Introduction

Implementation of and experiments based on the QSMNet (https://arxiv.org/abs/1803.05627).
Now CAD-QSMNet is not an implementation of the original QSMNet any more, but we indeed heavily drew inspiration from it.
This work was done to submit to the QSM Reconstruction Challenge 2.0.

Contributions

Names are listed in alphabetical order:

• Jaewon Chung (@jaywonchung)
• Jinho Park (@jinh0park)
• Yunchan Hwang(@yunchfruit)

Special thanks to Woojin Jung (@wjjung93) for his guidance and helpful advice during the challenge.

Test your idea

First Time Only

1. Clone this repository.

   git clone https://github.com/jinh0park/QSMFirstOrDeath.git

2. Install prerequisites.

   Using Anaconda3.
   torch >= 1.1.0

   • Cuda version == 8.0

   conda install pytorch=0.4.1 cuda80 -c pytorch
   pip install -r requirements.txt

   • Cuda version == 9.0

   conda install pytorch torchvision cudatoolkit=9.0 -c pytorch
   pip install -r requirements.txt

   • Cuda version == 10.0

   conda install pytorch torchvision cudatoolkit=10.0 -c pytorch
   pip3 install -r requirements.txt

Already cloned?

1. Pull the latest changes from the repository.

   git pull

2. Create a new branch in the repository.

   git checkout -b your_branch_name

3. Change parameters in constants.py.

4. Change the code directly.

   • data_utils.py defines how the code loads data from .mat files.
   • model.py defines the QSMNet model layers and their connections.
   • train.py defines the training logic. You can change the optimizer here.
   • loss.py defines the loss functions. train.py calls the total_loss function for the final loss.

5. Start training from scratch with

   python3 train.py

   Or, resume training from a saved training state with

   python3 train.py "path/to/saved/file"

   FYI, train.py saves the training state (epochs done, model weights, optimizer, and learning rate scheduler) every epoch (by default).

6. Check your training status with PyTorch TensorBoard.

   tensorboard --logdir runs --host localhost

   Once tensorboard runs, navigate to the given address (e.g. localhost:6006) on your machine's web browser.

7. When you're done, you may merge the branch to master if it should be part of the baseline model by sending a pull request. If not, you can delete the branch with

   git checkout master
   git branch -d your_branch_name

   But be careful; this will remove all the changes you made to that branch!

Create inferences

inference.py helps you perform inferences with the checkpoint file. It creates .mat and .nii files with inference results on the challenge data and the test set.

1. Requirements
   • The checkpoint file and its path in the form Checkpoints/<model_architecture>/<experience_name>/<model_name>.ckpt.
   • The config.json file for the checkpoint, in the Checkpoints/<model_architecture>/<experience_name> folder.
   • The training data used to train the model. We need the input and output mean/std from it.
   • Run bash ../QSMFirstOrDeath/prepare_inference.sh /path/to/NAS/list3 in the Data folder. This will (hopefully) have you meet all the requirements below this line.
   • In the ../Data folder:
     • Challenge_stage1_sim1_fix.mat
     • Challenge_stage1_sim2_fix.mat
     • Challenge_stage1_ROI_mask.mat
     • Test_0718.mat
     • Test_ROI_mask.mat
   • In the ../Data/Conventional folder:
     • Sim1_Input.nii and Sim2_Input.nii
     • Sim1_{method}.nii and Sim2_{method}.nii, where method is the reconstruction method.
     • Test_{method}.mat where method is the reconstruction method.
     • Also check and edit the two lists named 'methods' in inference.py appropriately. The first one is for the challenge inference, and the second one is for the test set inference.
2. Usage

   python inference.py 'Checkpoints/<model_architecture>/<experience_name>/<model_name>.ckpt'

3. A quick inference function: inf
   • Import:

     from inference import inf

   • Arguments:
     • checkpoint_file: path to the checkpoint file, relative to the QSMFirstOrDeath folder
     • file: (optional) network input of extension .mat, .nii or .nii.gz
     • numpy: (optional) network input of type np.ndarray
     • tensor: (optional) network input of type torch.Tensor
     • device: (optional) which device to use for inference. If not given, takes value from constants.py.
   • Usage: Provide one, two, or all of file, numpy, or tensor.
     • When only one input argument is given, inf returns its inference value. For file and numpy, the output is a np.ndarray. For tensor, the output is a torch.Tensor.
     • When more than one of the input arguments are given, inf returns a tuple of outputs. Inferences are made separately.