Beauchamp:ROIanalysis
If you would like to compare activity across subjects in threshold-defined areas, an ROI analysis may be right for you!
For example, we may be interested in comparing amplitude of activity in auditory cortex during different types of speech stimuli in people with different behavioral characteristics. While we could use a group whole-brain analysis to see if there are any voxels in a combined map within auditory cortex that show a differential response, it may be better to separately define each auditory ROI in each subject using strict and reproducible criteria, such as T stat > 3.
To facilitate this process, the surfaces in SUMA come already parcellated into useful regions. With the parcellation, each anatomical region is assigned a value (same in left and right hemisphere). So, for each subject, we can find active areas within these anatomical locations to use as ROIs. Depending on when the surface was made, the parcellation is done either with a 2005s or 2009s atlas.
Here are the parcellation values corresponding to 5 regions of interest using the 2005s atlas:
Auditory cortex 33, 81 Extrastriate visual cortex (V5/MT) 60 Fusiform gyrus 17, 63 Inferior frontal gyrus 6, 72 Superior temporal sulcus 80
And for the 2009s atlas:
Auditory cortex 466, 514 Extrastriate visual cortex (V5/MT) 493 Fusiform gyrus 450 Inferior frontal gyrus 285, 505 Superior temporal sulcus 513
To use these parcellations, first copy over the parcellation files from the SUMA folder:
cd /Volumes/data9/surfaces/last_name/subjID/SUMA/
cp lh.aparc.a2005s.annot.1D.roi /Volumes/data1/UT/{$ec}/afni
cp rh.aparc.a2005s.annot.1D.roi /Volumes/data1/UT/{$ec}/afni
cd /Volumes/data1/UT/{$ec}/afni
Then, convert them into BRIK files for AfNI analysis:
3dSurf2Vol -spec /Volumes/data9/surfaces/netek_anne/HO/SUMA/both.spec -surf_A lh.smoothwm.asc \
-grid_parent {$ec}v1mr+orig -sv {$ec}_SurfVol_Alnd_Exp+orig -map_func max -prefix L_parc -sdata_1D lh.aparc.a2005s.annot.1D.roi
3dSurf2Vol -spec /Volumes/data9/surfaces/netek_anne/HO/SUMA/both.spec -surf_A rh.smoothwm.asc \
-grid_parent {$ec}v1mr+orig -sv {$ec}_SurfVol_Alnd_Exp+orig -map_func max -prefix R_parc -sdata_1D rh.aparc.a2005s.annot.1D.roi
The Freesurfer parcellation gives us the entire length of the STS, but what if you are interested only in the posterior portion of this sulcus? It seems fair to consider only the posterior half of the STS for further analysis when studying STSms. To do this, first create an STS ROI for the entire length. In this example, the left STS in a 2009s parcellation of subject HT:
3dcalc -prefix {$ec}_L_STS_anat_parc -a0 L_parc+orig -expr "step(equals(a,513))"
Now, let's see how far anterior and posterior this ROI actually goes:
3dclust 5 100 {$ec}_L_STS_anat_parc+orig
with output of
#Volume CM LR CM PA CM IS minLR maxLR minPA maxPA minIS maxIS Mean SEM Max Int MI LR MI PA MI IS #------ ----- ----- ----- ----- ----- ----- ----- ----- ----- ------- ------- ------- ----- ----- ----- 16516 -52.8 -22.7 -6.9 -66.0 -38.5 -74.2 33.1 -25.9 13.1 1 0 1 -44.0 27.6 -25.9
This is the entire STS. If we examine the conjunction of auditory and visual activity (t > 2) in the STS, it extends from -31 mm to + 6 mm. There are a few approaches to handling this blob:
- Circle the entire blob manually using the draw tool in SUMA
- The blob is essentially in the mid STS, but FreeSurfer does not parcellate the STS into different sub-regions. One way to do this would be to divide it up ourselves based on the A-P extent, in this case from -74 to 33.
Divide up the STS into halves or thirds, and use that as the cutoff.
Here is an example of how to do this with 3dcalc. We can determine the halfway "cutoff" for y-values in the posterior segment:
cutoff = ((maxPA-minPA) * 0.5) + minPA
In this example, cutoff = (abs(33.1) + abs(-74.2)) * 0.5 - 74.2 = -20.55.
