itk:GradientDescentOptimizer seems not to return deterministic results with identical start conditions

You results are very close. To put that magnitude in perspective consider the estimated size of a sub atomic quark is 1e-15 meters. For any practically driven application, this would be defined well with in acceptable tolerances to say the least.

Two things to be aware of with reproducibility of the registration:

  • Look for randomness in the components of the registration such as ImageRegistrationMethodv4:: MetricSamplingReinitializeSeed
  • Consider only using one thread to maximize reproducibility, because only within the limits floating point resolution operations are commutative.

There was also a similar recent post here: Reproducibility of image registration

2 Likes