This is a small piece of the project I'm working on currently, but still I'd like to publish it here because I hope this may help a little bit for people who're interested in playing with the BraTS 2019 dataset. (UPDATE: That project has been finished by now. The article and source code have been published already. You may find some update of this pipeline in that repository.)
In general, I've reconstructed the pipeline code in this repository with h5py.
To get started, you just need to unzip the downloaded datasets into the brats_2019-data_pipeline/data folder. Then to do the preprocess you can go with
from preprocess import preprocess
preprocess()
The preprocess involves calculating the mean values and the standard deviations of all brain area voxels and making the z-score normalization for each modality. We also do the min-max scaling on every single image. For more details please refer to this article. preprocess() also includes the split for cross validations.
The Dataset class in generator.py is designed for training and validation processes.
import generator
dataset = generator.Dataset()
You could get the iterable generator for training by dataset.train_generator and for validation by dataset.val_generator.
Again, this work is based on my former repository which inherited from ellisdg's repository. Deeply appreciate ellisdg's contributions to the community.
Many thanks to the host of the BraTS challenges.