Random seed seems to have no effect on ANTs with SyN and CC

[gandreadis]

Describe the problem

I'm using the ANTs antsRegistration CMD-line executable, with SyN as transform and CC as metric. I'm using the --random-seed option to set the value, but no matter what I set it (0, which should use system time, or any positive non-0 value, which should set that as seed), the output and metric/convergence values over time are always exactly the same. I've also tried the ANTS_RANDOM_SEED environment variable, but no success with this, either.

Following the wiki, I've also tried to specify a sampling strategy for CC (random with percentage [0-1]), but this does not change the result. I've used --verbose to verify that this sampling strategy is indeed set.

To Reproduce Use the antsRegistration script with two images, using the SyN transform with CC, and try to set the random seed with --random-seed [int].

System information (please complete the following information)

• OS: Linux CentOS
• OS version: Stream release 8
• Type of system: HPC cluster

ANTs version information

• ANTs code version: v2.4.2.post39-gbaa2bd1
• ANTs installation type: Compiled from source

Additional information

[gdevenyi]

Have you read https://github.com/ANTsX/ANTs/wiki/antsRegistration-reproducibility-issues

[stnava]

short story - this is expected as CC is not randomized

[gandreadis]

@stnava That explains it, thanks. Is it possible to use a different sampling strategy with CC? Or is CC always deterministic? Because in that case it might be good to mention this somewhere.

[cookpa]

This is a bit confusing because the metric sampling strategy is an option of the metric (eg, CC) in antsRegistration, but it's a property of the registration type (eg, Rigid, SyN) in ITK.

If you choose any sampling strategy other than "None", the registration method may use random sampling to choose and/or perturb the points. This leads to variation in results. But the SyN method doesn't do that.

You can make CC non-deterministic with another transform, eg -t Affine[0.1] -m CC[ fixed, moving, 1, Random, 0.5 ] will be non-deterministic, because the affine registration method calls ImageRegistrationMethodv4:SetMetricSamplePoints(). I don't know if any of the deformable methods do.

[gandreadis]

@cookpa Thanks for the elaboration, that makes it clear. Maybe something to this effect could be added to the respective wiki entries about reproducibility? Because I can't imagine being the only one to only use the SyN stage (without affine) and trying to set the random seed parameter.

[cookpa]

Yes, I will try to explain it better

[cookpa]

Update, I tried to improve the wiki page

https://github.com/ANTsX/ANTs/wiki/antsRegistration-reproducibility-issues

I looked at the code again and I think SyN is the only transform option that doesn't call PrepareRegistrationMethod, so I think it's the only method that won't produce any variation with random subsampling in the metric.