A fair comparison in deep learning MRI reconstruction

This is a source code utilized for experiments of "A fair comparison in deep learning MRI reconstruction"

Requirements

• Conda (https://docs.conda.io/en/latest/)
• MATLAB 2016b (https://www.mathworks.com/products/matlab.html)
• FSL (https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/FSL)
• ANTs (https://github.com/ANTsX/ANTs)
• CUDA 10.0
• NVIDIA TITAN Xp GPU

Usage

Step 0: First time only

1. Create conda environment

conda env create -n qsmnet -f requirements.yaml

2. Data: please organize the qsmnet data according to the format below

data  
  └ D111
      └ train2  
          └ IMG.mat  
      └ train3   
          └ IMG.mat    
      └ ...
  └ D113
  └ ...

Step 1: dataset generation (MATLAB & Python)

1. Dataset generation & augmentation (matlab)

cd code_for_dataset_generation
matlab
final_run

2. Generate patches for deep learning (python)

cd code_for_dataset_generation
conda activate qsmnet
python patch.py

Step 2: QSMnet training & inference (Python)

1. Activate conda environment

cd code_for_neural_network
conda activate qsmnet

2. QSMnet training
   If you want to change hyperparameters, please replace the default value in parser arguments of "train.py"

python train.py -g 0 -s ../save/qsmnet111

3. QSMnet inference
   If you want to change hyperparameters, please replace the default value in parser arguments of "test.py"

python test.py -g 0 -s ../save/qsmnet111/result -e 25

4. Calculate evaluation metrics (NRMSE, SSIM, PSNR, HFEN) of inferenced maps

python testmetric.py -p ../network/qsmnet111/result