MNI normalization question with 3dNwarpApply

AFNI version info (22.1.08):

Dear all,

I am using @SSwarper and 3dNwarpApply to normalize EPI data.

The 3dNwarpApply command looks as follows

3dNwarpApply -overwrite -nwarp "$curr_warp $curr_aff12_1D" -source $mask_path/mask_in_epi.nii.gz -master mni+tlrc. -dxyz 1 -prefix mask_norm -verb

where $curr_warp is the warp file from @SSwarper and $curr_aff12_1D is the matrix from @SSwarper.

Is this a valid approach or am I missing something?

Thank you,

Hi, Philipp-

What do the QC images created by @SSwarper look like? Do they look good?

Note that normally the subject T1w anatomical is input into @SSwarper, and that that is aligned to a standard space template---that is something with enough spatial detail to have a good chance of good alignment to template space.

The subject EPI is typically much poorer resolution and contrast, and is not directly aligned to standard space. Instead, one would align the EPI to the anatomical (say, an affine alignment with; align the anatomical to the template (nonlinear alignment with @SSwarper), and then concatenate the warps before applying it to the EPI.

In most FMRI processing, there is the additional alignment step of motion "correction", which is usually rigid body alignment across time. That, too, would be concatenated with the other warps, before applying to the EPI, to minimize blurring from multiple regriddings.

This full barrage of alignments, as well as their proper concatenation, is one of the many convenient things that will generally take care of for you.


Hi Paul,

as you can see from the picture below, I'd say sswarper did a good job.

I agree that epi usually have lower resolution, so I am surprised that the normalization looks fine as far as I can tell.

Using the tranformation matrix from and the warp image from sswarper

3dNwarpApply -overwrite -nwarp "anat2mni_WARP.nii.gz epi_al_mat.aff12.1D" -source epi.nii.gz -master mni+tlrc. -dxyz 1 -prefix epi_norm -verb

yielded bad results:

even though the alignment of EPI to ANAT worked well:

Hence, I don't know what the problem exactly is.


Hi, Philipp-

It's not clear to me what the inputs of these various things are. Did you use:

  • @SSwarper to align the anatomical to template
  • to align the EPI to the anatomical

(which is basically the standard way of going)? From the two good alignments you show, that appears to be the case.

If so, you need to concatenate all alignment/warp pieces: EPI->anat affine, anat->template affine, anat->template nonlinear warp. From your codes here, it looks like these various alignments that are not good have not done so. Also, note that often with, we align the anat->EPI, so that matrix has to be inverted. But that question depends in detail on the transforms you made.

So, if you could copy+paste your commands that you ran here (both to generate the alignments and to try to apply them), that would help clarify things.

But again, this kind of task would be done most easily using, where you can even just specify alignment blocks and no regression even, if you don't want (but including that would get you the QC HTML).


@SSwarper produces both an affine transformation and a nonlinear warp - two different files. With EPI to anatomical alignment transformations and motion correction transformations too, all the transformations are chained together into a single transformation. usually takes care of this complicated procedure - getting the order and direction of each right.

Also I don't think MNI+tlrc is a template we distribute, so it's not clear whether that dataset was used for the alignment in @SSwarper. Post your actual command to do the alignment, maybe sticking to the MNI152_2009_template_SSW.nii.gz as your template base for SSW and 3dNwarpApply. Your MNI+tlrc might be just fine; I'm not familiar with that particular dataset, and there are lots of MNI templates available that can yield different results.

Hi all,

thanks for your responses so far!

I will clarify things to make it easier for you to see where I am coming from:

I did some resting-state functional connectivity analyses in native EPI space space and now I want to do a group comparison, hence I want to normalize the corrZ maps that I got. That implies that the data have already been preprocessed and analyzed on the 1st level.

What I did so far was using SSwarper on the anatomical image to normalize it to MNI

@SSwarper -input anat+orig. -base MNI152_2009_template_SSW.nii.gz -subid sswarp_anat2mni -SSopt '-blur_fwhm 2' -verb

, which gave me

  • anatQQ.sswarp_anat2mni_WARP.nii
  • anatQQ.sswarp_anat2mni.aff12.1D

Then I ran -overwrite -anat anatSS+orig. -epi epi+orig. -epi_base 0 -master_epi anatSS+orig. -epi2anat -dset1_strip None -dset2_strip None -partial_axial -edge -cost lpa -big_move

which gave me

  • epi_al_mat.aff12.1D, which should be the EPI-> affine matrix as a specified so in the command. Therefore, an inversion of this matrix is not necessary, I think?

