Complex-valued Multi-scale Hybrid Attention Network for Fast MRI via Sparsified Data Learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Accelerated MRI enables faster scans but often results in undersampled k-space data.
• Reconstruction of high-quality images from undersampled data is crucial for clinical applications.
• Existing methods struggle with preserving image quality and phase information under aggressive undersampling.

Purpose of the Study:

• To propose and validate a novel complex-valued multi-scale attention U-Net framework (SCMAU-Net) for accelerated MRI reconstruction.
• To enhance feature extraction and fusion for improved image reconstruction accuracy.
• To evaluate the performance of SCMAU-Net against existing state-of-the-art methods.

Main Methods:

• Development of a complex-valued multi-scale attention U-Net (SCMAU-Net) architecture.
• Integration of multi-scale dilated convolutions for enlarged receptive fields and enhanced feature extraction.
• Application of hybrid channel and spatial attention mechanisms for multi-scale feature fusion and complex-valued attention gates.

Main Results:

• SCMAU-Net demonstrated significant improvements in SSIM (1.6%), PSNR (0.87 dB), and NMSE (16.2%) compared to SCU-Net at R=4.27.
• Achieved up to 20.6% reduction in absolute phase disparity (APD) compared to E2E-VarNet, particularly in brain regions.
• Maintained competitive SSIM, PSNR, and NMSE under aggressive undersampling, outperforming existing methods in phase-sensitive applications.

Conclusions:

• SCMAU-Net effectively reconstructs high-quality images from undersampled k-space data.
• The proposed architecture leverages multi-scale convolutions and attention mechanisms for superior performance.
• SCMAU-Net shows particular promise for phase-sensitive MRI applications requiring accurate phase reconstruction.