AFNI_23.1.05 'Publius Helvius Pertinax'

Hi all,

I runned the 3dDeconvolve with TENT function and I found significant low pattern similarity between TR0 and TR1 for certain stimulus.

The stimulus did not showed the significant low pattern similarity is the stimuli that has appeared at the first of the run.

I've tried to change the onset (-2s or +2s, which are -1TR and +1TR) but I could get the same results.
e.g) When cue onset is 10s, the pattern similarity between at the 10s and 12s were too low (near -0.7, -0.8)
When I change the cue onset as -2s, the pattern similarity between the 8s and 10s were too low (near -0.7, -0.8)
When I changed the cue onset as +2s, the pattern similarity between the 12s and 14s were too low (near -0.7, -0.8)

Is there any reseaon that this low pattern similarity could appear? and is this the correct neural activity?
(But I thought this result were too weired, because I could see the reversed activity of entire brain at TR0 and TR1 when I observed the result of the deconvolution by afni)

If this is the limitation of the TENT function, can I deconvolve the neural actvitiy before 1TR from the cue onset, and eliminate the first output to get the neural activity at cue onset?

+) There is one stimuli that is presented at the first of the run, which means onset time is 0s. Is it possible to deconvolve from the 1TR before the onset (which means, using the TENT like TENT(-2,16,10) for the given stimuli time 0). I conducted that, and there was no warning (could be due to there are repeats) but I can not sure for the output of this.

+) I will conduct the time correction with different time range, to make sure there is some values before the 0s onset time at the beginning of the run. (e.g. cutting less bookend TR). Is it better way to get proper output from the deconvolution, if I want to deconvolve 1TR before the onset?

Below part is my script for the deconvolution

#!/bin/tcsh

set analName = REAC

set analDir 		= $subjectDir/REAC_add1TR
set resultsDir 		= $analDir/Results_scaled
set stimDir 		= $analDir/Stim_Concat
set contrastDir 	= $analDir/Contrasts

echo 1

set inputFile 		= $resultsDir/${analName}_mc_concat_scaled+orig.HEAD
#set inputFile 		= $resultsDir/${analName}_mc_concat+orig.HEAD

echo $inputFile

cd $resultsDir
rm ${analName}_eachstim_TENTM10_poly4*

3dDeconvolve -xout -input ${inputFile} \
	-xjpeg $resultsDir/matrix \
	-mask $maskDir/brainMask+orig \
	-polort 4 \
	-jobs 32 \
	-num_stimts 18 \
	-float \
	-tout \
	-basis_normall 1 \
	-stim_times 1 ${stimDir}/${participant}-${analName}-1 'TENT(-2,16,10)' -stim_label 1 1 \
	-stim_times 2 ${stimDir}/${participant}-${analName}-2 'TENT(-2,16,10)' -stim_label 2 2 \
	-stim_times 3 ${stimDir}/${participant}-${analName}-3 'TENT(-2,16,10)' -stim_label 3 3 \
	-stim_times 4 ${stimDir}/${participant}-${analName}-4 'TENT(-2,16,10)' -stim_label 4 4 \
	-stim_times 5 ${stimDir}/${participant}-${analName}-5 'TENT(-2,16,10)' -stim_label 5 5 \
	-stim_times 6 ${stimDir}/${participant}-${analName}-6 'TENT(-2,16,10)' -stim_label 6 6 \
	-stim_times 7 ${stimDir}/${participant}-${analName}-7 'TENT(-2,16,10)' -stim_label 7 7 \
	-stim_times 8 ${stimDir}/${participant}-${analName}-8 'TENT(-2,16,10)' -stim_label 8 8 \
	-stim_times 9 ${stimDir}/${participant}-${analName}-9 'TENT(-2,16,10)' -stim_label 9 9 \
	-stim_times 10 ${stimDir}/${participant}-${analName}-10 'TENT(-2,16,10)' -stim_label 10 10 \
	-stim_times 11 ${stimDir}/${participant}-${analName}-11 'TENT(-2,16,10)' -stim_label 11 11 \
	-stim_times 12 ${stimDir}/${participant}-${analName}-12 'TENT(-2,16,10)' -stim_label 12 12 \
	-stim_file 13 ${resultsDir}/REACMotion'[1]' \
	-stim_base 13 \
	-stim_file 14 ${resultsDir}/${analName}Motion'[2]' \
	-stim_base 14 \
	-stim_file 15 ${resultsDir}/${analName}Motion'[3]' \
	-stim_base 15 \
	-stim_file 16 ${resultsDir}/${analName}Motion'[4]' \
	-stim_base 16 \
	-stim_file 17 ${resultsDir}/${analName}Motion'[5]' \
	-stim_base 17 \
	-stim_file 18 ${resultsDir}/${analName}Motion'[6]' \
	-stim_base 18 \
	-concat $contrastDir/runs_${analName}.1D \
	-full_first -fout -tout \
	-bucket $resultsDir/${analName}_eachstim_TENTM10_poly4

cd

Best Regards,
Minjae Kwon

Gang · February 18, 2024, 4:09pm

Minjae Kwon,

I may need some clarification on your question. The hemodynamic response typically exhibits a sluggish pattern, evident in the standard response profile. It is presumed that the response at the stimulus onset is zero, and the subsequent time point may show a small or even negative response, possibly indicating an initial dip, as discussed in this recent paper. Could you provide more details on why you anticipate a similarity between the estimated response at the onset and the subsequent time point?

Gang Chen

Nevi · February 19, 2024, 12:21am

Thank you for your response.

I understand what you said that there could be an initial dip in the hemodynamic reponse. The thing I want to ask is more about the mechanic of TENT function, that can a low pattern similarity between timepoint occur at the onset time could happen when using TENT.

