Oops, Tim just pointed out that you also need to normalize according to the
baseline to convert these to % signal change. In FSL these are also around
10000 but if you want to be exact, you need to divide the scaled PE image
by the mean time series value per voxel. Then you can compute copes by
combining these images.