SAGAN

SAGAN - Sharpness-aware Low dose CT denoising using Conditional generative adversarial network

Our goal is to reproduce SAGAN. However, only the pix2pix learning concept is applied.

1. Prerequistites

1.1 NVIDIA GPU

1.2 Python 3.x

2. Getting Started

2.1 Install Python3.x (recommend using Anaconda)

2.2 Install python dependencies

2.3 Clone this repo

2.4 Prepare Dataset

1) The piglet dataset we used in the SAGAN paper is now open for download in here

2) The file extension of this dataset is DICOM. Therefore, the extension should be converted to PNG. We used mritopng to convert images.

3) Dataset Diretory Hierarchy

- CT scans of a deceased piglet

--- Input(Dose + Recon) : 5% DOSE ASIR x820

--- Label(Dose + Recon) : 100% DOSE ASIR x820

*Validation dataset randomly selected 123 images from the train dataset and excluded them from the train dataset.

3. Model Architecture

3.1 Generator

3.2 Discriminator(Image)

4. Overview of Training Concept

5. How to use

5.1 Train

--output_dir (models are saved here.)

--is_test (if is_test==False, train.)

5.2 Test

--is_test (if is_test==True, test.)

--test_dir (test input files directory.)

--model_dir (saved models directory.)

--output_dir (test results directory, test results are saved here.)

--mn (model checkpoint -> all_model_checkpoint_paths index.)

--fn (folder name integer.)

5.3 Convert

--is_convert (if is_convert==True, convert.)

--test_dir (test input files directory.)

--model_dir (saved models directory.)

--output_dir (conversion results directory, conversion results are saved here.)

--mn (model checkpoint -> all_model_checkpoint_paths index.)

6. Overview of source codes

-- sa_Tensorflow_utils.py contains the layer elements that make up the network.

-- sa_convert.py outputs a generated image by inputting a single image or a folder containing many images (evaluation result is omitted)

-- sa_dataset.py configures the datasets used for train, validation, and test.

-- sa_main.py contains the flags of the source codes and executes the sa_solver.py code.

-- sa_model.py constructs and designs the model.

-- sa_solver.py designs the overall process.

-- sa_utils.py consists of utils needed in source codes.

7. Reference

7.1 SAGAN Paper

https://doi.org/10.1007/s10278-018-0056-0 , journal of Digital imaging, October 2018

https://arxiv.org/pdf/1708.06453.pdf, arxiv, October 2017

7.2 Github repository

https://github.com/xinario/SAGAN, This repo provides the trained denoising model and testing code.

[https://github.com/xinario/defocus_segmentation], This repo provides the S network algorithm hints (See lbpSharpness.py)

7.3 Acknowlegements

Pix2Pix, arXiv, November 2018

8. Experience Results

1) We used the *PSNR and *SSIM for quantitative evaluation.

2) The ratio of GAN loss to Segmentation loss was adjusted and tested.

*PSNR (peak signal to noise ratio)

*SSIM (The structural similarity)

3) If you want to change the weights of the two loss values, you can change the values ​​of lambda1 (seg loss) and lambda2 (gan loss) in the flags.

4) The three images used in the test are in test_imgs (gan _ 1 + seg _ 100) ... / input file and are the average of the three results.

5) Table.1

*The five values ​​in a cell are test results from 4000, 8000, 12000, 16000 and 20000 learning iterations from the left.

cGANs-tensorflow-Python