What you effectively need to do is create a lookup table for every pixel value that encode both the activity and the density. This might not be possible. I would start by assuming the lowest value (of the three color channels) is the HU and the rest is the colormap of the activity. This will not discriminate between high density and high activity.
Since you're working with a NEMA phantom, wouldn't it be easier to fit a scan of the phantom to your data and subtract it?
I guess this shows why you should get original data, instead of fused images.
