Physics-Aware Motion Simulation for T2*-Weighted Brain MRI

Hannah Eichhorn, Kerstin Hammernik, Veronika Spieker, Samira M. Epp, Daniel Rueckert, Christine Preibisch, Julia A. Schnabel

Accepted at MICCAI 2023 SASHIMI workshop | Link to paper.

Abstract: In this work, we propose a realistic, physics-aware motion simulation procedure for T2*-weighted magnetic resonance imaging (MRI) to improve learning-based motion correction. As T2*-weighted MRI is highly sensitive to motion-related changes in magnetic field inhomogeneities, it is of utmost importance to include physics information in the simulation. Additionally, current motion simulations often only assume simplified motion patterns. Our simulations, on the other hand, include real recorded subject motion and realistic effects of motion-induced magnetic field inhomogeneity changes. We demonstrate the use of such simulated data by training a convolutional neural network to detect the presence of motion in affected k-space lines. The network accurately detects motion-affected k-space lines for simulated displacements down to ≥ 0.5mm (accuracy on test set: 92.5%). Finally, our results demonstrate exciting opportunities of simulation-based k-space line detection combined with more powerful reconstruction methods.

Citation

If you use this code, please cite our paper:

@InProceedings{eichhorn2023deep,
      title={Physics-Aware Motion Simulation for {T2*}-Weighted Brain {MRI}}, 
      author={Hannah Eichhorn and Kerstin Hammernik and Veronika Spieker and Samira M. Epp and Daniel Rueckert and Christine Preibisch and Julia A. Schnabel},
      booktitle="Simulation and Synthesis in Medical Imaging. SASHIMI 2023. Lecture Notes in Computer Science",
      year={2023},
      publisher={Springer International Publishing}
}

Contents of this repository:

• motion_simulation: simulating realistic motion artefacts in T2*w GRE MRI data
• line_det_network: training and testing a k-space line detection network, using the IML-CompAI Framework
• evaluation: evaluating the proposed method, using the medutils package for the TV reconstruction

All computations were performed using Python 3.8.12 and PyTorch 1.13.0.

Setup:

1. Create a virtual environment with the required packages:

   cd ${TARGET_DIR}/T2starLineDet
   conda env create -f conda_environment.yaml
   source activate t2star_linedet *or* conda activate t2star_linedet
   

2. Install pytorch with cuda:

   conda install pytorch torchvision torchaudio pytorch-cuda=11.7 -c pytorch -c nvidia
   pip install torchinfo
   conda install -c conda-forge pytorch-lightning
   

3. For setting up wandb please refer to the IML-CompAI Framework.

Steps to reproduce the analysis:

1. Motion Simulation:

   i) Run motion_simulation/PCA_motion_curves_CV.py to perform a prinicpal component analysis on the motion data

   ii) Run motion_simulation/Scan_order.py to extract the acquisition order of all k-space lines from a raw file (ISMRMRD format)

   iii) Run motion_simulation/Prepare_config.py to generate configuration files needed for running the motion simulation

   iv) Run motion_simulation/Simulate_Motion_Whole_Dataset.py to run the motion simulations

2. Line Detection Network:

   Follow the instructions in line_det_network/README.md

3. Final Evaluations:

   i) Run evaluation/EvaluatePredictions.py to evaluate the network performance and reconstruct example motion corrected images

Illustration of the motion simulation procedure:

Illustration of the network architecture:

Results

Classification performance is decreasing for decreasing simulation thresholds:

Weighted reconstructions show subtly redued artefacts:

For a more detailed description of the results please refer to our preprint.