A Deep Dual-Domain Interaction Reconstruction Framework With Adaptive Gating Fusion for Low-Field MRI.
Authors
Affiliations (4)
Affiliations (4)
- School of Instrument Science and Engineering, State Key Laboratory of Comprehensive PNT Network and Equipment Technology, Southeast University, Nanjing, China.
- Jiangsu Key Laboratory for Design and Manufacturing of Precision Medicine Equipment, Southeast University, Nanjing, Jiangsu, China.
- National Platform for Medical Engineering Education Integration, Southeast University, Nanjing, Jiangsu, China.
- School of Mechanical Engineering, Southeast University, Nanjing, Jiangsu, China.
Abstract
Low-field magnetic resonance imaging (MRI) offers a cost-effective and accessible alternative to high-field systems but inherently suffers from low signal-to-noise ratio (SNR) and prolonged scan times. Although k-space under-sampling shortens scan time, the resulting incomplete data typically introduces resolution loss and artifacts in the reconstructed images. Deep learning-based MRI reconstruction methods primarily focus on high-field, high-quality MRI data and image-domain reconstruction, with limited attention to the unique characteristics of low-field data and the relationship between k-space and image-domain representations. Furthermore, existing k-space reconstruction approaches often overlook the intrinsic encoding dependencies within k-space and lack mechanisms for learnable feature fusion across the two domains. To address the limitations, we propose DUAG, a deep dual-domain interaction reconstruction framework with adaptive gating fusion for low-field MRI. The framework employs a cascaded deep architecture with multi-scale U-Net structures to achieve hierarchical feature representation and incorporates a hybrid dual-domain interaction module. Attention mechanisms are introduced to model long-range dependencies, enabling precise capture of underlying correlations among k-space frequency encodings and image-domain pixels. In addition, an adaptive gating fusion strategy is designed for dynamical weighting and cross-domain feature fusions. In order to enhance feature reuse and improve the generalization ability of the model. Experiments on public 0.3 T low-field dataset show that DUAG reaches 42.79 ± 0.75 PSNR and 0.920 ± 0.010 SSIM. To further verify the generalization capability, we conducted real-world experiments on the laboratory-collected 0.5 T low-field MRI scanner data, DUAG achieves superior reconstruction performance of 34.13 ± 1.02 PSNR and 0.889 ± 0.013 SSIM. The proposed framework provides a promising solution for high-quality low-field MRI reconstruction, which is expected to promote the deployment of cost-effective MRI systems in resource-constrained clinical settings.