Spatial Normalization: AFNI, SPM, and FSL – Effect on Functional Data
Purpose:
Characterize relative accuracy of SPM5’s high-order, FSL’s fnirt, and AFNI’s @auto_tlrc methods for spatially normalizing images. @auto_tlrc is at a disadvantage because it is a 9-parameter affine warp while both SPM and FSL use many more parameters. The method in SPM5 uses a discrete cosine transform as a basis set, and this is parameterized by roughly 450 3-element vectors. FSL’s fnirt uses a basis set that appears to have a similar spatial resolution but it enforces a diffeomorphic constraint. See the FSL website fnirt and the SPM website
Goal
Evaluate effect of normalization method on group-level t-statistics
Methods
- Data from a study of children from ages 10-14
- N = 48 subjects
- Angry/happy/neutral faces task
- Compute Main effect of task using AFNI
- Correct for multiple comparisons using a Bonferroni correction.
- Compare 4 methods of spatial normalization:
- AFNI: @auto_tlrc
- SPM: SPM5 high parameter warp.
- FSL — flirt
- Use flirt to register EPIs to the structural
- Then use applywarp to apply high-parameter warp computed by fnirt.
- FSL — resample
- Use 3dresample to register EPIs to structural using a rigid body transformation (assumes no motion between T1 and EPI)
- Use fnirt to apply high parameter warp
Results
Comparison of activation maps computed with each method at 8 levels.
Summary
- Differences not tremendous among afni, spm, fsl — resample
- FSL — flirt yields higher t-statistics except for regions that disappear (see insula at z = +10 mm)