Solving Fourier ptychographic imaging problems via neural network modeling and TensorFlow.

Area of Science:

• Microscopy and Imaging Technologies
• Computational Imaging
• Machine Learning in Science

Background:

• Fourier ptychography enables large field-of-view, high-resolution microscopy.
• Modeling imaging processes with neural networks offers new computational approaches.

Purpose of the Study:

• To model the Fourier ptychographic forward imaging process using a convolutional neural network (CNN).
• To recover complex object information through network training and accelerate phase retrieval.
• To demonstrate the network's applicability to other imaging modalities like SIM.

Main Methods:

• A CNN was developed where the input is the point spread function or coherent transfer function, and the object is treated as learnable weights.
• The network was trained using TensorFlow, optimizing parameters like learning rate, solver, and batch size.
• Fourier-magnitude projection was implemented using a multiplication neural network model to accelerate phase retrieval.

Main Results:

• A large batch size with the Adam optimizer generally yielded the best performance for phase retrieval.
• The CNN model successfully demonstrated 4-frame resolution doubling in structured illumination microscopy (SIM).
• The approach provides a framework for modeling various coherent and incoherent imaging systems.

Conclusions:

• Modeling imaging systems with neural networks, particularly CNNs, offers a powerful approach for image reconstruction and acceleration.
• This framework facilitates the use of specialized hardware (e.g., TPUs) for faster image processing.
• The open-source implementation encourages further research in machine learning for microscopy.