I realize that much of this is a matter of preference, with no absolutely correct or incorrect approaches, but I'd like to get some opinions on how regions of interest ought to be defined. Based on reading, conversations with others who are more experienced at ROI analyses than I am, and trial and error, I've identified a few different methods for defining ROIs. I'll list and briefly describe each. If possible, I'd like others' thoughts/critiques on each method (and please suggest other methods). Again, I realize that there's not a right or wrong way, but I'd like to get some feedback in order to improve my procedures and to try to anticipate reviewers' critiques before submitting anything for publication. My ultimate goal is to have a reasonable number of ROIs for which I can compare IRFs and/or number of significant voxels and/or mean voxel intensity for each condition.
1. Create a mask for each subject by drawing the ROI(s) on each subject's un-Talairached brain, based on a priori hypotheses about structures of interest. This is time-consuming, and for some structures, may require better knowledge of anatomy than I currently possess.
2. Create the mask by using a Talairach atlas. This has the advantage of requiring that only one mask dataset be created, which can be applied to each subject individually. However, I'm wondering (perhaps based on a post that I vaguely remember reading here) if there are problems with Talairach transforming IRF functions. If not, I like this method, because it's fast and objective in the sense that it doesn't require that I identify structures.
3. Define ROIs functionally, based on significant clusters of activation (as determined by AlphaSim), and identifying structures based on where the most intense voxel or geometric center is per cluster. I like this method, but I've had problems with determining how to separate what may be distinct but adjacent ROIs that are technically part of the same cluster.
4. Creating spheres of some diameter around local maxima (for the combined effect of all conditions vs. baseline, with a very conservative and somewhat arbitrary threshold). I've experimented with this method using the Maxima plugin, on the advice of a colleague. With some datasets, it seems to work well. With others, I get an impractical number (dozens, or even hundreds) of spherical ROIs to sort through. I can focus on spheres that appear in structures of interest, but what do I do with the others? Also, if a local maximum is near the surface, much of the sphere will be outside the brain.
Ultimately, I would like to come up with a standardized method, or at least have a system for when to take what approach. Thanks for any suggestions. Sorry I couldn't be more concise about asking this.