3D two-channel network for post-processing learning

[martaranzini]

Implement and test a 3D network that takes the image and the output of the 2D segmentation network to learn the post-processing. This network should act as a segmentation refinement which takes into account the 3D information discarded by the 2D network.

This can be thought as an alternative to the 3D UNet for heatmap predictions (i.e. sequential to the 2D segmentation network instead of parallel to it).

Issues to be considered:

• Correct dimensioning of the architecture (two-channel and 3D)
• Loss function?
• Sampling scheme?
• Normalisation of the channels