High-dimensional embedding network derived prior for compressive sensing MRI reconstruction

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Deep learning significantly accelerates medical imaging, but current methods require retraining for new undersampling patterns.
• A flexible prior is needed to adapt deep learning models to various magnetic resonance imaging (MRI) acquisition parameters.

Purpose of the Study:

• To develop a general deep learning prior for magnetic resonance imaging (MRI) reconstruction that is adaptable to diverse undersampling patterns.
• To improve the flexibility and performance of deep learning-based MRI reconstruction.

Main Methods:

• Introduced the multi-channel enhanced Deep Mean-Shift Prior (MEDMSP) by integrating multi-model aggregation and multi-channel network learning.
• Formulated a high-dimensional embedding network derived prior.
• Applied the learned prior to single-channel MRI reconstruction using a variable augmentation technique and solved using proximal gradient descent.

Main Results:

• The MEDMSP model demonstrated superior performance across various undersampling trajectories and acceleration factors.
• Consistent improvements in image reconstruction quality were observed compared to existing methods.
• The proposed method effectively leverages learned priors for flexible and robust MRI reconstruction.

Conclusions:

• The MEDMSP provides a general and flexible prior for highly undersampled MRI reconstruction.
• This approach overcomes the limitations of retraining deep learning models for specific sampling patterns.
• MEDMSP shows significant potential for accelerating MRI acquisition while maintaining high image fidelity.