Official TensorFlow implementation of the unsupervised reconstruction model using zero-Shot Learned Adversarial TransformERs (SLATER). (https://arxiv.org/abs/2105.08059)
Korkmaz, Y., Dar, S. U., Yurt, M., Ozbey, M., & Cukur, T. (2021). Unsupervised MRI Reconstruction via Zero-Shot Learned Adversarial Transformers. arXiv preprint arXiv:2105.08059.
Demo
The following commands are used to train and test SLATER to reconstruct undersampled MR acquisitions from single- and multi-coil datasets. You can download pretrained network snaphots and sample datasets from the links given below.
For training the MRI prior we use fully-sampled images, for testing undersampling is performed based on selected acceleration rate. We have used AdamOptimizer in training, RMSPropOptimizer with momentum parameter 0.9 in testing/inference. In the current settings AdamOptimizer is used, you can change underlying optimizer class in dnnlib/tflib/optimizer.py file. You can insert additional paramaters like momentum to the line 87 in the optimizer.py file.
Sample training command for multi-coil (fastMRI) dataset:
python run_network.py --train --gpus=0 --expname=fastmri_t1_train --dataset=fastmri-t1 --data-dir=datasets/multi-coil-datasets/train
Sample reconstruction/test command for fastMRI dataset:
python run_recon_multi_coil.py reconstruct-complex-images --network=pretrained_snapshots/fastmri-t1/network-snapshot-001282.pkl --dataset=fastmri-t1 --acc-rate=4 --contrast=t1 --data-dir=datasets/multi-coil-datasets/test
Sample training command for single-coil (IXI) dataset:
python run_network.py --train --gpus=0 --expname=ixi_t1_train --dataset=ixi_t1 --data-dir=datasets/single-coil-datasets/train
Sample reconstruction/test command for IXI dataset:
python run_recon_single_coil.py reconstruct-magnitude-images --network=pretrained_snapshots/ixi-t1/network-snapshot-001282.pkl --dataset=ixi_t1_test --acc-rate=4 --contrast=t1 --data-dir=datasets/single-coil-datasets/test
Datasets
- IXI dataset: https://brain-development.org/ixi-dataset/
- fastMRI Brain dataset: https://fastmri.med.nyu.edu/
For IXI dataset image dimensions are 256x256. For fastMRI dataset image dimensions vary with contrasts. (T1: 256x320, T2: 288x384, FLAIR: 256x320).
SLATER requires datasets in the tfrecords format. To create tfrecords file containing new datasets you can use dataset_tool.py:
To create single-coil datasets you need to give magnitude images to dataset_tool.py with create_from_images function by just giving image directory containing images in .png format. We included undersampling masks under datasets/single-coil-datasets/test.
To create multi-coil datasets you need to provide hdf5 files containing fully sampled coil-combined complex images in a variable named 'images_fs' with shape [num_of_images,x,y] (can be modified accordingly). To do this, you can use create_from_hdf5 function in dataset_tool.py.
The MRI priors are trained on coil-combined datasets that are saved in tfrecords files with a 3-channel order of [real, imaginary, dummy]. For test purposes, we included sample coil-sensitivity maps (complex variable with 4-dimensions [x,y,num_of_image,num_of_coils] named 'coil_maps') and undersampling masks (3-dimensions [x,y, num_of_image] named 'map') in the datasets/multi-coil-datasets/test folder in hdf5 format.
Coil-sensitivity-maps are estimated using ESPIRIT (http://people.eecs.berkeley.edu/~mlustig/Software.html). Network implementations use libraries from Gansformer (https://github.com/dorarad/gansformer) and Stylegan-2 (https://github.com/NVlabs/stylegan2) repositories.
Pretrained networks
You can download pretrained network snapshots and datasets from these links. You need to place downloaded folders (datasets and pretrained_snapshots folders) under the main repo to run those sample test commands given above.
Pretrained network snapshots for IXI-T1 and fastMRI-T1 can be downloaded from Google Drive: https://drive.google.com/drive/folders/1_69T1KUeSZCpKD3G37qgDyAilWynKhEc?usp=sharing
Sample training and test datasets for IXI-T1 and fastMRI-T1 can be downloaded from Google Drive: https://drive.google.com/drive/folders/1hLC8Pv7EzAH03tpHquDUuP-lLBasQ23Z?usp=sharing
Notice for training with multi-coil datasets
To train multi-coil (complex) datasets you need to remove/add some lines in training_loop.py:
- Comment out line 8.
- Delete comment at line 9.
- Comment out line 23.
Citation
You are encouraged to modify/distribute this code. However, please acknowledge this code and cite the paper appropriately.
@article{korkmaz2021unsupervised,
title={Unsupervised MRI Reconstruction via Zero-Shot Learned Adversarial Transformers},
author={Korkmaz, Yilmaz and Dar, Salman UH and Yurt, Mahmut and {\"O}zbey, Muzaffer and {\c{C}}ukur, Tolga},
journal={arXiv preprint arXiv:2105.08059},
year={2021}
}
(c) ICON Lab 2021
Prerequisites
- Python 3.6 --
- CuDNN 10.1 --
- TensorFlow 1.14 or 1.15
Acknowledgements
This code uses libraries from the StyleGAN-2 (https://github.com/NVlabs/stylegan2) and Gansformer (https://github.com/dorarad/gansformer) repositories.
For questions/comments please send me an email: [email protected]