A multi-domain feature alignment and hierarchical cross feature enhancement network for under-sampled magnetic image reconstruction

Magnetic resonance imaging (MRI) is widely used in clinical diagnosis due to its high resolution and non-invasive scanning capabilities. However, long scanning times limit its development. To reduce acquisition time and obtain high-quality reconstructed images, a novel multi-domain MRI reconstruction network that fully utilizes the image domain, k-space, and wavelet domain is proposed. This network includes a parallel convolutional neural network (CNN) with k-space and wavelet domain branches, as well as a U-shaped image domain network. Following the parallel dual-domain CNN, a dual-domain feature alignment module aligns features from the k-space and wavelet domains into a unified representation space, mitigating artifact impacts. This design enhances the model’s understanding of multi-domain signals and improves generalization. Additionally, in the image domain, a hierarchical cross-feature enhancement module, based on Nested UNet, incorporates two cross-attention modules into different hierarchical skip connections of the Nested U-Net to reduce information propagation loss and enhance feature representation. Deep supervision within the image domain network further boosts the network’s performance and robustness. Extensive experiments on two public MRI datasets, FastMRI and CC359, as well as the private clinical dataset, validate the proposed method. Compared to several state-of-the-art deep learning methods, our approach demonstrates good reconstruction performance in both numerical assessments and visual effects.