[Feature] Automated registration target selection #98
Description
The current implementation relies on a manual selection of the specific 2D region, or 3D subvolume of the atlas to be used as the registration target. This is a time-consuming process, and can be error-prone.
We should explore methods to automatically select the region to be registered, based on the data itself. This could be done by using an adaptive grid search, ML based techniques, or Bayesian optimisation.
Adaptive/telescoping grid search could iterate through different combinations of parameters (pitch, yaw, roll, z-position) reducing the step size as results improve.
Bayesian optimisation would use a python implementation to run successive registrations to update the overall loss landscape to find the correct hyperparameters.
All of this relies on settling on a metric, or combination of metrics that correlates to how "good" a registration is.