If I understood correctly, I have to somehow concatenate the two matrices epi_al_mat.aff12.1D and anatQQ.sswarp_anat2mni.aff12.1D into one file and use it together with anatQQ.sswarp_anat2mni_WARP.nii in 3dNwarpApply?

3dNwarpApply -overwrite -nwarp "anatQQ.sswarp_anat2mni_WARP.nii concat_matrices.1D" -source epi+orig. -master MNI152_2009_template

My question would be how to concatenate the 2 matrices properly and input it correctly into the 3dNwarpApply command?

Finally, I understand that afni_proc is for proper fMRI preprocessing, which I don't have to use, therefore I was wondering whether it makes sense to use it?


You're on the right track. Here is an excerpt from the proc.FT.NL script distributed with the class data with the missing details.

    # catenate volreg/epi2anat/tlrc xforms
    cat_matvec -ONELINE                                                  \
               anatQQ.FT.aff12.1D                                        \
               anatSS.FT_al_junk_mat.aff12.1D -I                         \
               mat.r$run.vr.aff12.1D > mat.r$run.warp.aff12.1D

    # apply catenated xform: volreg/epi2anat/tlrc/NLtlrc
    # then apply non-linear standard-space warp
    3dNwarpApply -master anatQQ.FT+tlrc -dxyz 2.5                        \
                 -source pb01.$subj.r$run.tshift+orig                    \
                 -nwarp "anatQQ.FT_WARP.nii mat.r$run.warp.aff12.1D"     \
                 -prefix rm.epi.nomask.r$run

This example includes both the motion correction (volume registration) transformations. In your case, you may decide to leave that out. Still, you would have to be careful to use the same registration base index as the alignment base (anat to EPI). The affine transformations that are together in the chain can be concatenated with cat_matvec. Note that in this example code with the anat to EPI (not EPI to anat) alignment affine transformation, you would invert that for alignment of the EPI to anat instead (that's what the -I ("eye") is for). If using the EPI to anatomical affine, then you would NOT invert that particular affine transformation.

Be careful with the naming of the base template. Your last snippet did not include the full filename.

Hi Daniel,

this indeed was the missing part.

I used

cat_matvec -ONELINE anatQQ.sswarp_anat2mni.aff12.1D epi_al_mat.aff12.1D > epi_anat_2_mni.aff12.1D

and plugged it into 3dNwarpApply

3dNwarpApply -overwrite -nwarp "anatQQ.sswarp_anat2mni_WARP.nii epi_anat_2_mni.aff12.1D" -source epi.nii.gz -master MNI152_2009_template -dxyz 1 -prefix epi_norm -verb

which gave me this nice result:

Is this how it should be done?

I am still suprised that it looks so similar to the one without epi alignment information:

Thanks for your help and bests,

If you are viewing the transformed volreg base, there should be no difference between them, as the volreg base is not affected by the volreg transformation.

Out of curiosity, why are you not using for this? Is it for some part of registration that does not do? Perhaps you could just use for the 'align tlrc volreg' blocks, at the least, since that seems to be what you are putting together by hand.

  • rick

I think you are probably still missing that option to invert the affine transformation with -I that Paul and I have mentioned.

If your procedure fits the pipeline possibilities, then that is indeed an easier choice.

