Dear SPMers,
I am interested in your thoughts on optimisation of between vs. within subject alignment. We have collected structural (MPM) and functional data from three imaging sessions, conducted within 7 days, the experimental intervention being training in a behavioural task (scans: baseline, pre-training, post-training).
The main comparisons we would like to make are within-subject, for example Scan 3 - Scan 2 vs. Scan 2 - Scan 1. Having tried different alignment methods, longitudinal alignment followed by standard normalisation maximises within-subject registration, while when using DARTEL the within subjects is slightly worse, but much better for between-subject (as you'd expect).
For a simple paired-ttest between two differences images, d1 = Scan 3 - Scan 2, d1 = Scan 2 - Scan 1.
f = (mean(d1) - mean(d2)) / SE(d1 - d1).
Here it is clearly important to maximise within-subject registration for the subtractions, but at the end the differences are averaged over voxels that will be out of alignment if between-subject alignment is not good. Smoothing will remedy this, but then again it will also remedy within-subject misalignment.
In general, what would your thoughts on this be? Is it worth sacrificing small within-subject alignment accuracy for much better between-subjects alignment? Or, for longitudinal studies such as this, within-subject alignment should always take priority.
Best Wishes,
Joe
|