Contact: marc.vornehm@fau.de, rizwan.ahmad@osumc.edu
Please cite the original publication as follows:
Vornehm M, Chen C, Sultan MA, et al. Multi-dynamic deep image prior for cardiac MRI. Magn Reson Med. 2025; 94(6): 2668-2679. doi: 10.1002/mrm.70000
Additionally, a preprint is available on arXiv.
Multi-Dynamic Deep Image Prior is a novel unsupervised reconstruction framework for accelerated realtime cardiac MRI. M-DIP employs a spatial dictionary to synthesize a time-dependent template image, which is further refined using time-dependent deformation fields that model cardiac and respiratory motion.
M-DIP is tested using Python 3.12 on Linux. Install the required packages using
pip install -r requirements.txt
To run LR-DIP for comparison, create a separate environment based on Python 3.12 and install the respective packages using
pip install -r requirements_lrdip.txt
Raw data should be placed in a subdirectory data/.
The in-vivo data used in our publication cannot be published.
Instead, we recommend using real-time cine data from the OCMR dataset, ideally with at least 100 frames.
- M-DIP: in
M-DIP.ipynb - Low rank plus sparse (Otazo et al., https://doi.org/10.1002/mrm.25240): in
M-DIP.ipynb, sectionL+S reconstruction. - LR-DIP (Hamilton et al., https://doi.org/10.1007/s10334-023-01088-w): in
LR-DIP.ipynb, see more details below.
All parameters and paths can be changed via the variables defined in the second cell. Specify the folder containing raw data in MRD format in raw_folder, the desired output folder in out_folder, and the name of the raw data file in filename.
The Jupyter Notebook can be executed from the command line using the papermill package (included in requirements.txt).
You can furthermore overwrite any of the parameters defined in the second cell using the -p switch, where each parameter has to be preceeded by the switch:
papermill M-DIP.ipynb temp.ipynb [-p <parameter> <value> ...]
Example:
papermill M-DIP.ipynb temp.ipynb -p n_bases 15 -p filename <filename>
Note that temp.ipynb can be any arbitrary Jupyter Notebook filename.
Code for the LR-DIP reconstruction was kindly provided by Jesse Hamilton (University of Michigan) and was adjusted for Cartesian data and for parameterization using papermill. Note that the code does not follow the descriptions from the paper in every aspect. However, the provided code gave the best results.
Similar to the procedure described for M-DIP above, the notebook can be executed via the command line using
papermill LR-DIP.ipynb temp.ipynb [-p <parameter> <value> ...]