Consistent Misalignment of Fieldmap to Target EPI

[dwong263]

What happened?

I am pre-processing an fMRI dataset using the latest version of fmriprep. It seems that for a significant portion of my dataset, during SDC correction, the alignment between the anatomical reference of the fieldmap and the target EPI is quite poor. Here is an example:

I believe this issue is related to #461 and #430. It appears the solution suggested in #461 (dilation of the fmap and epi masks) has been implemented in the latest version of fmriprep to make this coregistration step more robust. However the discussion in #463 raised a concern regarding whether all the different types of fieldmap are as resilient to a dilated mask.

My dataset uses the Siemens' phase-difference method (i.e. fmap data consists of two magnitude images and one phase-difference image), where the tissue contrast in the magnitude images is inverted compared to the tissue contrast in the EPI images.

Additionally, as you can see in the screenshot above, the fmap data and the EPI data are fairly well aligned to start off with. Participants have fmap and EPI data acquired shortly after another, so there are only small rotations/translations between images. As such, I would expect that this would not be a difficult registration problem.

However the output of the registration is a significant misalignment (BOLD data in red, clipped_trans.nii.gz (fmap data) in grey):

What are some potential solutions to make this coregistration step more robust?

What command did you use?

# fmriprep_singlesubject.sh

DATA_DIRECTORY='./hande_fmri'
OUTPUT_DIRECTORY='./hande_fmri/derivatives/fmriprep'
WORKING_DIRECTORY='./fmriprep_scratch'
SUBJECT="64"
LICENSE='./license.txt'
N_THREADS=4

find "$DATA_DIRECTORY" -name "*._*" -delete

fmriprep-docker "$DATA_DIRECTORY" "$OUTPUT_DIRECTORY" \
	participant \
    --skip-bids-validation \
	--participant-label "$SUBJECT" \
    --task-id "taprest" \
    --output-spaces MNI152NLin2009cAsym:res-2 \
    --bold2anat-init "t1w" \
	--md-only-boilerplate \
	--fs-license-file "$LICENSE" \
	--stop-on-first-crash \
    --work-dir "$WORKING_DIRECTORY"

What version of the software are you running?

25.1.4

How are you running this software?

Docker

Is your data BIDS valid?

Yes

Are you reusing any previously computed results?

No

Please copy and paste any relevant log output.

# The following log output shows that dilate_fmap_mask is implemented ...

250807-04:28:43,648 nipype.workflow INFO:
	 [Node] Setting-up "fmriprep_25_1_wf.sub_03_wf.bold_ses_12_task_taprest_run_1_wf.bold_fit_wf.fmapreg_wf.dilate_fmap_mask" in "/scratch/fmriprep_25_1_wf/sub_03_wf/bold_ses_12_task_taprest_run_1_wf/bold_fit_wf/fmapreg_wf/dilate_fmap_mask".


# The following log output shows that fmriprep completed without errors ...

250807-22:47:08,284 nipype.workflow IMPORTANT:
	 fMRIPrep finished successfully!

Additional information / screenshots

The relevant ANTs registration command generated by fmriprep is as follows:

