RT-SMS-ML

Code for training a deep artifact suppression reconstruction method for SMS bSSFP imaging, as described in "Rapid online deep artifact suppression for real-time spiral bSSFP CMR with blipped-CAIPI simultaneous multi-slice imaging at 1.5 T".

---

Installation

Note: The code is currently only compatible with Linux, and requires Visual Studio Code and Docker.

To securely install and run the code in this repository:

1. Clone the Github repository

2. Navigate to the project folder in Visual Studio Code

3. "Reopen the folder in container" to run the devcontainer

Training a model

To train a model on a subset of the the OCMR dataset (www.ocmr.info):

1. Download a subset of multi-slice raw k-space data from the OCMR dataset (www.ocmr.info). In the terminal:

    bash download_OCMR.sh

2. Prepare the OCMR data by writing it as .npz-files by running:

    python prepare_data.py

3. Run the training script to train the model. In the terminal:

    python training.py

4. Navigate to the folder "test_set_images" to see the performance of your model on the test set, as gif:s showing the input image with artifacts, the output image from the network, and the ground truth image, stacked from top to bottom.

Inference

To run inference on a prospectively acquired dataset (prospective_example.h5, containing one SMS-2 slice) from the real-time SMS-2 sequence using your trained model and a pretrained model:

1. Run the inference script through

    python run_inference.py -m NAME_OF_YOUR_TRAINED_MODEL

For example, if you trained model in the folder "training" has the name "my_model_20260110_083618", run:

'python run_inference.py -m "my_model_20260110_083618"'

2. Find the reconstructed images from your model and from the pretrained model in the folder "reconstructed_images". The results are saved as mp4:s, where the input image is shown on top and the output image on bottom, as shown here:

prospective_example-ML-5.mp4