3D UNet for heatmap prediction

[martaranzini]

Develop training and testing code for a 3D UNet that can serve as post-processing helper - to include 3D information in the connected component analysis.
Goal: to output a probability map (rather than a one-hot segmentation) indicating regions most likely to contain the fetal brain. This can be combined with the output of the 2D segmentation network to select the connect component most likely to correspond to the brain.

Issues to be considered:

• Patch size?
• Balance positive vs negative samples?
• Loss function?