Thanks very much for your suggestions.
To test your suggestion, I created six concatenated files of unique brain images (n=13-16 brains each). I calculated agreement of FWHMx smoothness estimates across the six files by taking the stdev / mean. I found that for my sets, I had the best agreement amongst the group when using the combination of (-2difMAD -detrend 2). I had lesser agreement for other combinations of options: (-2difMAD -demed) > (-demed) > (-detrend 2).
As an aside, I am actually using this metric to analyze confocal microscopy data, not actually PET. The similarity is that I have a single image of intensity-based data for each brain somewhat confounded by structural information.