Parallel qMRI Reconstruction from 4x Accelerated Acquisitions

Paper

• Parallel qMRI Reconstruction from 4x Accelerated Acquisitions, Mingi Kang [arXiv]

Funding & Acknowledgements

This project was supported by the Washington University in St. Louis Undergraduate Summer Engineering Research Fellowship (WUSEF) 2025. [WUSEF Website]

Project was under the supervision of Prof. Ulugbek S. Kamilov at Washington University in St. Louis (now at University of Wisconsin-Madison) and the Computational Imaging Group (CIG) lab.

Project period: Summer 2025

Overview

Magnetic Resonance Imaging (MRI) acquisitions require extensive scan times, limiting patient throughput and increasing susceptibility to motion artifacts. This project introduces a miniature version of SPICER: self-supervised learning for MRI with automatic coil sensitivity estimation and reconstruction. The proposed model effectively reduces the parameter size by 4x of deep unfolding network utilized in SPICER while maintaining reconstruction performance. Our two-module architecture consists of Coil Sensitivity Map (CSM) estimation modeuls and a U-Net-based MRI reconstruction module.

Models

1. unet.py -

   • Implementation of the standard U-Net architecture for image segmentation tasks.

2. attention_unet.py -

   • Implementation of the Attention U-Net architecture, which incorporates attention gates to enhance feature learning.

3. deep_unfolding.py -

   • Implementation of a deep unfolding network for MRI reconstruction, combining model-based and data-driven approaches. Implementations from SPICER paper.

4. recon_unet.py -

   • Implementation of a U-Net-based MRI reconstruction model, designed for efficient and accurate image reconstruction.

5. recon_attention_unet.py -

   • Implementation of an Attention U-Net-based MRI reconstruction model, enhancing the standard U-Net with attention gates for improved performance.

Project Structure

.
├── models/
│   ├── unet.py             # U-Net architecture implementation
│   ├── attention_unet.py   # Attention U-Net architecture implementation
│   ├── deep_unfolding.py   # Deep unfolding network implementation
│   ├── recon_unet.py       # U-Net-based MRI reconstruction model
│   └── recon_attention_unet.py # Attention U-Net-based MRI reconstruction model
├── utils/
│   ├── fourier.py          # Fourier transform utilities
│   ├── loss.py             # Loss functions for training
│   ├── metric.py           # Evaluation metrics for MRI reconstruction
│   ├── operator.py         # MRI reconstruction operators
│   └── util.py             # Utility functions
├── dataset.py              # Multi-coil Echo MRI dataset
├── main.py                 # Main training script
├── test.py                 # Testing script for evaluation
├── visualize.py            # Visualization functions (save images, videos)
├── README.md               # ← you are here
└── LICENSE                 # MIT License

License

MRI-Reconstruction is released under the MIT License. Please see the LICENSE file for more information.