Setup fastMRI environment

conda create --name fastmri
conda activate fastmri
conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch # haven't tested with cudatoolkit=11.3 yet...
# conda install pytorch=1.9.1 torchvision torchaudio cudatoolkit=11.1 -c pytorch -c nvidia # that's what I use right now...

Install requirements

pip install -r requirements.txt

Environment variables

export FASTMRI_ROOT=<your-fastmri-root-dir>

fastMRI setup

• $FASTMRI_ROOT
  • knee
    • data
  • adni
    • data

Generate csv file

FASTMRI

python ./data_processing/create_dataset_csv.py --data-path ${FASTMRI_ROOT} --csv-file=../datasets/singlecoil_knee_train.csv --dataset knee/data/singlecoil_train  

python ./data_processing/create_dataset_csv.py --data-path ${FASTMRI_ROOT} --csv-file=../datasets/singlecoil_knee_val.csv --dataset knee/data/singlecoil_val  

python ./data_processing/create_dataset_csv.py --data-path ${FASTMRI_ROOT} --csv-file=../datasets/singlecoil_knee_test.csv --dataset knee/data/singlecoil_test  

ADNI

python ./data_processing/create_dataset_adni.py --data-path ${FASTMRI_ROOT} --dataset adni/data/singlecoil --original-data-path <ADNI Path> --input-csv-file <ADNI CSV Path> --split-file ./datasets/singlecoil_adni_split.json --csv-file ./datasets/singlecoil_adni 

ADNI Path must contain the image wise xml files and the image wise folders ADNI CSV Path must contain "DXSUM_PDXCONV_ADNIALL.csv", "FOXLABBSI_.csv", "idaSearch_.csv" where * is replaced by a date

Test dataloader

First, adapt the <fastmri-root-dir> in fastmri_dataloader/config.yml

python -m unittest fastmri_dataloader/fastmri_dataloader_th.py

Train

rename config_adni_train.yml or config_knee_train.yml to config.yml

python train.py --regularizer Real2chCNN --num_workers 1 --aleatoric --l2 --save_checkpoint True --combined-extra-network --run-name multivariate_aleatoric

Predict

rename config_adni_predict.yml or config_knee_predict.yml to config.yml change both accelerations: and center_fractions to a single value within config.yml

python test.py --regularizer Real2chCNN --aleatoric --l2 --non-diag-rank 8 --log-low-rank --ckpt_path <path to checkpoint>/checkpoint_epoch100.pth --cal-metrics --save-hdf5 --mode singlecoil_train1 --save-gnd --combined-extra-network --num_workers 8 

repeat for all accelerations