Hi Jen,
In addition to Rick's recommendation, I should mention that usually you need not map smoothed time series back to the volume. Most 3d* programs, the ones that do not involve inter-voxel operations (like 3dclust, or 3dmerge -1blur*), will work on surface-based data.
Also, if you are just mapping meridia in this experiment, it might be easier to use linear regression with 3dDeconvolve to identify responses to the two or 4 stimulus classes you used.
cheers,
Ziad