Joint Under-Sampling Pattern and Dual-Domain Reconstruction for Accelerating Multi-Contrast MRI.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Multi-Contrast Magnetic Resonance Imaging (MCMRI) accelerates MRI scans by using reference images.
• Existing MCMRI methods have limitations including redundant data, under-exploration of k-space information, and lack of physical interpretability in network design.

Purpose of the Study:

• To develop a novel deep-unfolding dual-domain network for accelerating MCMRI.
• To jointly optimize under-sampling patterns and image reconstruction for improved MCMRI performance.

Main Methods:

• A framework to learn optimal under-sampling patterns for MCMRI, reducing redundancy and sampling contrast-specific information.
• A dual-domain network model reconstructing images in both image and k-space domains, incorporating spatial transformation for structural inconsistencies and k-space learning for global information.
• Optimization using proximal gradient algorithm and unfolding into a deep network (MC-DuDoN).

Main Results:

• The proposed MC-DuDoN demonstrates superior performance in MCMRI super-resolution and reconstruction tasks.
• The method shows robustness on spatially misaligned MCMRI due to explicit modeling of inconsistent structures.
• Learned masks in reconstruction tasks restore more realistic images, even at ultra-high acceleration ratios (×30).

Conclusions:

• The developed MC-DuDoN effectively accelerates MCMRI by optimizing under-sampling and utilizing dual-domain reconstruction.
• The approach offers improved accuracy, robustness to misalignment, and enhanced image realism at high acceleration factors.