An unrolled neural network for accelerated dynamic MRI based on second-order half-quadratic splitting model

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Reconstruction

Background:

• Dynamic MRI reconstruction from incomplete k-space data is crucial for reducing scan times.
• Traditional methods struggle with high acceleration factors, while deep learning models face challenges with parameter size and robustness.
• Unrolled deep learning models offer improved stability and flexibility for MRI reconstruction.

Purpose of the Study:

• To develop an advanced unrolled deep learning network for dynamic MRI reconstruction.
• To enhance image fidelity and reduce artifacts caused by undersampling in accelerated MRI.
• To improve the robustness and generalization capabilities of MRI reconstruction algorithms.

Main Methods:

• Proposed an unrolled deep learning network integrating a second-order Half-Quadratic Splitting (HQS) algorithm.
• Introduced a degradation-sense module to guide the iterative process using random sampling patterns.
• Incorporated an Information Fusion Transformer (IFT) for extracting local and non-local priors to mitigate aliasing artifacts.
• Applied low-rank constraints within the HQS algorithm to further optimize reconstruction quality.

Main Results:

• Each module of the proposed model demonstrated a contribution to improved reconstruction performance.
• The method achieved comparable results to state-of-the-art techniques with excellent generalization across various sampling masks.
• Significant Peak Signal-to-Noise Ratio (PSNR) improvements were observed: 0.7% at low acceleration, 3.4% at acceleration factor 8, and 5.8% at acceleration factor 12.

Conclusions:

• The proposed unrolled deep learning network effectively reconstructs dynamic MRI from undersampled k-space data.
• The integration of HQS, a degradation-sense module, IFT, and low-rank constraints enhances reconstruction accuracy and robustness.
• This approach offers a promising solution for accelerating MRI scans while maintaining high image quality and generalizability.