hellopipu / hqs-net

[MIDL2022] Learned Half-Quadratic Splitting Network for Magnetic Resonance Image Reconstruction

HQS-Net

pytorch implementation of the paper Learned Half-Quadratic Splitting Network for Magnetic Resonance Image Reconstruction (https://openreview.net/pdf?id=h7rXUbALijU)

Install

python>=3.7.11 is required with all requirements.txt installed including pytorch>=1.10.0

git clone https://github.com/hellopipu/HQS-Net.git
cd HQS-Net
pip install -r requirements.txt

Prepare dataset

you can find more information about OCMR dataset at https://ocmr.info/

## download dataset
wget -nc https://ocmr.s3.amazonaws.com/data/ocmr_cine.tar.gz -P data/
## download dataset attributes csv file
wget -nc https://raw.githubusercontent.com/MRIOSU/OCMR/master/ocmr_data_attributes.csv -P data/
## untar dataset 
tar -xzvf data/ocmr_cine.tar.gz -C data/
## preprocess and split dataset, it takes several hours
python preprocess_ocmr.py

Or you can directly download the preprocessed dataset here, and then put them to data/ folder

Training

Training and testing Scripts for all experiments in the paper can be found in folder run_sh. For example, if you want to train HQS-Net on accleration factor of 5x, you can run:

sh run_sh/acc_5/train/train_hqs_5.sh

or if you want to train Unet based HQS-Net on accleration factors 10x, you can run:

sh run_sh/acc_10/train/train_hqs_unet_10.sh

Testing

For example, if you want to test HQS-Net on accleration factor of 5x, you can run:

sh run_sh/acc_5/test/test_hqs_5.sh

All pretrained models in the paper can be downlowned here, then you should put them to weight/ folder.

We also provide an Colab demo

.

Tensorboard

tensorboard for checking the curves while training

tensorboard --logdir log

hqs-net's Issues

question in formula (6)

Hello, may I ask how the second half of formula (6) is derived in the paper? How did you get rid of the inverse?

DC not used in DCCNN

Hi! I went through your code and I found that the implementation of DCCNN is not using data consistency layer.

I think maybe in the forward you should add this:

dcs.append(DCLayer(norm='ortho'))

...


def forward(self, x, k, m):
    for i in range(self.n_iter):
        x_cnn = self.rec_blocks[i](x)
        x += x_cnn
        x = self.update_operation(x, k, m)
        x = self.dcs[i](x, k, m)