An empirical result, for future SPM list archaeologists:
I tried out converting the deformation field to int16. I did this using SPM's nifti function, changing N.dat.dtype and creating a new image, then copying in the data. I believe this sets the scale factor accordingly. For my test image, the warps I got out (in my case atlases being warped to native space) looked indistinguishable to the eye from those using the float32 deformation field. My test criterion was that observed vs expected images should have a Pearson correlation > 0.99. It failed because it was 0.98. For my use case I decided to just replace the expected image. I'm only changing the data type to get around the size limit in the testing system. For real runs, it will stay float32.
Paul
|