This code implements a variational network for image reconstruction as described by Sriram A et al in End-to-End Variational Networks for Accelerated MRI Reconstruction.
This app takes as undersampled raw k-space data as input and outputs VarNet reconstructed images. Auto-calibration signal (ACS) data from a fully-sampled central k-space with integrated reference lines is required. Multi-slice data are supported.
A test data set is provided in the app container at /test_data/t1_tse_R6.mrd, a T1-weighted TSE sequence with R=6 uniform acceleration and 26 in-place auto-calibration signal (ACS) lines acquired on 18 slices in a phantom at 1.5T. It contains two datasets:
dataset_1, an adjustment scan containing noise calibration datadataset_2, k-space data from the TSE sequence
The varnet config is used for this workflow. The acceleration rate and size of the ACS region can be retrospectively downsampled by appending _Rx or _ACSy where x is the acceleration rate and y is the number of ACS lines. For example, a config of varnet_R8_ACS24 indicates retrospective downsampling to rate 8 acceleration with 24 ACS lines.
This MRD app can be downloaded from Docker Hub at https://hub.docker.com/r/kspacekelvin/varnet-mrd-app. In a command prompt on a system with Docker installed, download the Docker image:
docker pull kspacekelvin/varnet-mrd-app
Start the Docker image with the container name varnet-app, a local folder named ~/data mounted inside the container at /data, and port 9002 shared:
docker run --rm --name varnet-app -v ~/data:/data -p 9002:9002 kspacekelvin/varnet-mrd-app
In another window, use an MRD client such as the one from the python-ismrmrd-server. A copy is included in the app.
Start another terminal inside the Docker started above:
docker exec -it varnet-app bash
Run the client and send the data to the server:
python3 /opt/code/python-ismrmrd-server/client.py -c varnet -o /data/t1_tse_R6_varnet.mrd -g dataset_2 /test_data/t1_tse_R6.mrd
The output file (e.g. t1_tse_R6_varnet.mrd) contains VarNet reconstructed images stored in the ~/data folder. Reconstructed images can be viewed with tools such as the ismrmrdviewer or by converting to GIF using mrd2gif, which is included in this app:
python3 /opt/code/python-ismrmrd-server/mrd2gif.py /data/t1_tse_R6_varnet.mrd
An animated .GIF file is created in the same folder:
The phantom data can also be reconstructed using standard GRAPPA using the Gadgetron which can also be used as an MRD App. Download the Gadgetron Docker image:
docker pull ghcr.io/gadgetron/gadgetron/gadgetron_ubuntu_rt_nocuda
In another terminal, start the Gadgetron server mapping port 9003 on the host to port 9002 inside the container:
docker run --rm -p 9003:9002 ghcr.io/gadgetron/gadgetron/gadgetron_ubuntu_rt_nocuda
In the window above where the client was previous run, re-run the client and send the data to Gadgetron. Note that both the noise and actual data must be sent in order to do SNR scaled reconstruction.
python3 /opt/code/python-ismrmrd-server/client.py -a host.docker.internal -p 9003 -c default_measurement_dependencies.xml -g dataset_1 /test_data/t1_tse_R6.mrd
python3 /opt/code/python-ismrmrd-server/client.py -a host.docker.internal -p 9003 -c Generic_Cartesian_Grappa_SNR.xml -o /data/t1_tse_R6_gadgetron.mrd -g dataset_2 /test_data/t1_tse_R6.mrd
Create a .GIF of the GRAPPA reconstructed images:
python3 /opt/code/python-ismrmrd-server/mrd2gif.py /data/t1_tse_R6_gadgetron.mrd
When processing is complete, the Docker container can be stopped by running:
docker kill varnet-app