Cwin-Net: A channel window attention network for magnetic resonance image super-resolution

In contemporary medical diagnostics and therapeutic interventions, high-fidelity magnetic resonance imaging (MRI) plays a pivotal role in ensuring accurate clinical assessments. However, prolonged MRI acquisition times present substantial challenges to both patient comfort and healthcare system efficiency. To address these limitations, we introduce Cwin-Net — a channel window attention network specifically designed for magnetic resonance image super-resolution. The proposed Cwin attention mechanism introduces two key innovations: First, by organically combining window-shift attention with channel weighting mechanisms, it achieves synergistic capture of both local details and global information. Second, the introduced dynamic channel weight adjustment mechanism enables the network to adaptively enhance the representation of key feature channels, thereby significantly improving detail preservation while maintaining global structure. Recognizing the inherent complexity of MR image characteristics, we implement a segmentation perceptual loss to incorporate anatomical prior knowledge, thereby prioritizing the reconstruction of clinically relevant textural patterns. The architecture further incorporates Feature Enhanced Blocks (FEB) to optimize deep feature integration, selectively amplifying diagnostically significant elements through learned weight parameters and frequency domain analysis. Extensive validation on the OASIS, ACDC, and Knee datasets demonstrates Cwin-Net’s superiority over state-of-the-art methods, achieving optimal performance in both quantitative metrics and visual quality assessments.