Hi!
I am also trying to figure out understandable easy ways to load my 3D MRI data into pytorch, here there is an amazing discussion with some code, https://forums.fast.ai/t/custom-data-loader-for-3d-data/27290/9
For keras I wrote a very simple kernel to make dataset from numpy arrays but currently I am changing it to a pandas based method https://www.kaggle.com/meltematay/dataset-maker-for-any-dim-data-mainly-3d-npy
And for handling nifti data (standard format of processed fmris or mris generally)
I know these are very basic…
For Q2, transfer learning approach would not work here since there is no such thing as imagenet in 3D… If someone trained larger datasets with 3D images it would be possible to use transfer learning. But implementing similar networks and trying to train from scratch can also work.