antsRegistration 
     --collapse-output-transforms 1 
     --dimensionality 3 
     --initial-moving-transform [ /scratch/fmriprep_25_1_wf/sub_03_wf/fmap_preproc_wf/wf_fmap_sub03_ses12/magnitude_wf/brainextraction_wf/clipper_post/clipped.nii.gz, /scratch/fmriprep_25_1_wf/sub_03_wf/bold_ses_12_task_taprest_run_1_wf/bold_fit_wf/enhance_and_skullstrip_bold_wf/n4_correct/sub-03_ses-12_task-taprest_run-1_bold_average_corrected.nii.gz, 1 ] 
     --initialize-transforms-per-stage 0 
     --interpolation LanczosWindowedSinc 
     --output transform 
     --transform Rigid[ 0.1 ] 
     --metric Mattes[ /scratch/fmriprep_25_1_wf/sub_03_wf/fmap_preproc_wf/wf_fmap_sub03_ses12/magnitude_wf/brainextraction_wf/clipper_post/clipped.nii.gz, /scratch/fmriprep_25_1_wf/sub_03_wf/bold_ses_12_task_taprest_run_1_wf/bold_fit_wf/enhance_and_skullstrip_bold_wf/n4_correct/sub-03_ses-12_task-taprest_run-1_bold_average_corrected.nii.gz, 1, 32, Random, 0.25 ] 
     --convergence [ 50, 1e-07, 5 ] 
     --smoothing-sigmas 8.0mm 
     --shrink-factors 2 
     --use-histogram-matching 1 
     --masks [ /scratch/fmriprep_25_1_wf/sub_03_wf/bold_ses_12_task_taprest_run_1_wf/bold_fit_wf/fmapreg_wf/dilate_fmap_mask/clipped_mask_dil.nii.gz, /scratch/fmriprep_25_1_wf/sub_03_wf/bold_ses_12_task_taprest_run_1_wf/bold_fit_wf/fmapreg_wf/dilate_target_mask/sub-03_ses-12_task-taprest_run-1_bold_average_corrected_brain_mask_maths_dil.nii.gz ] 
     --transform Rigid[ 0.1 ] 
     --metric Mattes[ /scratch/fmriprep_25_1_wf/sub_03_wf/fmap_preproc_wf/wf_fmap_sub03_ses12/magnitude_wf/brainextraction_wf/clipper_post/clipped.nii.gz, /scratch/fmriprep_25_1_wf/sub_03_wf/bold_ses_12_task_taprest_run_1_wf/bold_fit_wf/enhance_and_skullstrip_bold_wf/n4_correct/sub-03_ses-12_task-taprest_run-1_bold_average_corrected.nii.gz, 1, 32, Random, 0.5 ] 
     --convergence [ 20, 1e-08, 3 ] 
     --smoothing-sigmas 2.0mm 
     --shrink-factors 1 
     --use-histogram-matching 1 
     --masks [ /scratch/fmriprep_25_1_wf/sub_03_wf/bold_ses_12_task_taprest_run_1_wf/bold_fit_wf/fmapreg_wf/dilate_fmap_mask/clipped_mask_dil.nii.gz, /scratch/fmriprep_25_1_wf/sub_03_wf/bold_ses_12_task_taprest_run_1_wf/bold_fit_wf/fmapreg_wf/dilate_target_mask/sub-03_ses-12_task-taprest_run-1_bold_average_corrected_brain_mask_maths_dil.nii.gz ] 
     --winsorize-image-intensities [ 0.001, 0.999 ]  
     --write-composite-transform 0

Looking into the masks used, it seems for the particular subject I am using as an example, the automatic masking seems to fail and includes many bright voxels in the scalp and neck. I wonder if this is maybe driving the problem. Screenshot of clipped_mask_dil.nii.gz:

[dwong263]

I have found a workaround that seems to robustly solve this problem. First, I run fmriprep as is, saving the scratch outputs. With the scratch outputs, I do my own brain masking (using FSL BET) and linear registration (using FSL FLIRT). I then specify the resulting transformation matrix as a derivative and then re-run fmriprep. Obviously this takes longer, since I have to run fmriprep twice, but it works. The combination of FSL BET and FSL FLIRT has not failed on any of my data, while the internal antsRegistration command in fmriprep fails quite often.

Details

Where are the relevant scratch outputs?

• Example Source Image: ./fmriprep2514_scratch/fmriprep_25_1_wf/sub_67_wf/fmap_preproc_wf/wf_fmap_sub67_ses12/magnitude_wf/brainextraction_wf/clipper_post/clipped.nii.gz
• Example Target Image: ./fmriprep2514_scratch/fmriprep_25_1_wf/sub_67_wf/bold_ses_12_task_taprest_run_1_wf/bold_fit_wf/enhance_and_skullstrip_bold_wf/n4_correct/sub-67_ses-12_task-taprest_run-1_bold_average_corrected.nii.gz

Example BET command

bet <input> <output> -f 0.3 -g 0 -m -R

Example FLIRT command

flirt -in <source> -ref <target> -out <out> -omat <fsl_transform> -cost mutualinfo -dof 6 -interp spline

Note that the resulting <fsl_transform> needs to be converted to ITK format for fmriprep. Using c3d_affine_tool:

c3d_affine_tool -ref <target> -src <source> <fsl_transform> -fsl2ras -oitk <itk_transform>

The resulting <itk_transform> file will need to have "MatrixOffsetTransformBase_double_3_3" replaced by "AffineTransform_double_3_3" for fmriprep to interpret the transform file correctly.

Example Result

fmriprep's Native Fieldmap Registration

My Custom Fieldmap Registration