Fast and High-Resolution Reconstruction for Laplace NMR based on Deep Learning

This is the original Pytorch implementation of DRECT in the following paper: High-Quality Reconstruction for Laplace NMR Based on Deep Learning.

Requirements

Dependencies can be installed using the following command:

pip3 install torch==1.11.0 --extra-index-url https://download.pytorch.org/whl/cu113
pip3 install -r requirements.txt

Test.ipynb (Using MATLAB to plot the last result，the display code for the comparative results of the paper can be found at "DRECT/MATLAB/Result/Results_Graph_Paper.mlx". To facilitate user convenience and verification, an example of how to draw your results graph is available at "DRECT/MATLAB/Result/Results_Graph.m".It is worth noting that you can achieve optimal display results by adjusting the contour_level parameter or by adjusting the parameters in preprocessing for better model output)

# For T1T2
python3 main.py --Type T1T2

# For DOSY
python3 main.py --Type DOSY

# For VD
python3 main.py --Type VD

Something wrong with git-lfs, you can get the pre-trained models on Google Drive instead.

The parameters can be configured in config.py. The details about the parameters are described below.

Parameter name	Description of parameter
n_layers	Num of encoder layers
n_heads	Num of heads
d_ff	Dimension of feed forward modules
lr	Learning rate
dropout	The probability of dropout
label_size	Number of points indicating the value of diffusion coefficient
batch_size	The batch size of training input data
mul_label	Label magnification
max_epochs	Max train epochs
n_fre	Number of points of chemical shift
o_fre	Number of simulated peaks
max_fre	Maximum value of simulated chemical shift
num_D	Maximum number of simulated molecular components
min_sep	Minimum separation between spikes, normalized by signal_dim
dB	Noise level
max_b	Maximum value of parameter b
num_samples	Number of simulated samples (including training and evaluation)
max_D	Maximum value of simulated diffuision coefficient
ratio	Ratio of validation dataset in all data
sig	Width of the simulated peak of Gaussian distribution.
signal_dim	Dimension of the input signal
sig_lorz	Width of the simulated peak of Lorentz distribution.
gpu	The gpu no, used for training and test
base_D	Minimum value of diffusion coefficient
lr_scheduler	learning rate scheduler(Cosine or ReduceLR)

Raw data for VD available at DOI: 10.15127/1.303820.(https://www.escholar.manchester.ac.uk/uk-ac-man-scw:303820 ), It needs to be processed by DOSYToolbox and our preprocessing module(MATLAB/Data/VD/Solve_data.m) after downloading.

If you have any questions, please contact us through Email (33320211150312@stu.xmu.edu.cn) or Github issues. Pull requests are highly welcomed!