Dynamic Focus Mechanism-Based Dual-Domain Reconstruction Network for Accelerated MRI.

The integration of information from both k-space and image domains has become a topic of growing interest in accelerated Magnetic Resonance Imaging (MRI) using dual-domain reconstruction networks. However, many existing methods treat sampled and unsampled measurements equally in the k-space domain, leading to progressive feature shift and suboptimal reconstruction performance. Additionally, networks with single-scale receptive field in the image domain are insufficient for fully exploring global-local information and lack channel-wise feature interaction to preserve structural details. To address these issues, we propose a novel dual-domain reconstruction network named DFRNet, which incorporates a dynamic focus mechanism. In the k-space domain, we introduce an area normalization module that dynamically eases feature shift and predicts reliable k-space measurements. In the image domain, our proposed dynamical attention module with channel-wise gating mechanism extracts rich globallocal features for detail recovery. Quantitative and qualitative experiments show that our proposed DFRNet achieves competitive performance to several state-of-the-art methods while being more lightweight.