top | previous | FSFAST Tutorial Top

Preprocessing in FS-FAST

1. Introduction to FSFAST Preprocessing

Once the data have been arranged in the proper directory structure and naming convention, they are ready to be preprocessed. Preprocessing includes

1. Template Creation
2. Brain Mask Creation

3. Registration with FreeSurfer Anatomical

4. Motion Correction
5. Slice Timing Correction (if using)
6. Spatial Normalization
7. Masking
8. Spatial Smoothing

In FS-FAST, it is assumed that each data set will be analyzed in three normalization spaces:

• Left Cortical Hemisphere
• Right Cortical Hemisphere
• Subcortical Structures (Volume-based)

You will need to decide how much to smooth the data and whether you want to do slice-timing correction. In this analysis, we will smooth the data by 5mm Full-Width/Half-Max (FWHM) and correct for slice timing. The slice-timing for this particular data set was 'Ascending', meaning that the first slice was acquired first, the second slice was acquired second, etc. To preprocess the data, run:

cd $FSFTUTDIR
preproc-sess -s sess01 -fsd bold -stc up -surface fsaverage lhrh -mni305 -fwhm 5 -per-run 

This data has already been preprocessed, so it should just verify that it is up-to-date, finishing in a few seconds. This command has several arguments:

• -s sess01 : indicates which session to preprocess
• -surface fsaverage lhrh : indicates that data should be sampled to the left and right hemispheres of the 'fsaverage' subject
• -mni305 : indicates that data should be sampled to the mni305 volume (2mm isotropic voxel size)
• -fwhm 5 : smooth by 5mm FWHM after resampling (volume and surface separately)
• -stc up : slice-timing correction with ascending ('up') slice order
• -fsd bold : indicate the functional subdirectory (FSD)
• -per-run : register each run separately

This command does a lot, and it can take quite a long time to run, especially for many subjects. Look at the contents of one of the run directories:

ls $FSFTUTDIR/sess01/bold/001

You will see many files there, but there are three important ones:

• fmcpr.up.sm5.fsaverage.lh.nii.gz - Left hemisphere of fsaverage
• fmcpr.up.sm5.fsaverage.rh.nii.gz = Right hemisphere of fsaverage
• fmcpr.sm5.mni305.2mm.nii.gz - Volume of fsaverage (MNI305 space) - for subcortical analyses

These are time series data, and their names indicate what has been performed on them:

• fmcpr - motion correction, per-run
• up - slice-timing correction using ascending
• sm5 - smoothed by 5mm FWHM (after resampling)
• fsaverage.lh - sampled on the surface of fsaverage left hemi
• fsaverage.rh - sampled on the surface of fsaverage right hemi
• mni305.2mm - sampled in the fsaverage (MNI305) volume at 2mm isotropic

To learn more see Preprocessing Details below.

2. Quality Assurance

2.1. Motion Correction

The motion plots can be viewed with:

plot-twf-sess -s sess01 -fsd bold -mc 

This gives the vector motion at each time point for each run. Note that it is always positive because this is a magnitude. It is also 0 at the middle time point because the middle time point is used as the reference.

There are no rules for how much motion is too much motion. Generally speaking, sudden motions are the worst as are task-related motion.

2.2. Functional-Anatomical Cross-modal Registration

You can get a summary of registration quality using the following command:

tkregister-sess -s sess01 -fsd bold -per-run -bbr-sum

This prints out a value for each run that indicates the quality of the registration with the anatomical.This will be a number between 0 and 1, with 0 being perfect and 1 being terrible. Actual values depend upon exactly how you have acquired your data. Generally, anything over 0.8 indicates that something is probably wrong.

You can view problematic registrations using the following command:

tkregister-sess -s sess02 -fsd bold -per-run

This will display each of the runs. Quit out of tkregister when you have glanced at the registration

3. Preprocessing Details

To understand what it does, we will go back into one of the run directories and see what it creates. To do this,

cd $FSFTUTDIR/sess01/bold/001
ls

This directory previously held only f.nii.gz, workmem.par, and wmfir.par before preprocessing. Now there are a lot of files, each indicative of a different preprocessing stage. Now type

ls -ltr

This command sorts the files by creation time with the oldest at the top and the newest at the bottom. The preprocessing is progressive, meaning that the output of one stage is the input to the next.

3.1. Template

This stage creates template.nii.gz (and template.log). This is the middle time point from the raw functional data (f.nii.gz). This is the reference used to motion correct and register the functionals for this run to the anatomical. It is also used to create masks of the brain.

• Verify that it has the same dimension and resolution as the raw data using mri_info.

3.2. Masking

The masks for this run are stored in the 'masks' directory. Run 'ls -ltr masks'. You will see a file called 'brain.nii.gz'. This is a binary mask created using the FSL BET program. There is also a file called 'brain.e3.nii.gz' which is the mask eroded by three voxels. These have the same dimensions as the template. View the masks with:

tkmedit -f template.nii.gz -overlay masks/brain.nii.gz -fthresh 0.5
tkmedit -f template.nii.gz -overlay masks/brain.e3.nii.gz -fthresh 0.5

The brain.nii.gz is used to constrain voxel-wise operations. The eroded mask (brain.e3.nii.gz) is used to compute the mean functional value used for intensity normalization and global mean time course. There are other masks there that we will get to later.

