QSMnet & QSMnet⁺

• The code is for reconstructing Quantitative Susceptibility Mapping by deep neural network (QSMnet) and QSMnet⁺. QSMnet⁺ covers a wider range of susceptibility than QSMnet, using data augmentation approach. Data preprocessing, training and inference code of QSMnet (.py) are availabe.
• The source data for training can be shared to academic institutions. Request should be sent to snu.list.software@gmail.com. For each request, internal approval from our Institutional Review Board (IRB) is required (i.e. takes time).
• Last update : 2022.06.09

References

• QSMnet </br> J. Yoon, E. Gong, I. Chatnuntawech, B. Bilgic, J. Lee, W. Jung, J. Ko, H. Jung, K. Setsompop, G. Zaharchuk, E.Y. Kim, J. Pauly, J. Lee. Quantitative susceptibility mapping using deep neural network: QSMnet. Neuroimage. 2018 Oct;179:199-206. https://www.sciencedirect.com/science/article/pii/S1053811918305378
• QSMnet+ </br> W. Jung, J. Yoon, S. Ji, J. Choi, J. Kim, Y. Nam, E. Kim, J. Lee. Exploring linearity of deep neural network trained QSM: QSMnet+. Neuroimage. 2020 May; 116619. https://www.sciencedirect.com/science/article/pii/S1053811920301063</br>
• Review of deep learning QSM </br> W. Jung, S. Bollmann, J. Lee. Overview of quantitative susceptibility mapping using deep learning: Current status, challenges and opportunities. NMR in Biomedicine. 2020 Mar; e4292. https://doi.org/10.1002/nbm.4292

Overview

(1) QSMnet

(2) QSMnet⁺

Requirements

• Python 3.7

• Tensorflow 1.14.0

• NVIDIA GPU (CUDA 10.0)

• MATLAB R2015b

Data acquisition

• Training data of QSMnet was acquired at 3T MRI (SIEMENS).
• 3D single-echo GRE scan with following sequence parameters: FOV = 256 x 224 x 176 mm³, voxel size = 1 x 1 x 1 x mm³, TR = 33 ms, TE = 25 ms, bandwidth = 100 Hz/pixel, flip angle = 15°.

Manual

First Time Only

1. Clone this repository

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

2. Download network </br> In Checkpoints directory,

• For Linux User,

sh download_network.sh

• For Windows User, </br> https://drive.google.com/drive/folders/1E7e9thvF5Zu68Sr9Mg3DBi-o4UdhWj-8 </br> and unzip the files in 'Checkpoints/' </br>

3. Create conda environment

conda create --name qsmnet -c conda-forge -c anaconda --file requirements.txt 

Phase processing

• Requirements

  • FSL (Ref: S.M. Smith. Fast robust automated brain extraction. Human brain mapping. 2002 Sep;17(3):143-155.)
  • STI Suite (Ref: W. Li, A.V. Avram, B. Wu, X. Xiao, C. Liu. Integrated Laplacian‐based phase unwrapping and background phase removal for quantitative susceptibility mapping. NMR in Biomedicine. 2014 Dec;27(2):219-227.)

• Process flow

  • Extract magnitude and phase image from DICOMs
  • Brain extraction : BET (FSL)
  • Phase unwrapping : Laplaican phase unwrapping (STI Suite)
  • Background field removal : 3D V-SHARP (STI Suite)

• Usage:

  • If you acquired data with different resolution from 1 x 1 x 1 mm³,</br> you need to interpolate the data into 1 mm isotropic resolution before phase processing.</br> (e.g. zero-padding or truncating in Fourier domain)

  • After 3D V-SHARP in MATLAB, run 'save_input_data_for_QSMnet.m</br>'. 'test_input{sub_num}.mat' and 'test_mask{sub_num}.mat' files will be saved in 'Data/Test/Input/'.

    save_input_data_for_QSMnet(TissuePhase, Mask, TE, B0, sub_num)
    % TissuePhase : Results of 3D V-SHARP
    % Mask : Results of 3D V-SHARP
    % TE (s)
    % B0 (T)
    % sub_num : subject number
    % Convert unit from Hz to ppm : field / (Sum(TE) * B0 * gyro) [ppm]
  

  • Save data with the same orientation and polarity as val_input.mat, val_mask.mat, and val_label.mat files in 'Data/Train/' folder. <img src="https://user-images.githubusercontent.com/29892433/64081330-5f2b9600-cd3a-11e9-9ff2-20e1e0ef2996.jpg" width="50%" height="50%">

Training process

• Activate qsmnet conda environment

  conda activate qsmnet 
  

• Usage

  • Before training, local field & susceptibility maps need to be dividied into 64 x 64 x 64 in Matlab

  python training_data_patch.py
  # PS : Patch size
  # net_name : Network name
  # sub_num : Number of subject to train
  # dir_num : Number of direction per subject
  # patch_num : Number of patches in [x, y, z]
  

  • Training process in python

  python train.py
  

Inference

• Activate qsmnet conda environment

  conda activate qsmnet 
  

• Usage

  python inference.py
  

  • 'subject#<network_name>-epochs.mat' & 'subject#<network_name>-epochs.nii' will be saved after QSMnet reconstruction.