The current N4 bias correction binary is not deterministic across platforms

ledigchr commented 3 years ago

This is an important issue reported with suggested fix here: https://github.com/MIRTK/MALPEM/issues/2

ledigchr commented 3 years ago

Some experiments around that seem to confirm what is also discussed in this thread by the N4 developers.

The N4 binary that is shipped with MALPEM yields slightly different results if it runs with a varying number of Threads - seems to be a problem with ITK.
The number of ITK Threads defaults to the number of CPUs on the system. So on the same system N4 (and as consequence MALPEM) should produce the same results.
The number of Threads can be specified with the environment variable ITK_GLOBAL_DEFAULT_NUMBER_OF_THREADS. So if this is set explicitly, N4 should be reproducible across systems with varying number of CPUs.
I've tested this on AWS instances and reproduced a default run on m5a.8xlarge (32 CPUs) on a t2.xlarge (4 CPUs, but with ITK_GLOBAL_DEFAULT_NUMBER_OF_THREADS=32).

My suggested fix for the MALPEM package is to update malpem-proot and explicitly set ITK_GLOBAL_DEFAULT_NUMBER_OF_THREADS. E.g.

'XAUTHORITY=/tmp/.mdmCHXAXX'  \
"${PROOT}" \

-->

'XAUTHORITY=/tmp/.mdmCHXAXX'  \
'ITK_GLOBAL_DEFAULT_NUMBER_OF_THREADS=32' \
"${PROOT}" \

ledigchr commented 3 years ago

Fixed in malpem-v1.3 by setting 'ITK_GLOBAL_DEFAULT_NUMBER_OF_THREADS=8' in malpem-proot

ledigchr / MALPEM