Set this cutoff in the script:
set cutoff = -20.55
Now, we can find voxels that are within the STS anatomical parcellation (in a 2009 parcellation in this case), active during auditory and visual blocks *and* are posterior to y = -20.55:
3dcalc -prefix {$ec}_L_pSTS_auto_parc -a0 L_parc+orig -b17 {$ec}v1mr+orig -c20 {$ec}v1mr+orig -expr "step(b-2)*step(c-2)*step(equals(a,513))*step(y+$cutoff)"
(I realize that the "step(y+$cutoff)" command is kind of backwards. If we want all the values more negative than -20.55, I thought it would've been "isnegative(y-$cutoff)". But, this way works.)
Another example: here we find voxels within the left auditory cortex that are active (T > 2) during auditory blocks in a subject with an old-school 2005s parcellation:
3dcalc -prefix {$ec}_L_aud_parc -a0 L_parc+orig -b17 {$ec}v1mr+orig -expr "step(b-2)*step(equals(a,33)+equals(a,81))"
Perhaps you would like to know the strength of activity averaged across these voxels. It's nice to use percent signal change values, since these are more easily comparable across subjects. First, use tent functions withing 3dDeconvolve to create a nice hemodynamic response for each voxel for each stimulus type:
set v = 2
3dDeconvolve -fout -tout -full_first -polort a -concat runs.txt \
-input {$ec}Albl+orig -num_stimts 12 -nfirst 0 -jobs 2 \
-mask {$ec}maskAlbl+orig \
-stim_times 1 McG.txt 'TENT(0,16,9)' -stim_label 1 McGurk \
-stim_times 2 NonMcG.txt 'TENT(0,16,9)' -stim_label 2 InC \
-stim_times 3 Cong.txt 'TENT(0,16,9)' -stim_label 3 Cong \
-stim_times 4 Target.txt 'TENT(0,16,9)' -stim_label 4 Target \
-stim_times 5 Ablock.txt 'TENT(0,26,14)' -stim_label 5 Ablock \
-stim_times 6 Vemotion.txt 'TENT(0,26,14)' -stim_label 6 Vemotion \
-stim_file 7 {$ec}rall_vr_motion.1D'[0]' -stim_base 7 \
-stim_file 8 {$ec}rall_vr_motion.1D'[1]' -stim_base 8 \
-stim_file 9 {$ec}rall_vr_motion.1D'[2]' -stim_base 9 \
-stim_file 10 {$ec}rall_vr_motion.1D'[3]' -stim_base 10 \
-stim_file 11 {$ec}rall_vr_motion.1D'[4]' -stim_base 11 \
-stim_file 12 {$ec}rall_vr_motion.1D'[5]' -stim_base 12 \
-iresp 1 irf1_tent -iresp 2 irf2_tent -iresp 3 irf3_tent -iresp 4 irf4_tent -iresp 5 irf5_tent -iresp 6 irf6_tent \
-prefix {$ec}v{$v}mr
3dTcat -prefix {$ec}v{$v}irf irf1_tent+orig irf2_tent+orig irf3_tent+orig irf4_tent+orig irf5_tent+orig irf6_tent+orig
Then, convert these impulse response functions into percent signal change:
3dcalc -a {$ec}v2irf+orig -b {$ec}EPIanatAlbl+orig -expr '100 * a/b * ispositive(10-a/b)' -prefix {$ec}psc_AV
Perhaps we are interested in the amplitude of response during each of four event-related conditions listed (McGurk, incongruent, congruent and target). The hemodynamic response peaks between 4-6 seconds, so it works to average the 3rd and 4th points together to estimate the height of the response in percent signal change:
# Psc files-- takes average of points 2 and 3 of the HRF (the numbering is 0-8, so points 2 and 3 are the 3rd and 4th points of the series)
3dMean -prefix {$ec}McGpsc {$ec}psc_AV+orig'[2]' {$ec}psc_AV+orig'[3]'
3dMean -prefix {$ec}InCpsc {$ec}psc_AV+orig'[11]' {$ec}psc_AV+orig'[12]'
3dMean -prefix {$ec}Congpsc {$ec}psc_AV+orig'[20]' {$ec}psc_AV+orig'[21]'
3dMean -prefix {$ec}Targetpsc {$ec}psc_AV+orig'[29]' {$ec}psc_AV+orig'[30]'
Now, you can find the average percent signal change in the anatomical-functional ROI during each of the four conditions:
3dROIstats -quiet -mask {$ec}_L_aud_parc+orig {$ec}McGpsc+orig
3dROIstats -quiet -mask {$ec}_L_aud_parc+orig {$ec}InCpsc+orig
3dROIstats -quiet -mask {$ec}_L_aud_parc+orig {$ec}Congpsc+orig
3dROIstats -quiet -mask {$ec}_L_aud_parc+orig {$ec}Targetpsc+orig
What hemodynamic response function (HRF) to use
The above examples use the 3dDeconvolve TENT functions to model the HRF.
