zaccharieramzi/tf-fastmri-data

TensorFlow data pipelines for the fastMRI dataset

tf-fastmri-data

Built around the tf.data API, tf-fastmri-data offers reliable, unit-tested, datasets for the fastMRI dataset.

Installation

Currently, you need to install the package from source:

git clone https://github.com/zaccharieramzi/tf-fastmri-data.git
cd tf-fastmri-data
pip install .

Example use

from tf_fastmri_data.datasets.cartesian import CartesianFastMRIDatasetBuilder

train_dataset = CartesianFastMRIDatasetBuilder(path='/path/to/singlecoil_train').preprocessed_ds

Data

To download the data, you need to consent to the fastMRI terms listed here. Afterwards, you should receive an email with data download links.

You can then use the environment variable FASTMRI_DATA_DIR to indicate where your fastMRI is. This will allow you to not have to specify the path when instantiating a FastMRIDatasetBuilder.

PyTorch

The PyTorch equivalent of this library is simply the official fastMRI repository. In particular, the data folder is where you find the data utils.

Benchmark

You can run the benchmark script with the following command:

FASTMRI_DATA_DIR=/path/to/fastmri python benchmark.py

Currently the benchmark gives the following output:

Multi coil with tfio loading (random slice): 0.369743709564209s per-file.
Single coil with tfio loading (random slice): 0.02855397939682007s per-file.
Multi coil with h5py loading (random slice, without preprocessing): 0.010439331208042165s per-file.
Single coil with h5py loading (random slice, without preprocessing): 0.0015996736497735258s per-file.
Single coil training with tfio loading: 0.04578723907470703s per-step.

You can also see the recommendation of TensorBoard regarding the single coil dataset (with a very simple model):

Citation

If you use the fastMRI data or this code in your research, please consider citing the fastMRI dataset paper:

@inproceedings{zbontar2018fastMRI,
  title={{fastMRI}: An Open Dataset and Benchmarks for Accelerated {MRI}},
  author={Jure Zbontar and Florian Knoll and Anuroop Sriram and Matthew J. Muckley and Mary Bruno and Aaron Defazio and Marc Parente and Krzysztof J. Geras and Joe Katsnelson and Hersh Chandarana and Zizhao Zhang and Michal Drozdzal and Adriana Romero and Michael Rabbat and Pascal Vincent and James Pinkerton and Duo Wang and Nafissa Yakubova and Erich Owens and C. Lawrence Zitnick and Michael P. Recht and Daniel K. Sodickson and Yvonne W. Lui},
  journal = {ArXiv e-prints},
  archivePrefix = "arXiv",
  eprint = {1811.08839},
  year={2018}
}

tf-fastmri-data's Issues

Document this

• function documentation
• Readme
• full example(s)

Benchmark this

In particular, a screenshot of tf profiling with a simple U-net could be nice.

Add possibility to have manual sharding for the data set

This should be a copy-paste of what has been done recently in fastmri-reproducible-benchmark here.

Implement fake equidistant mask

Right now we only have the possibility to implement the real equidistant mask scheme for retrospective undersampling.

However, due to a slight coding mistake, the so-called equidistant scheme for undersampling in the fastMRI dataset is not really equidistant.

We should implement this mask by simply copy-pasting what was done here, and maybe refactoring a bit.

Why do we restrict to 320x320?

tf-fastmri-data/tf_fastmri_data/dataset_builder.py

Lines 146 to 150 in 17592b3

	self._raw_ds = self._raw_ds.map(
	lambda _, kspace: crop(ortho_ifft2d(kspace), (320, 320)),
	num_parallel_calls=self.num_parallel_calls,
	deterministic=True,
	)

In above code, I feel its best to just not hardcode and rather just crop the data by later pipelines and not initially itself as this makes the data shapes as None, None etc. Also, using adjust_image_size is surely a better alternative (however, I dont want to crop here at all)

Package this

Make this into a pip installable package

Add possibility to batch fastMRI complex data

This should basically be a copy-paste of what has been done lately in the fastmri-reproducible-benchmark, here, with the appropriate adjust size function.

Add auto-increment for version

This can surely be done with a combination of git hooks and bumpversion.

I should try to look for some examples.