Rethinking Dual-Domain Undersampled MRI Reconstruction: Domain-Specific Design from the Perspective of the Receptive Field

Undersampled MRI reconstruction is crucial for accelerating clinical scanning. Dual-domain reconstruction networks are performant among SoTA deep learning methods by leveraging complementary information of two domains. In this paper, we rethink dual-domain network design from the perspective of the receptive field. We formulate k-space reconstruction as an interpolation problem and use the probability for a successful interpolation given a receptive field as a heuristic for building k-space networks. We divide image recovery into two sub-problems: artifact removal and detail refinement, corresponding to a global and a local task, respectively. To this end, we introduce two domain-specific modules for dual-domain networks, namely k-space global initialization and image-domain parallel local detail enhancement. We evaluate our modules by translating a SoTA method Du-DoRNet under different conventions of MRI reconstruction including image-domain, dual-domain, and reference-guided reconstruction on the public IXI dataset. Our model DuDoR-Net+ achieves significant improvements over competing deep learning methods.