Dual-path network with dual-domain fusion and cross-attention for MRI reconstruction

Accelerated magnetic resonance imaging (MRI) involves mapping the under-sampled k-space representation to reconstruct a high-quality image, and it remains a central challenge in MRI reconstruction. In recent years, deep learning has significantly improved the reconstruction performance in MRI. However, there is still a concern in the field regarding how to enhance the global feature learning ability of deep networks to further improve the quality of reconstructed images. To address this issue, this paper proposes a novel MRI reconstruction model called DDFCA-Net, which is based on a dual-path network with dual-domain fusion and cross-attention. The proposed DDFCA-Net model consists of two parallel and interactive paths. One path utilizes a convolutional neural network (CNN) to extract deep features from the k-space domain, while the other path employs a vision transformer (ViT) to extract deep features from the image domain. The image domain features and k-space features are mutually enhanced through cross-fusion. In the proposed model, the ViT network adopts a cross-attention learning strategy. The query matrix is derived from the image domain feature, while both the key matrix and value matrix are obtained from the fusion of the image domain feature and the k-space feature. This cross-attention mechanism promotes effective interaction between the two domains, further enhancing the feature extraction ability. Extensive experimental results on the CC359-Brain and FastMRI single-coil knee datasets, on different sampling rates and strategies, validate the effectiveness of the proposed method. The DDFCA-Net outperforms state-of-the-art methods, demonstrating superior reconstruction performance in MRI. • A novel parallel dual-domain network, dubbed DDFCA-Net, is proposed for MRI reconstruction that combines CNN and Transformer, which enables effective feature extraction from both the image domain and the k-space domain. • A cross-domain fusion learning approach is introduced for MRI reconstruction based on cross-attention Transformer, which facilitates the interaction and enhancement of feature extraction. • Extensive experiments are conducted on multiple publicly available datasets that demonstrate the effectiveness of the proposed method.