Why the Van Essen Lab and Avi Snyder (wustl.edu) Recommend Affine Registration to a Stereotaxic Space before Segmentation in Caret/SureFit

This is probably a little more important for anatomical studies using sulcal depth, but most of these reasons generalize to all studies.  It generally boils down to minimizing noise.  While AC-PC alignment has some advantages over starting with native input, the residual noise due to scale differences is undesirable and unnecessary.

Here are the main reasons we recommend spatial normalization using 12-parameter affine transform -- not just AC-PC alignment:

* You'll probably want to generate an average fiducial surface for your subjects.  You wouldn't average your orig/native volumes, so don't average your orig/native surfaces.

* You'll probably want to average your tlrc/atlas-registered anatomical volumes, to facilitate volume-surface interaction with your average fiducial surface.

* You'll probably want to toggle quickly between subjects' anatomical volumes/surfaces, while keeping the voxel index / node location constant, comparing specific regions across subjects. You'll need tlrc/atlas-registered volumes and surfaces for this purpose.

* If the segmentations are not tlrc/atlas-registered, then there will be added intersubject variability/noise in sulcal depth due to differences in brain size.

Most of these reasons become important at the stage where you are analyzing your surface-registered data (e.g., sulcal depth or fMRI).  It is possible to apply the affine transform to the surfaces post-registration, but for practical reasons, this is a nightmare for wustl.edu users.  It's less painful for AFNI users, but you need to know what you're doing.  Ask yourself why you didn't just normalize before segmenting?  Make sure a linear transformation will really confound your results before going down this path.  For most of the questions we study, we find this linear transformation improves -- not confounds -- our results.

The following benefits of normalizing before segmenting are really benefits of AC-PC alignment (i.e., they're less affected by scale differences across subjects):

* You'll prevent problems with Caret's SureFit segmentation routines (e.g., disconnecting eye, skull, hindbrain, and contralateral hemisphere; filling ventricles).

* You'll ensure that the template frontal cut intersects the medial wall posterior to olfactory sulcus / medial orbital gyrus, thereby standardizing where the medial wall splits into dorsal and ventral segments.  (The calcarine-medial wall junction determines the other split point for the posterior end.)

* You won't need to rotate your surface to approximate AC-PC alignment when drawing registration borders, to ensure your reference distances and aSTG starting point are correct.

How much does it really matter?  For a UCDavis autism study, we initially segmented AC-PC aligned input, rather than affine-registered input.  We later registered the volumes to avg152T1 and recalculated the sulcal depth measures, to remove noise due to scale.  The differences in the t-maps were disappointing, but noticeable.  The normalized maps have wider range (e.g, for LFA9vCON19 t-maps before norm -7.378 to 5.149; after norm: -9.2 to 5.7) and generally stronger signal:

jpg: tmaps before and after

The differences in the MDS plots, however, were more substantial:

jpg: MDS plots before and after

We trust the post-normalization results more than the pre-normalization results, so we used those results for our analyses.

We Don't Use Nonlinear Methods to Volume Register Our Input

For most of our studies, local deformations in our normalized input confound rather than improve our findings.  We let surface-based registration do the lion's share of intersubject alignment, but we value the noise reduction that an affine volume registration provides.
