DCRNet
March 20, 2026 · View on GitHub
Accelerating Quantitative Susceptibility and R2* Mapping using Incoherent Undersampling and Deep Neural Network Reconstruction
NeuroImage 2021 | arXiv | data & checkpoints | deepMRI collection
DCRNet recovers both MR magnitude and quantitative phase images from compressed-sensing undersampled k-space data, accelerating QSM and R2* acquisitions using a deep complex residual network.
Note: QSM post-processing (from phase images) requires Linux. Magnitude/phase reconstruction from undersampled data works on all platforms.
Overview
Framework
Fig. 1: Overview of the proposed QSM acceleration scheme.
Network Architecture
Fig. 2: DCRNet architecture built on a deep residual network using complex convolutional operations.
Requirements
For DL-based reconstruction:
- Python 3.7+, PyTorch 1.8+
- NVIDIA GPU (CUDA 10.0+)
- MATLAB R2017b+
For QSM post-processing (Linux only):
Tested on: CentOS 7.8 (Tesla V100), Windows 10 / Ubuntu 19.10 (GTX 1060).
Quick Start
1. Clone and set up environment
git clone https://github.com/sunhongfu/DCRNet.git
cd DCRNet
conda create -n DCRNet python=3.8
conda activate DCRNet
conda install pytorch cudatoolkit=10.2 -c pytorch
conda install scipy
pip install mat73
2. Download demo data and checkpoints
Download from Dropbox and place in the repo root.
3. Run demo
conda activate DCRNet
# Single-channel
matlab -nodisplay -r demo_single_channel
# Multi-channel
matlab -nodisplay -r demo_multi_channel
Reconstruction on Your Own Data
Edit parameters in lines 10–20 of the demo script, then run:
# Single-channel
conda activate DCRNet
matlab -nodisplay -r demo_single_channel
# Multi-channel
conda activate DCRNet
matlab -nodisplay -r demo_multi_channel
Training
% 1. Prepare training data
matlab -nodisplay -r PrepareTrainingData
# 2. Train DCRNet
cd PythonCodes/training
python TrainDCRNet.py
Citation
@article{dcrnet2021,
title={Accelerating quantitative susceptibility and R2* mapping using incoherent undersampling and deep neural network reconstruction},
journal={NeuroImage},
year={2021},
doi={10.1016/j.neuroimage.2021.118404}
}