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

Area of Science:

• Optics and Photonics
• Computational Imaging
• Machine Learning Applications

Background:

• Coherent Diffraction Imaging (CDI) offers lens-free microscopy, overcoming lens-based chromatic aberration and diffraction limits.
• Traditional CDI requires multiple diffraction patterns for phase retrieval, hindering dynamic imaging applications.
• The ill-posed nature of phase retrieval in CDI presents a significant challenge for real-time reconstruction.

Purpose of the Study:

• To develop a novel, physics-embedded, untrained neural network for snapshot Coherent Diffraction Imaging (CDI).
• To enable single-shot image reconstruction in CDI, overcoming limitations of multi-pattern acquisition.
• To provide a flexible and robust method applicable to complex-valued samples and diverse imaging setups.

Main Methods:

• A physics-embedded neural network architecture was designed, incorporating the physical model of diffraction propagation.
• The network was trained using an unsupervised learning paradigm, eliminating the need for labeled data.
• The method was validated through both computational simulations and experimental setups.

Main Results:

• The proposed physics-embedded network achieved state-of-the-art results in snapshot CDI.
• Demonstrated superior performance compared to existing unsupervised methods for single-shot image reconstruction.
• Successfully reconstructed images from complex-valued samples with high fidelity.

Conclusions:

• The developed physics-embedded neural network offers a significant advancement for snapshot CDI.
• This unsupervised approach enhances the speed and applicability of CDI for dynamic imaging.
• The method shows great potential for various imaging settings and complex sample analysis.