DDC-Net: Dual-Domain Cascaded Network with POCS Prior for Fast MRI Reconstruction

Magnetic Resonance Imaging (MRI) is a powerful diagnostic tool that provides high-resolution images, but it requires a time-consuming scanning procedure. To reduce the acquisition time, various methods have been proposed to reconstruct images from under-sampled k-space. However, most dual-domain learning-based methods have two primary limitations: 1) employing the same reconstruction network for different domains may hinder the learning of targeted information for each domain; 2) the reconstruction network uses only highly undersampled data as input, leaving the use of prior as complementary information highly underexplored. To alleviate the above issues, we propose a Dual-Domain Cascaded Network with Projection onto Convex Set (POCS) prior, called DDC-Net. It consists of cascaded blocks with I-Net and K-Net reconstruction sub-networks. Specifically, the customized I-Net for the image domain employs a two-encoder-one-decoder architecture to enrich detailed structures. On the other hand, the customized K-Net for the frequency (k-space) domain replaces the traditional encoder-decoder with a cross-domain encoder-decoder to alleviate the difficult modeling problem of k-space data. Additionally, the POCS prior as complementary information is deeply embedded in the dual domain to simultaneously guide the k-space and image reconstructions. Extensive experiments on fastMRI and CC359 datasets demonstrate that our proposed DDC-Net achieves state-of-the-art qualitative and quantitative reconstruction results, outperforming previous deep learning-based methods.