Comparing SPM and FSL

1
Comparing SPM and FSL

• Chris Rorden
• Contrasting SPM to FSL

2
FSL vs SPM5

• We have focused on FSL
• Completely free
• Allows students to get a feel for fMRI analysis
• SPM is the most popular tool
• Free, but requires Matlab to run
• Here we contrast these tools

3
FSL3.3 vs SPM5
SPM5
FSL

• Typical fMRI processing pipeline is similar.
• In FSL, normalization is done after initial
  statistics.
• Allows you to see activation on original raw
  scans.
• Faster, as we usually super sample images during
  normalization (i.e. increase field of view and
  resolution).

• Motion Correction

Motion Correction
Slice Timing Correction
Slice Timing Correction
Normalization
Smoothing
Smoothing
Individual Statistics
Individual Statistics
Normalization
Group Statistics
Group Statistics
4
Motion Correction (realignment in SPM)

• Both SPM and FSL use rigid body registration.
  Different cost functions are used (SPM variance
  FSL Normalised Correlation).
• By default, SPM aligns all images to first 4D
  volume, while FSL aligns to the middle 4D volume.
  
• Optionally, SPM can realign and unwarp which
  attempts to correct head motion related changes
  in image intensity (see spatial processing
  lecture).
• FSLs optional solution is to add motion
  parameters to statistical model (the Stats tab of
  FEAT).
• SPMs solution more sophisticated, but time
  consuming.
• Both techniques can reduce noise, but will reduce
  power if head motion correlates with task (e.g.
  head moves with button presses).

5
Slice Timing Correction

• Slice Timing Correction attempts to make all
  slices in a 3D volume appear as if they were
  collected simultaneously (see temporal processing
  lecture).
• Dilemma required but inaccurate for long TR,
  accurate but not influential with short TR.
• SPM and FSL use the same algorithm. By
  convention, most SPM users employ STC for event
  related designs, while FSL users do not.
• We are fortunate to have a scanner that can
  provide full brain coverage with a short TR.
  Therefore, I would use a rapid (2sec) TR for
  event related designs and not use STC.

6
Smoothing

• During the smoothing stage, both FSL and SPM
  apply a spatial smoothing.
• FSL also applies a 100s highpass temporal filter
  to remove low frequency artefacts.
• SPMs temporal filtering occurs during the
  individual statistics stage, with a default 128s
  highpass.
• For both tools, a low pass filter is optional,
  and can help block designs.

7
Normalization

• Normalization aligns the individuals brain to
  stereotaxic space (warping the orientation and
  size), allowing comparison between people.
• SPM and FSL have very different approaches to
  normalization.
• In general, FSL is very robust (always
  approximately right), but pretty constrained
  (there tends to be a lot of residual error).
• SPM is very aggressive, and can do better than
  FSL in ideal circumstances (i.e. good data).

8
Normalization

• Two competing approaches for normalization (found
  both in SPM and FSL)
• Direct normalization
• Normalize T2 fMRI data directly to stereotaxic
  space
• Indirect normalization
• Coregister T2 fMRI data to T1 scan
• Normalize high resolution T1 to stereotaxic space
• Use parameters from step 2 to normalize fMRI data
  to stereotaxic space
• The second approach is better in theory. However,
  it does require a good structural scan, and has
  more chances to fail catastrophically.

9
Normalization

• FSL uses FLIRT (FMRIB's Linear Image Registration
  Tool) to normalize
• Only linear normalization.
• By default, direct normalization uses a T1
  template image, so a between-modality
  (correlation ratio) cost function is used.
• FSLs analysis of DTI data uses a non-linear
  registration tool (IRTK), but this is not
  typically employed for fMRI data.
• SPMs normalization initially uses linear
  transforms, and then applies non-linear
  transforms.
• By default, direct normalization uses a T2
  template, so a variance cost function can be
  applied.

10
Normalization

• SPM5 introduces a very aggressive indirect
  normalization.
• The T1 scan is bias corrected and segmented to
  gray matter, white matter and CSF probability
  maps (see VBM lecture).
• Warping these tissue maps to standard space can
  provide more accurate normalization (as non-brain
  tissue does not influence parameters).
• FSL does skull strip data for normalization,
  but this is more constrained than SPMs method.

11
Individual Statistics

• SPM and FSL apply general linear model to data
• SPM models HRF using double gamma function
  (blue) by default, FSL uses a single gamma
  function (red).
• Both include temporal derivatives(turn these off
  for block designs)

SPM and FSL have different approaches for
autocorrelation see the temporal processing
lecture.
12
Group statistics

• FSL uses estimates of each individuals contrast
  parameter estimates (copes) and variability
  (varcopes), SPM only uses estimates of contrast.
  (see statistics lecture).
• In theory, FSL might be a bit more sensitive. In
  practice, it is much slower.

Z stats
Group
copes varcopes
copes varcopes
copes varcopes
copes varcopes
Sub 4
Sub 1
Sub 2
Sub 3
13
Learning SPM

• SPM5 comes with an excellent manual
• Chapter 25 walks you through analysis of a block
  design.
• Chapter 26 guides you through the analysis of an
  event-related design.