Related Experiment Video
Updated: Jul 9, 2025
08:51
Magnetic Resonance Imaging of Multiple Sclerosis at 7.0 Tesla
Published on: February 19, 2021
9.0K
McSTRA: A multi-branch cascaded swin transformer for point spread function-guided robust MRI reconstruction
Mevan Ekanayake1, Kamlesh Pawar2, Mehrtash Harandi3
1Monash Biomedical Imaging, Monash University, Australia; Department of Electrical and Computer Systems Engineering, Monash University, Australia.
Computers in Biology and Medicine
|December 7, 2023
Summary
This study introduces McSTRA, a novel transformer model for Magnetic Resonance Imaging (MRI) reconstruction. McSTRA enhances MRI image quality by integrating physics principles with advanced AI, outperforming existing methods in challenging conditions.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Biomedical Engineering
Background:
- Convolutional Neural Network (CNN) models, common in deep learning MRI reconstruction, struggle with global feature correlations due to their localized operations.
- Vision Transformer (ViT) models excel at capturing global correlations but often neglect MRI-specific physics in reconstruction tasks.
Purpose of the Study:
- To develop a novel physics-based transformer model for robust Magnetic Resonance Imaging (MRI) reconstruction.
- To address the limitations of existing methods in capturing global correlations and incorporating MRI physics.
Main Methods:
- Proposed the Multi-branch Cascaded Swin Transformers (McSTRA) model, integrating MRI physics with Swin transformers.
- Utilized shifted window self-attention for global MRI feature extraction and a multi-branch setup for spectral component analysis.
- Implemented a cascaded network with intermediate de-aliasing, data consistency, and loss computations.
- Introduced a point spread function-guided positional embedding for effective aliasing artifact reconstruction.
Main Results:
- McSTRA demonstrated superior performance compared to state-of-the-art MRI reconstruction methods.
- The model exhibited robustness under various challenging conditions, including high acceleration, noisy data, and anatomical variations.
- Achieved effective reconstruction by leveraging aliasing artifact spread through the proposed positional embedding mechanism.
Conclusions:
- McSTRA represents a significant advancement in deep learning-based MRI reconstruction by effectively combining transformer architecture with MRI physics.
- The proposed model offers robust and high-quality MRI reconstruction, outperforming existing methods in diverse and adverse scenarios.
- This physics-informed approach holds promise for improving diagnostic accuracy and efficiency in clinical MRI applications.
