Real-time or offline Data Augmentation for segmenting microscopy images?

I’m doing semantic segmentation(for cells) using microscopy images. I’m exploring U-net and FCN DenseNets for the task. In the U-net paper the authors have trained their model only from 30 images but have used data augmentation extensively specially elastic deformations. I too have very little data, 15-20 annotated images. Thus I too plan to use data augmentation. I’m also using elastic deformations.

Other techniques I’m using:

1. Flipping, Rotation, Translation, Shearing
2. Random crops, zoom-out, stretching
3. Image Contrast and Gaussian noise

This will increase my dataset 10-20 folds.

Q.1 My question is shall I use offline augmentation or on-the-fly(Real time) augmentation?

From what I’ve read so far: use on-the-fly augmentation if dataset if large so to not make the dataset explode in size. I know this is very generalised. I’ve also seen this on Kaggle: For online data augmentation, the model see one random generated sample only once hence generalises better.

Q.2 Also I’m using so many data augmentation techniques will that be even useful as the data set so small, or shall I restrict myself to only some of these techniques?

Any help is highly appreciated.