Unfolding High-Order Correlations for Interpretable Multi-Contrast MRI Super-Resolution.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep unfolding networks are effective for MRI super-resolution (SR).
• Existing methods overlook high-order correlations within volumetric MRI data and use proximal gradient algorithms (PGA) with manual hyperparameters, leading to suboptimal solutions.

Purpose of the Study:

• To propose HocMRI, a deep unfolding multi-contrast MRI SR framework.
• To integrate dual-prior modeling and a hyperparameter-free PGA for enhanced MRI reconstruction.

Main Methods:

• A novel degradation model using dual-prior mechanisms: low-rank tensor factorization for slice dependencies and a Mamba-based network with 3D scanning for high-order correlations.
• Derivation of a hyperparameter-free PGA using a hyperbolic tangent function for dynamic step control, eliminating manual tuning.
• Unfolding the optimization algorithm into a multi-stage deep network for efficient iterative reconstruction.

Main Results:

• HocMRI effectively captures intra- and inter-slice dependencies and high-order correlations across slices.
• The hyperparameter-free PGA ensures stable convergence without manual tuning, outperforming traditional PGA.
• Experimental results show HocMRI achieves superior performance and enhanced efficiency compared to state-of-the-art methods on widely used MRI datasets.

Conclusions:

• HocMRI represents a significant advancement in multi-contrast MRI super-resolution.
• The proposed framework addresses limitations of existing methods by incorporating high-order correlations and a robust optimization algorithm.
• HocMRI offers a promising solution for high-quality MRI reconstruction with improved efficiency.