MRI Super-Resolution via Hybrid Information Enhancement Network based on Multi-Attention and Adaptive Convolution

Deep learning-based super-resolution (SR) reconstruction is a critical approach which is used to generate high-resolution images from corresponding low-resolution images. However, the CNN-based methods are ineffective in capturing global information, whereas the Transformer-based methods have limited ability to model long-range dependencies caused by window self-attention. Besides, it is a challenging task to recover lost high-frequency information from downsampled images. In this paper, a Hybrid Information Enhanced Network (HIEN) is proposed for MRI super-resolution task. Specifically, we propose a Spatial-Channel Hybrid Attention (SCHA) to enhance specific semantics representation by combining spatial and channel self-attention together. To recover more high-frequency components, we propose a Dynamic High-Frequency Pass Filter (DHPF) to preserve high-frequency information adaptively in pixel-wise. The results of our extensive experiments indicate that HIEN outperforms other state-of-the-art methods.