$\chi$-sepnet (chi-sepnet)

• The code is for reconstructing susceptibility source-separated maps by deep neural network ($\chi$-sepnet, chi-sepnet).
• Matlab toolbox including conventional source separation method ($\chi$-separation) is also available (https://github.com/SNU-LIST/chi-separation.git).
• The source data for training can be shared to academic institutions. Request should be sent to [email protected]. For each request, internal approval from our Institutional Review Board (IRB) is required (i.e. takes time).
• Don't hesitate to contact for usage and bugs: [email protected]
• Last update : Nov 11, 2024

Usage

⭐ If you have both GRE and SE data, you have option for chi-sepnet-R2' (better quality).

⭐ If you only have GRE data, neural network (chi-sepnet-R2*) will deliver high quality susceptibility source-separated maps.

⭐ If you acquired data with different resolution from 1 x 1 x 1 mm³, the resolution generalization method (can process resolution > 0.6 mm; check the reference) is required.

⭐ If you acquired data with different B₀ direction from [0, 0, 1], the B₀ direction correction to [0, 0, 1] is required.

⭐ Input data with the same orientation with trained data (check the figure below) is recommended.

Inference

You can follow the steps below for the inference.

1. Clone this repository

   git clone https://github.com/SNU-LIST/QSMnet.git

2. Create conda environment via downloaded yaml file

   conda env create -f chisepnet_env.yaml

3. Activate xsepnet conda environment

   conda activate xsepnet

4. Run the inference code

   python test.py

References

M. Kim, S. Ji, J. Kim, K. Min, H. Jeong, J. Youn, T. Kim, J. Jang, B. Bilgic, H. Shin, J. Lee, $\chi$-sepnet: Deep neural network for magnetic susceptibility source separation, arXiv prepring, 2024

S. Ji, J. Park, H.-G. Shin, J. Youn, M. Kim and J. Lee, Successful generalization for data with higher or lower resolution than training data resolution in deep learning powered QSM reconstruction, ISMRM, 2023