To clarify my question, I have to answer your question. Actually, I did not have any anticipation a similarlity between reponse at the cue onset and subsequent time point. I just have a concern about the output of TENT function due to I found weird thing in my data.

In my expriment, there is a cue onset and the action(recall) starts 3s after cue. The action lasted for 4s, I just wanted to investigate the difference of neural pattern at the cue onset and the action. To figure out the temporal pattern change, I got the neural data from cue onset to end of the recall (16s from cue to 4s after recall) using TENT function. The pattern similarity is one of the analysis for the investigation, and I found that there is a too low pattern similarity between cue onset and 2s after cue onset(first and second timepoint of the deconvolution output) for certain stimuli, for all ROIs, for all participants.

To confirm my data, I did some other things. I found that the stimulus did not show the low pattern simiarltiy between first and second time point, were stimulus which were presented at the first of the each run. When I deconvolve the neural data from the 2s after cue onset timepoint, the pattern similarity between 2s after cue onset and 4s after cue onset showed low pattern similarity for the same stimulus. Moreover, when I deconvolve the neural data from the 2s before cue onset timepoint, the pattern similarity between 2s before cue onset and cue onset showed low pattern similarity for the same stimulus.
(The reason why I changed beginning point of the deconvolution as 2s is because 1TR is 2s)

This is the reason why I asked you about low pattern similarity between close timepoint at the cue onset. I dont't have to test any hypothesis about the pattern similarity between them, but I don't know the reason of this effect, and this makes me concern about the using the output of TENT at the cue onset.

Thank you
Minjae Kwon

Gang · February 20, 2024, 10:41pm

Minjae Kwon,

The neural response at the cue onset and 2 seconds later is expected to be correlated. However, it's important to note that the BOLD signal, although related to neural response, operates on a slower timescale, peaking around 3 to 5 seconds. Hence, it is not surprising that a strong similarity in the estimated hemodynamic response between cue onset and 2 seconds was not observed.

Gang Chen

Nevi · February 21, 2024, 4:24am

Gang Chen,

I understood your response to mean that there could be or could not the correlation between BOLD signal at cue onset and 2 seconds later, and this is not a ciritical problem. Did I understand this correctly?

But one thing I curious about is why that(low pattern similarlity ~-0.8) happens at different time point when I change the starting point of the deconvolution.

For example in my data, there were two case of decovolution, which start from the cue onset and start from the 2s before cue onset.

Case 1. Deconvolution from the cue onset

I deconvolved the BOLD signal using TENT script

-stim_times 1 {stimDir}/{participant}-${analName}-1 'TENT(0,18,10)' -stim_label 1 1 \

the given stim file was the onset of the cue, so the output of this script was 10 BOLD signal from the cue onset.

The first and second coef (which are the cue onset and 2s after cue onset) appeared to be almost inversely related to each other.

Case 2. Deconvolution from the 2s before cue onset
I deconvolved the BOLD signal using TENT script

-stim_times 1 {stimDir}/{participant}-${analName}-1 'TENT(-2,16,10)' -stim_label 1 1 \

the given stim file was the onset of the cue(same as case1), so the output of this script was 10 BOLD signal from the 2s before cue onset.

In the comparison between the second and third coef (which are the output that matches the time points of above two figures: cue onset and 2s after cue onset), the reverse pattern was not observed at this result.

However when I compare the first and second coef, (which are the 2s before cue onset and cue onset) they appear to be almost inversely related to each other.

These results makes me concern to use which pattern (first output in case 1 vs second output in case 2) as base.

I understood that the output of deconvoltuon at the cue onset did not capture the neural activity of the cue onset. Then is there any recommendation for base BOLD signal to compare base vs cue onset / base vs action?

Gang · February 22, 2024, 4:15pm

If you don't anticipate any participant expectation effects, the estimated BOLD signal at the onset time point (given your case 1 specification TENT(0,18,10)) might capture something unrelated to the task. Similarly, the estimated signal for the first two time points with your case 2 specification TENT(-2,16,10) could be just noise. To avoid wasting effort on the data points before or at the onset time, I recommend using TENT(2,18,9). This adjustment allows you to focus on the expected hemodynamic response profile.

Gang Chen

Nevi · February 23, 2024, 6:20am

Gang,

I conducted what you suggest, and I found that there was also negative correlation between first and second output when I used TENT(2,18,9) (The reversed pattern was observed at the first and second output of the TENT). Did this mean that the outputs of TENT near onset were most likely just noise?

This pictures were all outputs of 3 #0_Coef and 3 #1_Coef from each method of TENT, and they looks reversed pattern between 3 #0_Coef and 3#1_Coef for all TENT method.

Addtional Question)
There were some stimuli that were presented at the first of the run, which means cue onset time is 0s. When conducted TENT(-2,16,10), it gave me the 10 output even though there is no data at the -2s before onset for those stimuli, without error messages of the TENT. How TENT function treat this case?

Thank you for your response,
Minjae Kwon

Gang · February 23, 2024, 1:47pm

Minjae Kwon,

Temporal profile of hemodynamic response:
- To gain more informative insights, consider examining the temporal profile of the estimated hemodynamic response.
- Follow these steps:
  - Extract the sub-bricks associated with the estimated hemodynamic response.
  - Load these sub-bricks and visualize them as UnderLay on the AFNI GUI.
  - Use the Graph button to explore the voxel-level temporal profile.
- This should help you explain the spatial correlation structure you observed.
First trial and hemodynamic response:
- The hemodynamic response at the condition level is estimated with many trials together, not separately for each single trial.
- It is not essential for a specific trial's response to be fully available.
- As long as all the combined trials cover the entire response course, the estimation remains valid.
- In other words, the absence of a few time points in the first trial does not pose an issue.

Gang Chen