Ring Artifact Correction in Photon-Counting Spectral CT Using a Convolutional Neural Network With Spectral Loss

This repo contains code acompanying Ring Artifact Correction in Photon-Counting Spectral CT Using a Convolutional Neural Network With Spectral Loss.

Run the code

Dependencies

Install necessary packages via

pip install -r requirements.txt

In addtion, you need to download pytorch_ssim and put it in your directory.

Usage

The two main scripts are train and evaluate. To train run

python train.py --FLAGS 

To see available flags run python train.py -h. Similarly, to evaluate our network we run

python evaluate.py --FLAGS 

Example

For instance, to train the top performing network, we ran

python train.py --net unet_alt --loss_fn vgg16_l1_alt --layer 9 --lambda_1 10 --lambda_2 1 --skip_connection --batch_sz 2 --init_features 64 --patch_sz 512 --train ./data/train_kits_img --valid ./data/val_kits_img --epochs 100 --n_samples 2 --log_interval 25

This model is then saved as unet_alt_64_vgg16_l1_alt_9_100_2_sc_512_10.0_1.0_2_train_kits_img in .\results. To evaluate this network run

python evaluate.py --net unet_alt_64_vgg16_l1_alt_9_100_2_sc_512_10.0_1.0_2_train_kits_img --loss_fn vgg16_alt --data ./data/test_kits_img --idx_print 34 --n_samples 1

Contact

Dennis Hein
[email protected]

Acknowledgements

The following sources were helpful for this project:

• Pytorch_ssim
• Implementation of UNet