❗ These implementations are based on the original fastMRI repository by Facebook Research. The original repository can be found here.

Deep Learning-based MRCP Reconstruction

Deep Learning-based (DL) reconstruction framework for Magnetic Resonance Cholangiopancreatography (MRCP) imaging. We use ResNet-based DL models for supervised (SV) and self-supervised training. We train DL models on the 3T six-fold retrospectively undersampled MRCP. We evaluate the model for six-fold retrospective and prospective undersampling acquired at 3T and 0.55T.

Features of the Framework

• Input: SENSE-based synthesized k-space data on six-fold retrospective undersampling instead of zero-filled k-space data
• Target: GRAPPA reconstruction of two-fold accelerated MRCP (clinical standard)
• Sensitivity maps: Predefined using an ESPIRiT algorithm
• Unrolled network: The Variational Network model

Installation

1. Clone the repository then navigate to the MRCP_DLRecon root directory.

git clone [email protected]:JinhoKim46/MRCP_DLRecon.git
cd MRCP_DLRecon

2. Create a new conda environment

conda env create -f environment.yml

1. Activate the conda environment

conda activate mrcp_dlrecon

1. Install a dlrecon package

pip install -e . 

Usage

Data Preparation

• MRI data are stored in the HDF5 format containing the following structures:
  • Datasets:
    • grappa: target data (y $\times$ x $\times$ Slice)
    • kdata_raw: Raw $k$-space data (x2) (nCoil $\times$ PE $\times$ RO $\times$ Slice)
    • kdata_fs: Fully-sampled $k$-space data from kdata_raw using GRAPPA (nCoil $\times$ PE $\times$ RO $\times$ Slice)
    • sm_espirit: ESPIRiT-based sensitivity maps (nCoil $\times$ y $\times$ x $\times$ Slice)
• The sample_data directory contains sample MRCP data for training, validation, and testing. We provide two two-fold (2x) 3D MRCP and one six-fold (6x) 3D MRCP. In the sample_data/dataset.csv, the two 2x MRCP data are defined for training and validation, and the 6x MRCP data is marked for testing. You can find the data here. Additional information, such as header information, is ignored in the sample data.
• Data splitting for training, validation, and testing is done by the sample_data/dataset.csv file.
• Replace data_path in the configs/paths.yaml file with the actual path to the data.

Run

Train

1. Define the training configurations in the configs/dlrecon.yaml file.
2. Run main.py by

python main.py fit --config configs/dlrecon.yaml

1. You can define the run name by adding the --name argument at run. Unless you define the run name, it is set to %Y%m%d_%H%M%S_{training_manner} with the current date and time.

python main.py fit --config configs/dlrecon.yaml --name test_run

1. You can overwrite the configurations in the configs/dlrecon.yaml file by adding arguments at run.

python main.py fit --config configs/dlrecon.yaml --model.training_manner ssv

5. Log files containing checkpoints/, lightning_logs/, and script_dump/ are stored in log_path/run_name. log_path is defined in the configs/paths.yaml file.
6. You can resume the training by giving run_name with fit command. *.ckpt file should be placed in run_name/checkpoints/ to resume the model.

python main.py fit --config configs/dlrecon.yaml --name run_name

Test

1. Run main.py with run_name by

python main.py test --config configs/dlrecon.yaml --name run_name

1. *.ckpt file should be placed in run_name/checkpoints/ to test the model.
2. The output files are saved in log_path/run_name/npys/FILENAME.