Hi all,

@rickr: I assume that afni_proc works with 4D images (time series only) and since I only have on 3D epi image it doesn't work and throws the following error:

3dcopy /data/LC2016/anat+orig 2016.results/anat
++ 3dcopy: AFNI version=AFNI_22.1.08 (May 3 2022) [64-bit]
3dTcat -prefix 2016.results/pb00.2016.r01.tcat /data/2016/epi+orig[0..$]
++ 3dTcat: AFNI version=AFNI_22.1.08 (May 3 2022) [64-bit]
++ elapsed time = 0.4 s
set tr_counts = ( 1 )
cd 2016.results
3dbucket -prefix vr_base pb00.2016.r01.tcat+orig[2]
++ 3dbucket: AFNI version=AFNI_22.1.08 (May 3 2022) [64-bit]
** ERROR: selector index 2 is out of range 0..0
can't decipher index codes from pb00.2016.r01.tcat+orig[2]

I used the following commands to create the script: -subj_id 2016 -copy_anat /data/LC2016/anat+orig -dsets /data/2016/epi+orig. -blocks align tlrc volreg -align_opts_aea -cost lpc+ZZ -giant_move -tlrc_base MNI152_T1_2009c+tlrc -volreg_align_e2a -volreg_tlrc_warp

@dglen: I used align_epi_anat to align the epi directly to anat, hence using -I would wrongly invert the matrix, I think?


Ah, yes, you are probably right here. The output from the is inverted for the epi2anat output, so you wouldn't need to invert it again. With just a single volume in the EPI, you wouldn't be carrying the motion correction with that.

Hi, Philipp-

The problem is that the default volume for EPI alignment is the [2]th one, so you need to specify something different. Here, add the opt -volreg_align_to first to your command.

I think you will also need to add this, to turn off a variance check:
-find_var_line_blocks NONE. If you aren't seeing that message already, I wonder if your AFNI version is a bit old?

I would also add this, because radial correlation calcs will whine if there is only 1 volume:
-radial_correlate_blocks NONE

Finally, I had added this in my test case, but since you aren't really checking for outliers, it might not matter:
-outlier_polort 0

You should also specify whether your anatomical has its skull on or not, for alignment:
-anat_has_skull no
-anat_has_skull yes.

That would all lead to this command for your data, with YES_OR_NO needing to be replaced with "yes" or "no":                                       \
    -subj_id           2016                        \
    -copy_anat         /data/LC2016/anat+orig      \
    -anat_has_skull   YES_OR_NO                    \
    -dsets             /data/2016/epi+orig.        \
    -blocks            align tlrc volreg           \
    -radial_correlate_blocks NONE                  \
    -align_opts_aea    -cost lpc+ZZ                \
                       -giant_move                 \
    -tlrc_base         MNI152_T1_2009c+tlrc        \
    -find_var_line_blocks NONE                     \
    -volreg_align_e2a                              \
    -volreg_tlrc_warp                              \
    -outlier_polort    0

Also, you could use @SSwarper for nonlinear anatomical-template alignment (warping), as well as skullstripping (SS).


Also, just to be sure, do you also have a series of EPI volumes that you will be including in this processing/alignment? Because if so, it is worth noting that motion correction is also an alignment process, and therefore should be concatenated in the giant set of alignment transforms before application. And again, this is something that would handle well, automatically and by design.


Hi all,

thanks again for your responses and help regarding the normalization!

@ptaylor: there will be no time series to normalize because the 1st level analysis (resting-state) has been done in native space and only the 3D Zcorrelation maps (from a seed-based FC analysis) will be normalized for group comparison, hence movement should not be an issue. However, that begs the question of how to normalize a Zcorr map with afni_proc given that it looks necessarily very different from a standard EPI?

In order to include @SSwarper, I'd modify the nice command you created by the following:                                       \
    -subj_id           2016                        \
    -copy_anat         /data/LC2016/anat+orig      \
    -anat_has_skull   YES_OR_NO                    \
    -dsets             /data/2016/epi+orig.        \
    -blocks            align tlrc volreg           \
    -radial_correlate_blocks NONE                  \
    -align_opts_aea    -cost lpc+ZZ                \
    -giant_move                 \
    -tlrc_base        MNI152_2009_template_SSW.nii.gz        \
    -tlrc_NL_warp        \
    -tlrc_NL_warped_dsets     Qwarp/anat_warped/anatQQ.2016.nii        \
                                      Qwarp/anat_warped/anatQQ.2016.aff12.1D   \
                                      Qwarp/anat_warped/anatQQ.2016_WARP.nii   \
    -find_var_line_blocks NONE                     \
    -volreg_align_e2a                              \
    -volreg_tlrc_warp                              \
    -outlier_polort    0


Hi, Philipp-

If your 1st level analysis is already done, I guess that means you have already done motion correction and the realignment that includes? Note that motion correction itself is an alignment process, one which regrids the EPI volumes. Doing so inherently blurs the data, and then later aligning the data to standard space will blur it again, which is something that should be avoided, if possible. And it is possible, because would combine all of the following alignments---motion correction, EPI-anatomical and anatomical-template---into a single transform.

