Why do you say that the "whole big blob is not meaningful"? If that is the case, why would some subset be any more meaningful?
FYI, the 3 peak voxels of the cluster are just based on the default setting of spm_list (I think that's the function), of displaying only the top 3 voxels per cluster that are >= 8mm apart. So even those 3 most significant voxels might not be meaningful, per se. i.e. how could it be determined that, if my local spm_list's default setting is to display the top 3 voxels >= 6mm apart, my choice is more/less legitimate than yours?
I think most people would agree that you should choose an ROI based on some a priori hypothesis. And not based on the SPM results. That might leave you open to an accusation of "double dipping".
More specifics on your end would certainly help in answering your question; what patient group are you studying, what is the experiment in question, etc. A very simple example would be if you are using a verb generation task; you would probably be most interested in left hemi language areas, and choose an ROI mask of those areas for your analysis.
________________________________________
From: SPM (Statistical Parametric Mapping) [[log in to unmask]] on behalf of Jonathan Beck [[log in to unmask]]
Sent: Tuesday, October 02, 2012 7:51 PM
To: [log in to unmask]
Subject: [SPM] ROI Definition
Dear SPM Experts,
These questions are not about SPM per se, but more about how best to
use SPM. I am trying to figure out the most principled way to examine
betas from second-level contrasts as part of an ROI analysis (i.e. the
betas that are frequently graphed alongside whole brain maps to
mitigate against blobology, and to show that findings are not spurious
due to subtractions of negative betas).
For a control-patient contrast, I have a fairly large cluster (~325
voxels) that is statistically significant (FWE-corrected, T=2.41). The
SPM results view shows the 3 peak voxels in this cluster. The whole
big blob is not meaningful, so I need to figure out a principled way
to break it down into smaller units. What is the best way to define
smaller-than-cluster ROIs from which to extract and graph betas?
(1) Should I investigate the most significant voxel with a sphere around it?
(2) Should I pull betas from the area defined by the overlap of
spheres around the 3 peak voxels?
(3) Should I define an ROI using the center of mass (determined in
Marsbar) with a sphere around that?
(4) Or should I just pull betas from the whole monster cluster?
I understand that using an atlas to mask an anatomically-defined
region can yield more meaningful results, and I know it is common to
investigate a conjuction of an atlas-defined region and a significant
cluster. The question then becomes: Which brain structure should I
choose? The structure that contains the most voxels in the
significant cluster? The structure that contains the peak voxel? The
structure of most interest based on a priori hypothesis? Perhaps it
all depends on the size of the cluster and/or the size of the brain
structure.
If anyone can shed some light on these questions, I would really
appreciate it! Any good resources I could turn to for guidance?
Thanks!
Jonathan
______________
Jonathan Beck
Junior Specialist
Department of Psychiatry and Behavioral Sciences
UC Davis MIND Institute
|