Paper | GitHub

A deep learning model trained on ImageNet for removing noise and Gibbs artifacts from diffusion images.

This repository contains code for reproducing models from the paper, "Training a Neural Network for Gibbs and Noise Removal in Diffusion MRI" by M. Muckley et al.

This package was developed entirely for research purposes at the NYU School of Medicine and is not affiliated with any other entity. It has not been validated for clinical use.

The workhorse script is degibbs_main.py. Configuration files are in the .yaml format and are stored in the configs/ folder. Options specified at command line overwrite options in the .yaml files, e.g.,

python dldegibbs_main.py magnitude --exp_dir='my_exp_dir/'

should overwrite

exp_dir: "~/data/logs/diffusion/res_256_to_130/"

in the .yaml files.

The package was tested using PyTorch 1.1. PyTorch and other packages were installed using Anaconda.

Pretrained models are available on GLOBUS. Models from the pretrained experiment are saved based on best MSE on the ImageNet validation data set. To use a pretrained model, download one of the model folders from GLOBUS (either res_256_to_100 if you're working on 100 x 100 images or res_256_to_130 if you're working on 130 x 130 images) and change the configs/data.yaml script exp_dir variable to where you save it in your local file system. When running degibbs_main.py, you should get a message that the script found the file and loaded the model.

@article{muckley2019training,
  title={Training a Neural Network for Gibbs and Noise Removal in Diffusion MRI},
  author={Muckley, Matthew J and Ades-Aron, Benjamin and Papaioannou, Antonios and Lemberskiy, Gregory and Solomon, Eddy and Lui, Yvonne W and Sodickson, Daniel K and Fieremans, Els and Novikov, Dmitry S and Knoll, Florian},
  journal={arXiv preprint arXiv:1905.04176},
  year={2019}
}