DVS-Net: Dual-domain Variable Splitting Network for Accelerated Parallel MRI Data.

Area of Science:

• Medical Imaging
• Magnetic Resonance Imaging
• Deep Learning

Background:

• Accelerated parallel imaging in MRI reduces scan times and motion artifacts.
• Undersampled multi-coil data presents reconstruction challenges.
• Existing methods may not fully leverage available data for optimal reconstruction.

Purpose of the Study:

• To develop a joint frequency and image domain (dual-domain) reconstruction method for undersampled multi-coil MRI data.
• To improve reconstruction accuracy and perceptual quality in accelerated parallel MRI.
• To create an efficient, end-to-end trainable deep neural network for dual-domain MRI reconstruction.

Main Methods:

• A novel dual-domain reconstruction approach incorporating a full sampling condition.
• Development of an iterative algorithm using variable splitting and ADMM.
• Unrolling the iterative algorithm into a deep neural network for end-to-end training.
• Evaluation on complex-valued multi-coil knee MRI data with 6-fold and 8-fold acceleration.

Main Results:

• The proposed dual-domain network achieved superior reconstruction accuracy compared to variational and deep learning methods.
• Enhanced perceptual quality was observed in the reconstructed images.
• The method effectively utilizes dual-domain information for improved MRI reconstruction.

Conclusions:

• The joint dual-domain reconstruction method significantly improves the quality of accelerated parallel MRI data.
• This approach offers both visual and quantitative benefits for clinical applications.
• Deep learning unrolling of iterative algorithms provides a powerful framework for advanced MRI reconstruction.