3dBlurToFWHM is designed to work on the noise in FMRI time series, and it smooths the residuals of the time series after removing a trend to get something like the "noise".
For PET images, where the concept of time series is less applicable, you could try using 1 input image at a time. See what the results look like. However, I'd have to think about it more -- but it's time for some eggnog with rum.