I'm curious: how did you perform the first level analysis? Is there a reason to not let do that, so all the processing features like alignment concatenation---as well as others, and particularly the useful QC checks automatically included in be done for you?


Hi Paul,

alright, I am happy to go into more detail to give a clearer picture of the situation: overall, I analyse resting-state data in terms of seed-to-whole-brain functional connectivity.

To this end, the data (which are multi echo data) have been preprocessed with multi-echo independent component analysis (ME-ICA; ) with the aim of separating BOLD from non-BOLD signal.
This toolbox actually uses several afni functions, such as 3dretroicor for respiratory and pulse data, 3dTShift+3dvolreg for estimating motion parameters, 3dSkullStip+3dAllineate for anatomical coregistration, where motion correction and anatomical coreg were done in one step using 3dAllineate; 3dBlurInMask is then used for 5mm FWHM smoothing and a high pass filter of 0.02Hz is finally applied before the operations regarding echo combination begin, which include denoising and the ME-ICA component sorting method.

The output file, which is 4D nifti of the preprocessed 4D time series is then subjected to the FC analysis including time series extraction from the seed region with 3dmaskave

3dmaskave -quiet -mask $mask $meica_preproc_BOLD_epi.nii.gz > seed_timeseries.txt

followed by the seed-to-whole-brain correlation

3dTcorr1D -pearson -prefix meica_preproc_BOLD_epi.Tcorr1D.nii.gz $meica_preproc_BOLD_epi.nii.gz seed_timeseries.txt

Finally, the image that I want to normalize is then computed from the correlation coefficient map

3dcalc -a  $meica_preproc_BOLD_epi.Tcorr1D.nii.gz -expr 'log ( ( 1+a ) / ( 1-a ) ) /2' -prefix meica_preproc_BOLD_epi_corrZ

I hope I that gives you a better idea of my situation and why have done things the way I did so far.

Please let me know your thoughts!


Hi, Philipp-

I will sound like a broken record, but multiecho FMRI can be processed as part of Here is an OHBM poster describing the demo we have about doing this:
... where the data and scripts can be downloaded with:


This includes an example of a full script (with @SSwarper run beforehand) for using Kundu et al's MEICA program (sidenote: that was developed down the hallway from the AFNI developers, and has been a part of the AFNI toolbox since about the time it was developed).

You can also implement the optimal combination (OC) echo combination mechanism (Posse et al., 1999), again within's easy to swap those out, and see which you prefer. There are further examples in the AP help.

There is also the much more modern tedana group version of MEICA (Du Pre et al., 2021), which is includable within

I think this would greatly simplify and streamline your processing. You can choose what primary parameters you want (blur radius, final voxel size, censoring thresholds, etc.), while letting doing the more burdensome concatenation/etc. tasks under the hood.


Hi Paul,

I am pleasantly surprised reading this! Sometimes the world is indeed a small place :)

I will dig into the demo, but I wonder how @Install_APMULTI_Demo1_rest relates to, for example,Example 13. Complicated ME, surface-based resting state example from the AP help, which seems to be the closest to my case?

As for now, it seems that in the demo, there are scripts for different scenarios in the scripts_desktop folder. How should I best move forward?

Thanks and bests,

Hi, Philipp-

Did you want to do a surface-based analysis? That example you picked is for that, but from previous discussions, it didn't sound like that was what you were headed towards.

What properties do you want for your analysis? For example:

  • volumetric analysis, or surface-based?
  • if volumetric, going to standard space, or leaving in subject space?
  • what form of combining echoes: Kundu et al. MEICA, DuPre et al. MEICA, or Posse et al. OC?
  • voxelwise analysis, or ROI-based (determines if you want to blur the data or not, respectively)?

Answering those will determine which demo script would be the best starting point. And, because it's, you could easily switch between those, too, relatively conveniently.