Object detection neural network improves Fourier ptychography reconstruction.

Area of Science:

• Microscopy
• Optical imaging
• Computational imaging

Background:

• Superresolution microscopy relies on high-quality optics and aberration correction.
• