Sparse phase retrieval using a physics-informed neural network for Fourier ptychographic microscopy.

Area of Science:

• Optics and Photonics
• Computational Imaging
• Artificial Intelligence in Microscopy

Background:

• Fourier ptychographic microscopy (FPM) is a powerful imaging technique.
• Phase retrieval is crucial for reconstructing high-resolution images in FPM.
• Conventional methods struggle with sparsely sampled data and low overlap rates.

Purpose of the Study:

• To develop a sparse phase retrieval framework for FPM.
• To leverage physics-informed neural networks (PINNs) for improved phase retrieval.
• To enable reconstruction from limited data without additional driving information.

Main Methods:

• Implemented a sparse phase retrieval framework using PINNs.
• Trained bidirectional mappings between noisy image and object spaces.
• Integrated image formation physics with convolutional neural networks.
• Modified the mean absolute error loss function for signal characteristics.

Main Results:

• Successfully reconstructed sparsely sampled FPM data.
• Achieved high-quality reconstructions with a small aperture overlapping rate.
• Demonstrated superior performance compared to conventional methods.
• Framework operates without requiring additional data driving.

Conclusions:

• The proposed PINN-based framework offers a robust solution for sparse phase retrieval in FPM.
• This approach overcomes limitations of traditional methods in data-scarce scenarios.
• PINNs provide a promising direction for advancing computational microscopy techniques.