The first fastMRI challenge provided a great opportunity to push the limits of MRI data acquisition speed further using deep learning. Instead of using a single best model, we investigate several network architectures for multicoil MR image reconstruction, defined in a sigmanet, including sensitivity networks and parallel coil networks, various data consistency layers, regularization networks and (semi-)supervised learning schemes. Sigmanet by our team holykspace is among the top-three entries in the public and challenge leaderboard for multicoil MR image reconstruction (December 2019).

Our framework allows for a systematic evaluation of various iterative deep neural networks for image enhancement and image reconstruction. This repository builds upon the fastMRI source code and contains code for data processing, training, testing and evaluation for the fastMRI dataset used in our publications

• Hammernik et al., Sigmanet: Systematic Evaluation of Iterative Deep Neural Networks for Fast Parallel MR Image Reconstruction, arXiv preprint arXiv:1912.09278, 2019.
• Schlemper et al., Sigmanet: Ensembled Iterative Deep Neural Networks for Accelerated Parallel MR Image Reconstruction, arXiv preprint arXiv:1912.05480, 2019.

If you use this code for your research please cite

@article{hammernik2019sigmanet,
    author = {Hammernik Kerstin and Schlemper Jo and Qin Chen and Duan Jingming and Summers Ronald M. and Rueckert Daniel},
    title = {$\Sigma$-net {Systematic Evaluation of Iterative Deep Neural Networks for Fast Parallel MR Image Reconstruction}},
    journal = {arXiv preprint arXiv:1912.09278},
    year = {2019},
    archivePrefix={arXiv},
    eprint={1912.09278},
}

• sigmanet (ROOT)
  • data_processing
  • ipynb
  • reconstruction (based on the fastMRI source code)
    • common
    • data
    • models

1. Setting up an environment following the fastMRI repository
2. BART toolbox for coil sensitivity map estimation using ESPIRiT.
3. medutils

1. It might be easiest to setup an environment variable ${FASTMRI} pointing to the fastMRI data root directory. We suggest following structure for the data

   • ${FASTMRI}
     • data
       • multicoil_train
       • multicoil_val
       • multicoil_test_v2
     • datasets

2. Creation of *.csv files with meta information about the datasets by running

   python data_processing/create_dataset_csv.py --csv-path ${FASTMRI}/datasets --data-path ${FASTMRI}/data --dataset multicoil_train

   We provide a jupyter notebook in ipynb/csv_stats.ipynb to get some statistics of the datasets.

3. Estimation of coil sensitivity maps using ESPIRiT by running the file

   python data_processing/estimate_sensitivities.py --csv-path ${FASTMRI}/datasets --data-path ${FASTMRI}/data --dataset multicoil_train --acl 15 30

   This requires a proper setup of the BART toolbox. The flag --acl specifies the number of low frequencies / auto-calibration lines for which the sensitivity maps should be computed and is only used for the train and validation dataset. For the test and challenge dataset, we use the provided number of low frequencies for sensitivity map estimation. For more details on the *.h5 files, we refer to our article and to the source code. The data is stored with an additional suffix _espirit in the datafolder.

4. (optional) However, in many cases reading the *.h5 data is extremely slow. To reduce the dataset and improve the dataloading, we recommend running

   python data_processing/copy_data_to_local.py --csv-path ${FASTMRI}/datasets --data-path ${FASTMRI}/data --out-path ${FASTMRI}/data_reduced --dataset multicoil_train

   This crops the dataset in Frequency Encoding (FE) direction without introducing new artifacts. If an additional --float16 flag is provided, data are stored as float16 which additionally reduces memory. If resources allow, try to copy as much data as you can to your local disk to further speed-up data loading.

5. (optional) For training and testing, we used foreground masks which we store for the individual datasets with the suffix _foreground in the datafolder. We provide the foreground masks for the knee dataset upon request.

Note: The data processing works for both the fastMRI knee and brain dataset. The number of ACL lines for training and testing was fixed for the knee dataset to 15 for R=8 and 30 for R=4 by examining the numbers of ACL lines given in the test dataset. A different setting might be chosen for the brain dataset.

Once the datasets are setup and sensitivity maps are estimated, the new sensitivity-weighted target for the evaulation on the validation set can be computed using

python data_processing/create_sense_target.py --data-path=${FASTMRI}/data --csv-path=${FASTMRI}/datasets --dataset=multicoil_val --acl 15 30 --acceleration 8 4 --out-path ${FASTMRI}/results/sense_target

If additionally the flag --mask-bg is provided, the background regions are masked out. Therefore, foreground masks are necessary, see previous section / 5. This is only used for final evaluation and not during training.

To make a first test of the implementation, we encourage to run the zero-filled reconstruction from reconstruction/models/zero_filled_sense/

CUDA_VISIBLE_DEVICES=0 python run_zero_filled_sense.py --challenge multicoil --data-path ${FASTMRI}/data --csv-path ${FASTMRI}/datasets --data-split val --out-dir ${FASTMRI}/results/zf/R8 --acceleration 8 --center-fractions 0.04

All networks can be run from the directory reconstruction/models/sigmanet. An example training to train a sensitivity network can be started as follows.

CUDA_VISIBLE_DEVICES=0 python train.py --challenge multicoil --data-path ${FASTMRI}/data --csv-path ${FASTMRI}/datasets --exp-dir ${FASTMRI}/trained_models/sn --num-iter 8 --data-term PROX --regularization-term dunet --batch-size 1 --num-epochs 40 --fe-patch-size 96 --optimizer rmsprop --lr 1e-4 --use-fg-mask --lambda-init 10 --learn-data-term --shared

• --use-fg-mask uses pre-computed foreground masks for training
• --data-term can be NONE, GD, PROX, VS
• --lambda-init defines the initialization of the dataterm weight. This is crucial for the learning behaviour.
• --learn-data-term Learns the dataterm weight during training
• --shared denotes shared parameters over the number of cascades defined via --num-iter
• --regularization can be dunet or unet

A detailed description of the flags can be found in the source code.

Note: The training and testing code was originally designed for the knee dataset with fixed base resolution of (320 x 320) and might not fully support the brain dataset yet.

Testing of sigmanet networks can be conducted in the directory reconstruction/models/sigmanet. To run a trained model you can use following command

CUDA_VISIBLE_DEVICES=0 python run.py --challenge multicoil --data-split val --data-path ${FASTMRI}/data --csv-path ${FASTMRI}/datasets --out-dir ${FASTMRI}/results/sn --checkpoint ${FASTMRI}/trained_models/sn/model.pt --accelerations 4 --center-fractions 0.08 --mask-bg

The flag --mask-bg applies the foreground mask (if available).

The evaluation script of fastMRI is extended to work with sensitivity-weighted targets. For evaluation, you can use following command from the root directory of this repository.

python reconstruction/common/evaluate.py --challenge=multicoil --sense-target --dataset knee --target-path=<sense-target-path> --predictions-path=<prediction-path>

The flag --sense-target indicates that the sensitivity-combined targets are used which are stored with the key reconstruction_sense when creating the new targets as described above. The evaluation script also works for root-sum-of-squares (RSS) targets as in the original version by removing the --sense-target flag.

Sigmanet is MIT licensed, see LICENSE file.

Sigmanet builds upon the fastMRI source code, which is also MIT licensed.