Be careful with Amide, some recent version had issues with endianess lately and I would double check the raw image if you have very erratic values. See [3] below:
With a LUT the indexing doesn’t matter, you could choose to randomly shuffle the detector index and you should get the same sensitivity image as computed from geometry. However it is true the aspect of this image will depend on your geometry, but if your positions seems correct I would not expect such noisy output except maybe if the number of detectors if very low.
Of course you can generate your own dataset and feed it to CASToR as a normalization datafile with the proper data channels (with or without correction factor) and give it to CASToR to generate the sensitivity image. This should actually be better than computing from geometry.
Best,
Thibaut