Python scripts for Dynamic Mode Decomposition (DMD) based low-latency online reconstruction and self-gating routines for real-time 2D cardiac MR imaging.

Reference:

• E. Yagiz et al. "Dynamic Mode Decomposition enables low-latency high temporal resolution reconstruction for golden-angle spiral real-time MRI". Int. Society of Mag Res in Med (ISMRM) 2024. Singapore.
• E Yagiz et al. "Dynamic Mode Decomposition (DMD) Cardiac Phase Estimation for adult and fetal real-time MRI". Int. Society of Mag Res in Med (ISMRM) 2024. Singapore.

Installation

Clone repo

git clone https://github.com/usc-mrel/DMD-RT.git

Create a conda environment and activate it.

conda create env -f env.yml
conda activate pydmd_env

[Optional] For visualization, you can install pyArrView from repo. Follow the instructions there.

Example Data

2 example datasets are shared here. Download these files into the DMD-RT folder.

• Adult (example_gridded_adult.h5 and example_ref_adult.h5)
• Fetal (example_gridded_fetal.h5 and example_ref_fetal.h5)

Gridded data:

• Data collected at a prototype 0.55 T MRI scanner at USC.
• Acquisition for both datasets is GA spiral. The details can be found in the reference.
• Per frame, 8 arms are gridded (Non-uniform Fourier Transform applied) to obtain an undersampled image series. Again, the details can be seen in the reference.
• Dimensions: [Nx, Ny, Nf] number of spatial pixels in x-, y- directions and number of frames

Reference data:

• A spatiotemporally constrained reconstruction (STCR) is applied.

Run

The relevant parameters for the scripts are read from the config.toml.

Low Latency

The parameters are listed in config.toml under lowlatency. Parameters are as follows:

• data_path = '' -> Absolute path to the data.
• in_filename = 'example_gridded_adult.h5' -> Filename of the data.
• out_filename = 'out_lowlatency.h5' -> Output h5.
• ref_filename = 'example_ref_adult.h5' -> Filename of the reference data
• f_th = 2 -> Threshold Frequency [Hz] (3 Hz for fetal, 2 Hz for adult)
• res_flag = false -> Residual Flag (True/False), (True for fetal)
• alpha = 0.3 -> Scalar for the residual part if res_flag. (can be tuned)
• gpu = 0 -> device to use. -1 => CPU, 0, ... => GPU index
• Nbuff = 20 -> Number of frames in the buffer. 20 is optimal for spiral GA

To run the low-latency script:

python lowlatency.py

This will use the GPU if the flag is set to >= 0 and the GPU index is valid. -1 will use CPU. The output will be saved as an h5 file with the given name.

If the pyArrView is installed, then the original input (gridded image series), low-latency reconstruction, and a reference reconstruction are displayed side by side.

Cardiac self-gating

The parameters are listed in config.toml under selfgating. Parameters are as follows:

• data_path = '' -> Absolute path to the data.
• in_filename = 'example_ref_fetal.h5' -> Filename of the input data
• frange = [1, 10] -> Frequency range for self-gating lower-upper. List of one is for upper th [Hz]
• Nbuff = 200 -> Number of frames to process in the buffer. If > number of frames, set to number of frames
• gpu = 0 -> device to use. -1 => CPU, 0, ... => GPU index

To run the self-gating script:

python selfgating.py

Outputs are estimated waveform and end-diastole, end-systole frames from the input data. The thresholds are not fully optimized, and the script hasn't been tested for undersampled fetal data.

Notes

• The code is tested on a server with an NVIDIA A100 GPU (40GB memory)—the required memory scales with the input image series. The shared example data is around 2GB.
• You are encouraged to modify/distribute this code. However, please acknowledge this code, cite the paper appropriately.
• This work is for research purposes and is a work in progress, so use it at your own risk.