Labels are off for DTI data and tracking starts outside the brain


I am battling tracking on HCP dataset with particular methods from dipy (e.g. CSA Peaks, Maximum Direction Getter). I am assuming that something really fishy is happening to the masks because streamlines at the end seem to start… outside the brain. Perhaps due to a small experience, I am doing something wrong on the way. I would really appreciate it if someone can review my approach. :wink:

HCP dataset
DTI: 1.5 mm
T1: 1mm

Coregistration and masks generation
For my specific approach I need to coregister T1 to DTI, which I am doing as follows:

flirt -in ./anat/T1/T1.nii.gz -ref ./diff/raw/mri/diff.nii.gz -out ./tracking_files/t1_coreg_to_diff.nii.gz -omat ./tracking_files/t1_coreg_to_diff.mif -dof 6
5ttgen fsl ./tracking_files/t1_coreg_to_diff.nii.gz ./tracking_files/5TT_coreg_to_diff.mif -nocrop -nthreads 6
mrconvert -coord 3 2 -axes 0,1,2 ./tracking_files/5TT_coreg_to_diff.mif ./tracking_files/wm.nii.gz
5tt2gmwmi -nthreads 10 ./tracking_files/5TT_coreg_to_diff.mif ./tracking_files/wmgm.nii.gz

Mask for GMWMI of a functional region
Now that I have T1 coregistered I want to get a mask of GMWMI of a particular functional region. For that purpose, I go to FSLEeyes and use standard data in MNI152 space, get the functional region masks that are saved as:


These are binarized masks, threshold value of 20.

Now I want to move them from MNI152 space to a subject space. For that I am following this tutorial from mrtrix using prepared bash script:

export templateImage="/usr/local/fsl/data/standard/MNI152_T1_1mm_brain.nii.gz"
export templateImageMask="/usr/local/fsl/data/standard/MNI152_T1_1mm_brain_mask.nii.gz"
export subjectT1=$1
export masksDIR=$2

echo ${masksDIR}BA44_L_thr20.nii.gz

mkdir mask_temp
echo " "
echo "Extracting brain from T1 scan"
echo " "
bet2 ${subjectT1} ./mask_temp/T1_brain -m

echo " "
echo "Calculating transforms and masks"
echo " "
mrcalc ${subjectT1} ./mask_temp/T1_brain_mask.nii.gz -mult ./mask_temp/T1_masked.nii.gz
mrcalc ${templateImage} ${templateImageMask} -mult ./mask_temp/template_image_masked.nii.gz
flirt -ref ./mask_temp/template_image_masked.nii.gz -in ./mask_temp/T1_masked.nii.gz -omat ./mask_temp/T1_to_template.mat -dof 12

echo " "
echo "Calculating dilated transforms, masks and warp"
echo " "

maskfilter ./mask_temp/T1_brain_mask.nii.gz dilate - -npass 3 | mrconvert - ./mask_temp/T1_brain_mask_dilated.nii.gz -strides -1,+2,+3
maskfilter ${templateImageMask} dilate - -npass 3 | mrconvert - ./mask_temp/template_mask_dilated.nii.gz -strides -1,+2,+3
fnirt --config=T1_2_MNI152_2mm.cnf --ref=${templateImage} --in=${subjectT1} --aff=./mask_temp/T1_to_template.mat --refmask=./mask_temp/template_mask_dilated.nii.gz --inmask=./mask_temp/T1_brain_mask.nii.gz --cout=./mask_temp/T1_to_template_warpcoef.nii

echo " "
echo "Getting inverse of a warp"
echo " "
invwarp --ref=${subjectT1} --warp=./mask_temp/T1_to_template_warpcoef.nii --out=./mask_temp/template_to_T1_warpcoef.nii
echo " "
echo "Applying warp to a mask file"
echo " "

applywarp --ref=${subjectT1} --in=${masksDIR}BA44_L_thr20.nii.gz --warp=./mask_temp/template_to_T1_warpcoef.nii --out=./mask_temp/BA44_L_thr20_subj_space.nii.gz --interp=nn
applywarp --ref=${subjectT1} --in=${masksDIR}BA44_R_thr20.nii.gz --warp=./mask_temp/template_to_T1_warpcoef.nii --out=./mask_temp/BA44_R_thr20_subj_space.nii.gz --interp=nn

echo " "
echo "Mixing masks"
echo " "
mkidr masks

mrcalc ./mask_temp/BA44_L_thr20_subj_space.nii.gz ./mask_temp/BA44_R_thr20_subj_space.nii.gz -max ./mask_temp/BA44_RL_thr20_subj_space.nii.gz
mrcalc ./mask_temp/BA44_RL_thr20_subj_space.nii.gz wmgm.nii.gz -and ./masks/BA44_wmgmi_thr20.nii.gz

Now comes the weird part. When I visualize mask on a T1 volume coregistered with T1 everything looks ok in Slicer:

However, when I throw in diffusion data (diff.nii.gz) the mask is really off:

Now the weird thing is when I visualize diffusion data with a mask in FSLEyes it seems to be ok:

although FSLEyes shows the message in location pane:

Displaying images with different orientations/fields of view!
[39 93 47]: 0.0
[100 46 47 0]: 666.577392578125

It seems then the coregistration is off?

When I try to track from this setup, using mask shown here as a source of seeds, it seems that it is really off:

It seems that tracking really started in a place indicated by mask overlayed with diffusion in slicer! How come?

I thought maybe I just need to transform the streamlines, so I did, using DTI affine:

from import load_nifti
from dipy.tracking.streamline import transform_streamlines

data, aff, img = load_nifti(arguments.dti_reference, return_img=True)

streamlines = transform_streamlines(tractogram.streamlines, aff)

but then it was even more off and transform of a inverse affine did not help as well:

streamlines = transform_streamlines(tractogram.streamlines, np.linalg.inv(aff))

Am I doing here something fundamentally wrong?

Best regards,


I might be missing something, but it seems you’ve coregistered the T1 to the diffusion and used that to derive the 5TT images, but your MNI ROIs are derived by coregistration of the template to the original T1 – not the one you coregistered to the diffusion. If that’s what you’re doing, I wouldn’t expect your ROI masks to be aligned with the diffusion.

The reason things display ok in FSLeyes is that it overlays images voxelwise, ignoring the transform (though it does warn you about it!).

Do you think that could be it?


I would daubt that, since as a $subjectT1 I am passing coregistered T1. Same is used for figures.

Hiw can I assure that the alignment is ok as in FSL? Is it the same procedure like moving a mask from MNI152 to subject? Shall I do the same for coregistered T1 andd DTI using e. g. Mean B0 volume?

Oh well, I guess that would have been too easy… I’m not familiar enough with the specifics of flirt and fnirt to advise much more here, but one thing I’m wondering is whether you shouldn’t provide the same image as --ref to the final applywarp as you did when invoking flirt & fnirt?

Otherwise, if you know the image overlay perfectly in FSLeyes, you can try to replace the image transform of your masks with the one from the diffusion. Something like this ought to do the trick:

mrtransform mask.nii -replace diff.nii.gz mask_out.nii

May need some experimentation though, I haven’t tested…