Dual-domain accelerated MRI reconstruction using transformers with learning-based undersampling.

Area of Science:

• Medical Imaging
• Deep Learning
• Artificial Intelligence

Background:

• Accelerated MRI is crucial for imaging large volumes and dynamic processes like the heart.
• Current deep learning methods often use CNNs and fixed undersampling, limiting long-range dependency capture and reconstruction optimality.
• Leveraging temporal correlations in dynamic MRI data for improved reconstruction remains an underexplored area.

Purpose of the Study:

• To develop a novel deep learning framework for accelerated MRI reconstruction.
• To address limitations of CNN-based architectures and fixed undersampling patterns.
• To incorporate temporal correlations from dynamic MRI sequences for enhanced reconstruction.

Main Methods:

• Introduced DuDReTLU-net, a dual-domain (image and k-space) transformer-based reconstruction network.
• Paired the network with a learning-based undersampling strategy for dynamic MRI reconstruction.
• Trained the network end-to-end using fully sampled ground truth data and tested on cardiac CINE images.

Main Results:

• DuDReTLU-net demonstrated superior performance compared to state-of-the-art methods (LOUPE, k-t SLR, BM3D-MRI).
• Transformer-based reconstruction outperformed CNN-based methods in both image and k-space domains.
• Dual-domain architectures showed better results than single-domain ones, and dynamic sequence input improved accuracy over single-frame input.

Conclusions:

• The study highlights the effectiveness of transformer-based, dual-domain architectures and learning-based undersampling for accelerated MRI.
• Findings encourage further research into these advanced techniques for dynamic MRI reconstruction.
• The developed model offers a promising direction for improving the efficiency and quality of MRI scans.