SART-GAN-for-CT-reconstruction

Readme

This project is from Ruiwen Xing’s master thesis: Deep Learning Based CT Image Reconstruction.

The project contains following code file: Main.py the start point Sinogram.py create sinogram from tomography image Reconstruction.py reconstruction process by SART algorithm imgPainter.py output image file imgFormatConvert.py adjust shape of image tensor imgEvaluation.py calculate PSNR & SSIM value between two image NNmodel.py create different neural net model NNstructure.py inner structures of neural networks NNstructure_xxxx.py inner structures of other neural networks Parameters.py handle hyper parameters SAGAN_ops.py support block from original SAGAN SAGAN_util.py support block from original SAGAN Debug.py some functions useful in debug process createTrainset.py to generate noisy or incomplete CT images from clean, high dose CT images dataLoader.py load data imgTrainset.py create paired image trainset

Prepare Dataset

In this project we use CT images from cancer image archive. CT images are created under standard dose scan. However, this project needs paired image dataset So, we use ASTRA toolbox to simulate different CT scan and generate low-dose, sparse view, and limited angle sinogram (original scan data) And then, we use standard SART algorithm to reconstruct these sinogram to CT image. Because of different setting during simulated scan (low-dose, sparse view, and limited angle), we loss some data in these processes. And when we reconstruct them, we will get noisy or incomplete CT images. These noisy or incomplete images will be paired with original normal-dose clean CT images to train our neural network.

$ python createTrainset.py --func [normal or full: full contains data augmentation process] --inFile [input folder path] --outFile [output folder path] --dataType [png or flt] --maxImg [maximum generated image number] --option [sinogram generate option] --iterNum [number of iterations in SART reconstruction] --ns [subset number in SART algorithm] --handelNum [number of images to handle at each time] --startFrom [number of image to start from input folder]

• data augmentation by rotate CT image
• for --option, currently we support "low-dose_1e4", "low-dose_1e5", "low-dose_1e6", "low-dose_2e5", "sparse_view_450", "sparse_view_180", "sparse_view_100", "sparse_view_60", "sparse_view_50", "sparse_view_40", "limited_angle_160", "limited_angle_140", "limited_angle_120", "limited_angle_100" detail of these options can be seen in parameters.py

For example:

$ python createTrainset.py
--func full
--inFile “../NDCT”
--outFile “../uiharuDataset/limitedAngle120ns50it20”
--dataType png
--option limited_angle_140
--iterNum 20
--ns 50
--handelNum 10
--startFrom 0

Prepare environment

In this project we use anaconda to create our python environment. Download anaconda installer from: https://www.anaconda.com/download/#linux To activate anaconda:

$ conda activate To create anaconda environment abc: $ conda create --name abc To activate anaconda environment abc: $ conda activate abc (abc)>$ To install numpy package in conda environment abc: (abc)>$ conda install numpy To clone an environment: $ conda create –name myclone –clone myenv To create an environment from yml file $ conda env create -f environment.yml To remove an environment: $ conda remove --name myenv --all To deactivate conda $ conda activate

See https://docs.conda.io and https://docs.anaconda.com/anaconda/user-guide/getting-started/ for more information.

Before run our code, the following python packages need to be installed: Astra-toolbox v 1.8.3 Cudatoolkit v 8.0 CUDNN v 7.1.3 Keras v 2.2.4 Matplotlib v 3.1.1 Numpy v 1.14.2 Pillow v 6.0.0 Tensorflow v 1.10.0

To install astra toolbox, see http://www.astra-toolbox.com/

We also provide a uiharu-k.yml environment for you.

Training process

For environment preparation, please see prepare environment.txt Some python packages need to be installed We also provide an environment in uiharu-k.yml

To start a training process, type:

$ python main.py --function trainNN
--cleanTrainset [trainset clean img path]
--cleanTrainsetDataType [png or flt]
--noisyTrainset [trainset noisy img path]
--noisyTrainsetDataType [png or flt]
--cleanTestset [validation set clean img path]
--cleanTestsetDataType [png or flt]
--noisyTestset [validation set noisy img path]
--noisyTestsetDataType [png or flt]
--checkpointFolder [path to save checkpoint]
--batchSize [batch size]
--NNtype [neural net structure in NNmodel.py]

Default hyperparameters are saved in parameters.py Neural networks type can be seen in NNmodel.py

For example:

$ python main.py --function trainNN
--cleanTrainset ../NDCT
--cleanTrainsetDataType png
--noisyTrainset “../uiharuDataset/sparseView100ns50iter20;../uiharuDataset/low-dose1e5ns50iter20;../uiharuDataset/limitedAngle140ns50iter20”
--noisyTrainsetDataType png
--cleanTestset “../NDCTval”
--cleanTestsetDataType png
--noisyTestset “../NDCTval”
--noisyTestsetDataType png
--checkpointFolder “./checkpoint”
--batchSize 1
--NNtype NN721

Test process

To test the model, type:

$ python main.py --function autoRecon
--inputFolder [path of ground truth CT images]
--dataType [png or flt]
--sinoFolder [path to save sinogram]
--noiseOption [scenarios, e.g. limited_angle_120]
--mnnOrder [order to reconstruct image, e.g. ‘sart5|(gan)./checkpoint|sart20|return’]
--outputFolder [path to save reconstructed image]
--ns [subset number of SART]

• --mnnOrder In this project we designed a simple command system to describe reconstruction process This command system is defined by reconstruction.py Here’s the rule: This system contains three kinds of command, commands are connected by pipeline “|” So a full mnnOrder should be: “command_1|command_2|command_3|command_4…….|command_n”

The first kind of command is sart command, it includes: "sart" Several iterations of the SART algorithm. “sart15ns20” means 15 iterations of SART with subsetNumber=20. "TVsart" Several iterations of the SART algorithm and optimize the result by reducing total variation. “TVsart15ns20” means 15 iterations of SART with subsetNumber=20, and optimization after each iteration.

The second kind of command is neural net command, it is formed by a neural net type name and a path that save the checkpoint of a trained neural net. It looks like: (cycleGAN)./checkpoint/cycleGAN

The third kind of command is return command. It simply returns the result.

Thus, a full mnnOrder could be: “sart30|(simpleGAN)../models/simpleGAN|sart20|(NN721)../models/nn721|sart20|(NN721)../models/nn721|sart20|(NN721)../models/nn721|return”

For example:

$ python main.py --function autoRecon
--inputFolder ../NDCTtest
--dataType png
--sinoFolder ./sinogram
--noiseOption limited_angle_120
--mnnOrder “sart30|(NN721)../models/simpleGAN|sart20|(NN721)../models/nn721|sart20|(NN721)../models/nn721|sart20|(NN721)../models/nn721|return”
--outputFolder ./result
--ns 20