Coherent modulation imaging using a physics-driven neural network.

Area of Science:

• Optics and Photonics
• Computational Imaging
• Machine Learning

Background:

• Coherent modulation imaging (CMI) is a diffraction imaging technique for complex field reconstruction.
• Traditional CMI relies on iterative algorithms to solve ill-conditioned phase retrieval problems.
• Deep learning offers powerful optimization for complex inverse problems in imaging.

Purpose of the Study:

• To develop a physics-driven neural network (CMINet) for complex-valued object reconstruction in CMI.
• To leverage deep learning for enhanced phase retrieval from single diffraction patterns.
• To improve reconstruction quality, speed, and robustness in CMI.

Main Methods:

• Developed CMINet, a physics-driven neural network tailored for CMI.
• Employed a customized physical-model-based loss function for network training, avoiding ground truth data.
• Validated the approach using simulation experiments for reconstruction quality and robustness.

Main Results:

• CMINet achieved high-quality reconstructions with reduced noise and improved robustness to physical parameter variations.
• The trained network enabled fast reconstruction of dynamic processes, bypassing frame-by-frame iterations.
• Biological experiments demonstrated CMINet's capability to reconstruct high-resolution amplitude and phase images with sharp details.

Conclusions:

• CMINet presents a novel and effective deep learning solution for CMI.
• The physics-driven approach enhances reconstruction accuracy and efficiency.
• CMINet shows significant potential for practical biological imaging applications requiring detailed amplitude and phase information.