Time for reading/writing uncompressed binary files is negligible (compared to registration or resampling time), even for very large files. Generating the resampled output and displacement field are time-consuming, so that deserves some attention.
- I think the resampled output volume is generated twice (once by elastix, at the end of each registration phase; once by transformix at the end - these are redundant).
- We could avoid id completely by changing Elastix to write ITK-compatible transforms or implement Elastix transform readers in SlicerElastix.
Contributions are welcome!