Fourier ptychographic microscopy reconstruction with multiscale deep residual network

Area of Science:

• Microscopy
• Image Reconstruction
• Computational Imaging

Background:

• Fourier ptychographic microscopy (FPM) enables high-resolution, wide-field, quantitative phase imaging.
• Conventional FPM reconstruction relies on iterative methods like Alternate Projection.
• FPM shows significant potential in biomedical fields such as hematology and pathology.

Purpose of the Study:

• To develop a novel deep learning framework for FPM reconstruction.
• To design and implement a multiscale, deep residual neural network for enhanced FPM imaging.
• To evaluate the performance of the deep learning model against conventional methods and other neural network approaches.

Main Methods:

• A multiscale, deep residual neural network was designed and implemented using PyTorch.
• A large-scale simulation dataset and an actual captured dataset were created for training and testing.
• The deep learning model was trained on simulated data and fine-tuned with actual FPM data.

Main Results:

• The proposed deep learning model achieved superior reconstruction quality compared to conventional methods.
• The model demonstrated significantly reduced reconstruction time.
• The deep learning approach exhibited enhanced robustness against system aberrations like noise and blurring.

Conclusions:

• Deep learning provides a powerful and efficient alternative for FPM reconstruction.
• The developed neural network framework offers improved performance and robustness for biomedical imaging applications.
• This method accelerates FPM analysis and enhances its reliability in challenging imaging conditions.