3.3. Intensity Normalization and Global Mean Time Course

By default, FSFAST will scale the intensities of all voxels and time points to help assure that they are of the same value across runs, sessions, and subjects. It does this by dividing by the mean across all voxels and time points inside the brain.e3.nii.gz mask, then multiplying by 100. This value is stored in global.meanval.dat. This is a simple text file which you can view. At this point, this value is stored and used later. A waveform is also constructed of the mean at each time point (text file global.waveform.dat). This can be used as a nuisance regressor.

• What are the global means for runs 1, 2, 3, and 4?

3.4. Functional-Anatomical Cross-modal Registration

The next six files (init.register.dof6.dat, register.dof6.dat, register.dof6.dat.mincost, register.dof6.dat.sum, register.dof6.dat.log, register.dof6.dat.param) deal with the registration from the functional to the same-subject FreeSurfer anatomical. There are only two files here that are really important: register.dof6.dat and register.dof6.dat.mincost.

• register.dof6.dat - is a text file that contains the registration matrix.
• register.dof6.mincost - is a text file that contains a measure of
  • the quality of the registration.

The registration is will be revisited below when we talk about Quality Assurance

3.5. Motion Correction (MC)

The motion correction stage produces these files: fmcpr.mat.aff12.1D, fmcpr.nii.gz, mcprextreg, mcdat2extreg.log, fmcpr.nii.gz.mclog, fmcpr.mcdat. There are only three important file here:

• fmcpr.nii.gz -- this is the motion corrected functional data. It has the same size and dimension as the raw functional data.
• fmcpr.mcdat - text file of the amount of motion at each time point. This is important for Quality Assurance (see below).
• mcprextreg - text file of the motion correction parameters assembled into an orthogonalized matrix that can be used as nuisance regressors
• Verify that fmcpr.nii.gz has the same dimension as f.nii.gz using mri_info.

3.6. Slice-Timing Correction (STC)

Slice-timing correction compensates for the fact that each of the 30 slices was acquired separately over the course of 2 sec. It does this by interpolating between time points to align each slice to the time of the middle of the TR. The file created with this is fmcpr.up.nii.gz (and fmcpr.up.nii.gz.log).

• Verify that fmcpr.up.nii.gz has the same dimension as f.nii.gz using mri_info.

3.7. Resampling to Common Spaces and Spatial Smoothing

At this point, the functional data has stayed in the 'native functional space', ie, 64x64x30, 3.4x3.4x5mm3. Now it will be sampled into the 'Common Space'. The Common Space is a geometry where all subjects are in voxel-for-voxel registration. There are three such spaces in FSFAST:

• Left hemisphere of fsaverage (fmcpr.up.sm5.fsaverage.lh.nii.gz)
• Right hemisphere of fsaverage (fmcpr.up.sm5.fsaverage.rh.nii.gz)
• Volume of fsaverage (MNI305 space) - for subcortical analyses (fmcpr.sm5.mni305.2mm.nii.gz)

Each of these is the entire 4D functional data set resampled into the common space. The spatial smoothing is performed after resampling. Surface-based (2D) smoothing is used for the surfaces; 3D for the volumes.

Check the dimensions of the MNI305 space volume:

mri_info --dim fmcpr.up.sm5.mni305.2mm.nii.gz
mri_info --res fmcpr.up.sm5.mni305.2mm.nii.gz

The dimension will be '76 76 93 142' meaning that there are 76 columns, 76 rows, 93 slices but still 142 time points (same as the raw data). The resolution will be '2.0 2.0 2.0 2000' meaning that each voxel is 2mm in size and the TR is still 2 sec. The transformation to this space is based on the 12DOF talairach.xfm created during the FreeSurfer reconstruction.

Check the dimensions of the Left Hemisphere 'volume':

mri_info --dim fmcpr.up.sm5.fsaverage.lh.nii.gz
mri_info --res fmcpr.up.sm5.fsaverage.lh.nii.gz

The dimension is '163842 1 1 142'. This 'volume' has 163842 'columns', 1 'row', and 1 'slice' (still 142 time points). You are probably confused right now. That's ok, it's natural. At this point the notion of a 'volume' has been lost. Each 'voxel' is actually a vertex (of which there are 163842 in the left hemisphere of fsaverage). Storing it in a NIFTI 'volume' is just a convenience.

The 'resolution' is '1.0 1.0 1.0 2000'. The values for the first 3 dimensions are meaningless because there are no columns, rows, or slices on the surface so the distances between them are meaningless. The last value indicates the time between frames and is still accurate (2 sec).

The transformation to this space is based on the surface-based intersubject registration created during the FreeSurfer reconstruction.