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

This study proposes a complex-valued multi-scale attention U-Net framework (SCMAU-Net) for accelerated MRI (magnetic resonance imaging) reconstruction from undersampled k-space data. The method employs a complex-valued difference transform to sparsify training, validation, and testing data, and extends the complex-valued U-Net architecture by integrating multi-scale dilated convolutions to enlarge the receptive field and enhance feature extraction, alongside hybrid channel and spatial attention mechanisms for multi-scale feature fusion and complex-valued attention gates in the decoder to prioritize salient features. An inverse filtering reconstruction restores image contrast. Quantitatively, compared with SCU-Net at the highest acceleration factor (R = 4.27), SCMAU-Net improves SSIM by 1.6%, PSNR by 0.87 dB, and reduces NMSE by 16.2%, confirming the effectiveness of the proposed architecture. Against E2E-VarNet, SCMAU-Net achieves up to a 20.6% reduction in APD (absolute phase disparity) in full-image evaluation and consistently yields lower APD within brain regions, while maintaining competitive SSIM, PSNR, and NMSE under aggressive undersampling. By leveraging multi-scale convolutions and attention mechanisms with sparsified data training, SCMAU-Net achieves superior reconstruction for undersampled k-space data, particularly in phase-sensitive MRI applications.