A recent addition is the TENTzero function, which forces the HRF to finish and end at zero value. This is sensible.
Another option is CSPLINE and CSPLINEzero. These are useful if you don't have enough degrees of freedom to do a full TENT function fit to the HRF.
Here is a picture of the HRF from a subject's STS calculate with TENTzero(0,16,9) and CSPLINEzero(0,16,4); 4 is the minimum n for CSPLINE; if 9 is used, it gives the same results as TENT.
Obtaining trial-specific estimates of response amplitudes
Usually with ROI analyses we are only interested in obtaining one average HRF for a given ROI. This is an average of each time a given stimulus was presented in the scanner. This might be done with the following command lines:
3dDeconvolve -fout -tout -full_first -polort a -concat runs.txt -x1D {$ec}method1.xmat.1D\
-input {$ec}Albl+orig -num_stimts 16 -nfirst 0 -jobs 2 \
-mask {$ec}maskAlbl+orig \
-stim_times 1 Arel.txt 'BLOCK(2,1)' -stim_label 1 Arel \
-stim_times 2 Vrel.txt 'BLOCK(2,1)' -stim_label 2 Vrel \
-stim_times 3 TargetRel.txt 'BLOCK(2,1)' -stim_label 3 TargetRel \
-stim_times 4 McG.txt 'BLOCK(2,1)' -stim_label 4 McGurk \
-stim_times 5 NonMcG.txt 'BLOCK(2,1)' -stim_label 5 InC \
-stim_times 6 Cong.txt 'BLOCK(2,1)' -stim_label 6 Cong \
-stim_times 7 Target.txt 'BLOCK(2,1)' -stim_label 7 Target \
-stim_times 8 Ablock.txt 'BLOCK(20,1)' -stim_label 8 Ablock \
-stim_times 9 Vblock.txt 'BLOCK(20,1)' -stim_label 9 Vblock \
-stim_times 10 AVblock.txt 'BLOCK(20,1)' -stim_label 10 AVblock \
-stim_file 11 {$ec}rall_vr_motion.1D'[0]' -stim_base 11 \
-stim_file 12 {$ec}rall_vr_motion.1D'[1]' -stim_base 12 \
-stim_file 13 {$ec}rall_vr_motion.1D'[2]' -stim_base 13 \
-stim_file 14 {$ec}rall_vr_motion.1D'[3]' -stim_base 14 \
-stim_file 15 {$ec}rall_vr_motion.1D'[4]' -stim_base 15 \
-stim_file 16 {$ec}rall_vr_motion.1D'[5]' -stim_base 16 \
-prefix {$ec}method1
Sometimes we may want to have an estimate for each individual trial. This can be useful for a connectivity analysis, MVPA, or getting more N for an individual subject.
Previously we have used 3dDeconvolve with -stim_times_IM (individual modulation) to create a bucket file with an estimate for each stimulus presentation for each voxel in the brain. You can then use 3dROIstats to average across a set of voxels. This method works, but can often have very high variance.
Another way to do this is to use 3dDeconvolve using -stim_times_IM for only one stimulus, while holding all other stimuli at -stim_times. This makes sure that the regression analysis accounts for the presentation of other stimuli (otherwise they are considered "baseline", which can skew your results) while only obtaining estimates for a single stimulus of interest.
The best way however is to take the results from the previous method and use 3dLSS (least squares-sum) to estimate trial-specific responses. This method will give you lower variance, which may help you detect changes between different stimulus conditions better.
What if the FreeSurfer parcellation doesn't work
FreeSurfer generates a .stats file which contains the GM volume and surface area for each area in the brain. Sometimes the FreeSurfer parcellation may fail, for instance if there is a lesion in the brain. In this case, you can manually edit the parcellation file and then calculate the volume for each area (see subject IE for details). The volumes calculate by this method are not as accurate because they don't take partial voluming into account. Another possibility is to edit the aseg FS parcellation files, e.g. aseg.mgz and then run the mri_segstats program. This will take partial voluming into account. One could also fill in the lesion in the wm.mgz file and then the surface should go outside the lesion.
