SAFNet: a spatial adaptive fusion network for dual-domain undersampled MRI reconstruction

Undersampled magnetic resonance imaging (MRI) reconstruction reduces scanning time while preserving image quality, improving patient comfort and clinical efficiency. Current parallel reconstruction strategies leverage k-space and image domains information to improve feature extraction and accuracy. However, most existing dual-domain reconstruction methods rely on simplistic fusion strategies that ignore spatial feature variations, suffer from constrained receptive fields limiting complex anatomical structure modeling, and employ static frameworks lacking adaptability to the heterogeneous artifact profiles induced by diverse undersampling patterns. This paper introduces a Spatial Adaptive Fusion Network (SAFNet) for dual-domain undersampled MRI reconstruction. SAFNet comprises two parallel reconstruction branches. A Dynamic Perception Initialization Module (DPIM) in each encoder enriches receptive fields for multi-scale information capture. Spatial Adaptive Fusion Modules (SAFM) within each branch's decoder achieve pixel-wise adaptive fusion of dual-domain features and incorporate original magnitude information, ensuring faithful preservation of intensity details. The Weighted Shortcut Module (WSM) enables dynamic strategy adaptation by scaling shortcut connections to adaptively balance residual learning and direct reconstruction. Experiments demonstrate SAFNet's superior accuracy and adaptability over state-of-the-art methods, offering valuable insights for image reconstruction and multimodal information fusion.