MRI Puns Project

Outline

Dependencies and Installation

We have tested this code using:

• Ubuntu 18.04 / Windows 10 / Mac OS X (Catalina)
• Python 3.8
• CUDA 10.1
• CUDNN 7.6.5
• 1 / 4 CUDA-enabled GPUs

First install PyTorch according to the directions at the PyTorch Website for your operating system and CUDA setup.

Then, navigate to the root directory and run:

pip install -e .
pip install -e dev-requirements.txt

pip will handle all package dependencies. After this you should be able to run most the code in the repository.

Download the dataset from the fastMRI dataset page, and unzip it appropriately. This project only uses the Single Coil dataset.

Training the models

Navigate inside the experimental/<model_name>/ folder and run the demo file inside with no arguments. For example:

cd experimental
cd nnret
python train_nnret_demo.py

Graphing the training results

After training a model, take the console output, place it in a textfile and place this file inside a new folder. Then place the folder inside results. Alter the run() method in results/grapher.py to include the output and run:

python grapher.py

Test using the models

Follow the same steps as training except call the demo file with --mode test. For example:

python train_nnret_demo.py --mode test

Extracting reconstructed images from test output

Navigate inside the experimental/<model_name>/ folder and run the demo file inside with arguments. For example:

cd experimental
cd unet
python train_unet_demo.py --data_path ..\..\data --mode test --recon True

--data_path specify the location of the file to test with. --mode specify the type which is test in our example. --recon specify the reconstruction parameter.

The python script will automatically pick the files to test under the specified directory from singlecoil_test directory.

If the reconstruction is successful, the images will be placed under <file_name> directory under experimental/<model_name>/

In our example, the file name is file1000000.h5 which has 36 slices. We pick slice 22 for reconstruction because it resembles a complete knee.

The images stored are the input image to the model(file1000000.h5_22_image), the output image from the model(file1000000.h5_22_output) and the target image(file1000000.h5_22_target).

Code References

The bulk of this repository is from https://github.com/facebookresearch/fastMRI. We made alterations/augmentations and added new models to the codebase.