LAMA: Learned Alternating Minimization Algorithm for Dual-domain Sparse-View CT Reconstruction

This is the official site for the implementation of the published paper in MICCAI 2023: "Learned Alternating Minimization Algorithm for Dual-domain Sparse-View CT Reconstruction."

Here is the link for the paper

Dependencies

pytorch==1.10.0
scipy==1.11.1
numpy==1.25.2
scikit-image==0.19.3
opencv-python==4.8.0
pytorch_msssim==1.0.0
CTLIB

Computational Tomography Library: CTLIB

Pytorch version from 1.10 to 1.13 are all tested and okay to use. But PyTorch 2.0 and above are not working with the CTLIB library.

Dataset

The two dataset used for experiments in the paper are provided in this OneDrive. To download this for Linux, you can use the following commands inside the LAMA-Learned-Alternating-Minimization-Algorithm/ directory:

wget -O dataset.zip "https://uflorida-my.sharepoint.com/:u:/g/personal/ding_chi_ufl_edu/ETaqu5-bebZOjr0Bws3VxU4BkOVqJcwK_M0AvvcCzsqfgQ?e=umy9G0&download=1"

To unzip and process the data, you can use the following commands:

unzip dataset.zip
python process-data.py

To process data on specific dataset, number of views, type of initialization network, train or test, you can use the following command as an example:

python process-data.py --dataset=mayo --n_views=64 --network=CNN --train=True

This will generate the following for the mayo clinic dataset: 64-view sinogram, FBP images, and the initialized sinogram and images after a simple CNN initialization network as the training data for LAMA. The data will be saved in separate folders in the directory dataset/mayo/train/ for the example.

CTLIB

CTLIB-demo.ipynb is a demo to help the readers to familiarize the usage of the CTLIB, and the application of sparse-view CT.

Training

Before training, you need to process the data first by running the following command as an example:

python process-data.py --dataset=mayo --n_views=64 --network=CNN --train=True

The training data will be saved in the directory dataset/mayo/train/ for the example. Then you can use the following command to train the network:

python demo-train.py --dataset=mayo --n_views=64

There are more hyperparameters that you can set for training. Please refer to the demo-train.py for more details. But if you change the dataset, number of views, or the type of initialization network, etc. you need to process the data again by changing the corresponding parameters. For example, if you want to train the network on the NBIA dataset with 128 views and the initialization network is a simple CNN, you can use the following command:

python process-data.py --dataset=NBIA --n_views=128 --network=CNN --train=True

So far we only support CNN as the initialization network, LAMA baseline as the main backbone. We will add more initialization networks and backbones in the future. More specifically, you can train LAMA without using any initializaton network by setting --img_dir FBP_64views and --prj_dir 64views.

Testing

Similarly, process data first by running the following command as an example:

python process-data.py --dataset=mayo --n_views=64 --network=CNN --train=False

Then you can use the following command to test the network:

python demo-test.py --dataset=mayo --n_views=64

Model

LAMA in one iteration:

Network unrolling:

Sparse-View CT Reconstruction

Mayo clinic LDCT grand challenge:

NBIA dataset:

Experiments

Mayo:

Metrics	Views	FBP	DDNet	LDA	DuDoTrans	Learn++	LAMA(ours)
PSNR	64 128	27.17 ± 1.11 33.28 ± 0.85	35.70 ± 1.50 42.73 ± 1.08	37.16 ± 1.33 43.00 ± 0.91	37.90 ± 1.44 43.48 ± 1.04	43.02 ± 2.08 49.77 ± 0.96	44.58 ± 1.15 50.01 ± 0.69
SSIM	64 128	0.596 ± 9e−4 0.759 ± 1e−3	0.923 ± 4e−5 0.974 ± 4e−5	0.932 ± 1e−4 0.976 ± 2e−5	0.952 ± 1.0e−4 0.985 ± 1e−5	0.980 ± 3e−5 0.995 ± 1e−6	0.986 ± 7e−6 0.995 ± 6e−7

NBIA:

Metrics	Views	FBP	DDNet	LDA	DuDoTrans	Learn++	LAMA(ours)
PSNR	64 128	25.72 ± 1.93 31.86 ± 1.27	35.59 ± 2.76 40.23 ± 1.98	34.31 ± 2.20 40.26 ± 2.57	35.53 ± 2.63 40.67 ± 2.84	38.53 ± 3.41 43.35 ± 4.02	41.40 ± 3.54 45.20 ± 4.23
SSIM	64 128	0.592 ± 2e−3 0.743 ± 2e−3	0.920 ± 3e−4 0.961 ± 1e−4	0.896 ± 4e−4 0.963 ± 1e−4	0.938 ± 2e−4 0.976 ± 6e−5	0.956 ± 2e−4 0.983 ± 5e−5	0.976 ± 8e−5 0.988 ± 3e−5

Citation

If you find this implementation useful, please consider citing our work:

@InProceedings{ding2023LAMA,
  title={Learned Alternating Minimization Algorithm for Dual-Domain Sparse-View CT Reconstruction},
  author={Ding, Chi and Zhang, Qingchao and Wang, Ge and Ye, Xiaojing and Chen, Yunmei},
  booktitle={Medical Image Computing and Computer Assisted Intervention -- MICCAI 2023},
  year={2023},
  publisher={Springer Nature Switzerland},
  pages={173--183},
  isbn={978-3-031-43999-5"}
}