Complementary time-frequency domain networks for dynamic parallel MR image reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Dynamic parallel MRI reconstruction is crucial for clinical applications.
• Current methods face challenges in balancing speed and image quality.
• Exploiting spatiotemporal correlations is key to improving reconstruction.

Purpose of the Study:

• To develop a novel deep learning-based approach for fast and high-quality dynamic multicoil MR reconstruction.
• To learn a complementary time-frequency domain network that exploits spatiotemporal correlations.
• To enable rapid, high-fidelity dynamic MRI acquisition.

Main Methods:

• Formulated dynamic parallel MR image reconstruction as a multivariable minimization problem.
• Developed an iterative algorithm using variable splitting in complementary domains (x-f and x-t).
• Embedded the iterative model into a deep recurrent neural network for image recovery.

Main Results:

• The proposed method outperformed state-of-the-art approaches quantitatively and qualitatively on cardiac cine MRI datasets.
• Demonstrated robust generalization to data from different scanners and unseen pathologies.
• Achieved high-quality reconstruction from highly undersampled multicoil data.

Conclusions:

• Deep neural networks effectively reconstruct dynamic parallel MRI in complementary time-frequency domains.
• The method enables fast (2.8 seconds) and robust high-quality image reconstruction from highly undersampled data.
• This facilitates fast single-breath-hold clinical 2D cardiac cine imaging.