This repository contains the PyTorch implementation of CT super-resolution using multiple dense residual block based GAN presented in [Zhang et al., 2020], with slight changes as recommended in [Gulrajani et al., 2017] and with architecture parameters adapted from [Ledig et al., 2017], in case they were not specified.

The data used represent a subset of MosMed Covid-19 dataset. It consist in multiple chest CT scans of healthy and infected patients in .nii.gz 3D-format, with different number of scan slices and slices thickness.
The 3-dimensional CTs are splitted in slices and each slice is saved in .npy 2D-format.

The data directory should have the following structure:

.
├── Data
    ├── Test
    │   ├── *.npy
    ├── Train
    │   ├── *.npy
    ├── Validation
    │   ├── *.npy

This will automatically start a Tensorboard session, with the following parameters:

--tb_port (specify the port number to be used by Tensorboard) 
--tb_logdir = 'runs' (directory to write logs to Tensorboard)  
--tb_plot_interval = 10 (interval for creating and writing generated CT images | the losses and PSNR scores are logged every <--log_interval> epochs)

Control the training by modifying the default values for the following parameters:

--device = cuda (train on cuda)  
--log_interval = 1 (print train & validation loss each epoch)
--checkpoint_interval = 100 (save trained model and optimizer parameters every 100 epochs)
--num_epochs = 500

Resume training by specifying a valid value for --restore_epoch.

The model saved as Models/checkpoint_<restore_epoch>.pt will be loaded

Test the model saved at training epoch --restore_epoch on CT images specified.

--test_ct_names accepts a list of the CT images without the .npy extension.

for example, the CT image located at Data/Test/ct_sample1.npy will be passed as ct_sample1.

The model saved as Models/checkpoint_<restore_epoch>.pt will be loaded.

Add pre-processing code for CT scans *.nii.gz --> *.npy

PyTorch, Tensorboard