Digging Deeper in Gradient for Unrolling-Based Accelerated MRI Reconstruction.

Area of Science:

• Medical Imaging
• Image Reconstruction
• Magnetic Resonance Imaging (MRI)

Background:

• Accelerated MRI reconstruction commonly uses parallel imaging and compressed sensing.
• Current methods often fail to adequately recover high-frequency image details.
• This limitation impacts the quality of fine details in reconstructed MR images.

Purpose of the Study:

• To develop a novel MRI reconstruction model that enhances the recovery of high-frequency image information.
• To address the limitations of existing methods in capturing fine image details.
• To improve the quality of accelerated MRI reconstruction.

Main Methods:

• Proposed a novel MRI reconstruction model based on Maximum a Posteriori (MAP) estimation.
• Established a Cumulative Deviation from Maximum Gradient magnitude (CDMG) prior for MR images.
• Incorporated both explicit CDMG and implicit deep priors, and used a multi-order gradient operator.
• Unrolled the MAP estimation into a convolutional neural network (DDGU-Net) for optimization.

Main Results:

• The proposed model effectively recovers meaningful high-frequency information.
• The DDGU-Net model demonstrates improved accuracy in the likelihood term.
• Experimental results show high-quality MR image reconstruction.
• Achieved state-of-the-art (SOTA) performance, especially at higher acceleration factors.

Conclusions:

• The novel MAP-based reconstruction model with combined priors enhances fine detail recovery in accelerated MRI.
• The DDGU-Net architecture provides an effective framework for optimizing this model.
• The approach significantly improves MRI reconstruction quality and performance, particularly under accelerated sampling conditions.