Snapshot coherent diffraction imaging via a physics-embedded untrained neural network

Coherent diffraction imaging (CDI) is a lens-free imaging paradigm that avoids the diffraction limit caused by lens chromatic aberration. However, due to the ill-posed problem of phase retrieval, experiments often require the collection of multiple diffraction patterns for image reconstruction, which limits its application in dynamic imaging scenarios. In this Letter, we propose a physics-embedded untrained neural network for snapshot coherent diffraction imaging. Our network embeds a physical model of diffraction propagation and can be trained in an unsupervised learning paradigm. Moreover, the proposed method is applicable to complex-valued samples and is flexible for various imaging settings. Simulation and experiments demonstrate that the proposed physics-embedded network architecture performs better and achieves state-of-the-art results in snapshot CDI, compared with existing unsupervised methods.