RASER AI Code Repository

Official code for the publication:

Deep learning corrects artifacts in RASER MRI profiles
Moritz Becker, Filip Arvidsson, Jonas Bertilson, Elene Aslanikashvili, Jan G. Korvink, Mazin Jouda, Sören Lehmkuhl
Magnetic Resonance Imaging, Volume 115, January 2025, 110247
https://doi.org/10.1016/j.mri.2024.110247

---

Table of Contents

• Project Overview
• Project Structure
• Dataset
• Installation & Dependencies
• Usage
  • Configuration
  • Data Preparation
  • Model Training & Evaluation
  • Image Reconstruction & Inference
• Reference
• Citation
• Contact

---

Project Overview

This repository provides the code for a deep learning pipeline that corrects severe artifacts in RASER MRI images using a two-step approach:

• 1D Correction: CNN for 1D RASER projections (nonlinear distortion correction)
• 2D Enhancement: U-net for 2D image enhancement

Models are trained on synthetic data but generalize to experimental data, enabling artifact correction in otherwise unusable RASER MRI images.

Project Structure

raser_mri_ai/
  AI.py                  # Model training & evaluation (1D/2D)
  image_AI.py            # Image reconstruction & inference
  image_data_processing.py # Data preprocessing
  NN_architectures.py    # PyTorch model definitions
  utility_functions.py   # Helper functions
  models/
    config_models.py     # Pydantic experiment/model configs
LICENSE
pyproject.toml
README.md

Dataset

The dataset (synthetic & experimental RASER MRI data) is openly available:

• https://publikationen.bibliothek.kit.edu/1000168053

See the dataset repository for structure and terms.

Installation

Install dependencies (using Poetry):

poetry install

Or with pip:

pip install .

Usage

1. Configuration

All main scripts use a Pydantic config object (RaserConfig), with an optional cnn field for CNN hyperparameters (kernel size, filters, activation, etc.). Adjust the initial_config variable in each script to control experiment setup, model, and training options.

2. Data Preparation

• Script: raser_mri_ai/image_data_processing.py
• Purpose: Preprocess raw/simulated data into NumPy arrays for training/testing.
• How to use:
  1. Edit initial_config in image_data_processing.py for your dataset/system.
  2. Run:

     python -m raser_mri_ai.image_data_processing

  3. Processed .npy files are saved to the output directory.

3. Model Training & Evaluation

• Script: raser_mri_ai/AI.py
• Purpose: Train 1D/2D neural networks and evaluate performance.
• How to use:
  1. Edit initial_config in AI.py for model/data/training options.
  2. Run:

     python -m raser_mri_ai.AI

  3. Results and trained weights are saved in the output directory.

4. Image Reconstruction & Inference

• Script: raser_mri_ai/image_AI.py
• Purpose: Apply trained models for image correction and evaluation.
• How to use:
  1. Edit initial_config in image_AI.py for model/input/output settings.
  2. Run:

     python -m raser_mri_ai.image_AI

  3. Outputs reconstructed images and evaluation results.

---

Reference

See the paper for detailed methods, datasets, and results:

Deep learning corrects artifacts in RASER MRI profiles
Moritz Becker, Filip Arvidsson, Jonas Bertilson, Elene Aslanikashvili, Jan G. Korvink, Mazin Jouda, Sören Lehmkuhl
Magnetic Resonance Imaging, Volume 115, January 2025, 110247
https://doi.org/10.1016/j.mri.2024.110247

Citation

If you use this code or data, please cite:

Becker, M., Arvidsson, F., Bertilson, J., Aslanikashvili, E., Korvink, J. G., Jouda, M., & Lehmkuhl, S. (2025). Deep learning corrects artifacts in RASER MRI profiles. Magnetic Resonance Imaging, 115, 110247. https://doi.org/10.1016/j.mri.2024.110247

Contact

For questions or collaborations, contact the corresponding author as listed in the paper.