Good point. I think it really depends on what you expect. In case you assume that the brain stays the same (more or less) over time, then rigid body transformations should be sufficient, and then it makes sense to create a mean out of the different volumes and one normalization based on this volume instead of performing a bunch of normalizations.
The problem is that coregistration between different modalities / weightings like EPI and T1 will always be imperfect to some degree. Whereas in the one session, the structural volume might be a little "too high" compared to the functional mean and tilted in one direction after coregistration, the other structural volume might be a little "too low" and tilted in the other direction. The normalizations (including a coregistration step between the structural volumes and the T1 template) might also result in slight differences. Thus it seems to be more promising to me to realign within modalities and perform one coregistration and one normalization only.
If the time points are very different, then the question is whether rigid body transformations are sufficient at all. If you really expect large changes in density, increases in ventricles, ..., then very probably they are not sufficient (in theory at least). Registrating to the middle time point instead of the mean would be suboptimal as well. Instead one should probably use non-linear registration methods to create some subject-specific "average brain" out of the different time points (by the way, in case of relevant brain changes one should account for changes in GM density/volume by including the GM density as regressor into the fMRI GLM probably).
If this is what one wants to. I could imagine conditions in which one might still prefer rigid body transformations between different time points for fMRI analysis. Imagine very specific atrophy in, let's say hippocampus. What would happen when applying non-linear transformations? Maybe this would result in an artificial increase/deformation of structures nearby unaffected by atrophy?
Anyway. One thing are detectable brain changes, the other thing is whether they are large enough to affect the normalization parameters as such. Scanning a large amount of healthy subjects one year apart might result in nice ageing effects in VBM, but probably there's hardly any difference in the normalization parameters.
Dunno whether this has ever been tested. Maybe people like John Ashburner (DARTEL algorithm) or Christian Gaser (VBM8 toolbox including preprocessing pipeline for longitudinal VBM) have run simulations on this topic? Or maybe the Alzheimer's Disease Neuroimaging Initiative?
|