Dynamic feature arbitration and synergistic attention for high-fidelity brain MRI super-resolution

High-fidelity super-resolution (SR) of brain structural MRI (sMRI) is crucial for advancing neuroimaging research and clinical applications. However, existing deep learning approaches often struggle to simultaneously recover fine textures and maintain global anatomical consistency, largely due to insufficient integration of spatial and channel information. To address these challenges, we propose a three-stage cross-fusion network for brain sMRI super-resolution, consisting of shallow feature extraction, deep feature extraction, and image reconstruction. The deep feature extraction stage is built around two key modules. First, an Adaptive Gated Cross-Fusion (AGCF) block introduces an Adaptive Gated Cross-Attention (AGX) mechanism, enabling dynamic, bidirectional interactions between spatial and channel pathways. Second, we propose a Lightweight Sliding Window Attention (Lite-SWA) block for efficient multi-branch fusion. The SWA branch focuses on fine-grained anatomical textures, while the DSConv branch provides low-cost contextual support. We then use a lightweight interleaved channel–spatial fusion (ICSF) gate to adaptively merge the two branches. Together, these designs provide a balanced representation that preserves local detail without compromising global anatomical integrity. We evaluate the proposed model on a heterogeneous cohort of 2501 subjects from eight public datasets. Results show consistent state-of-the-art performance across multiple upscaling factors. Extensive ablation studies further validate the effectiveness of the proposed modules, underscoring their potential to enhance the fidelity and clinical reliability of brain MRI super-resolution.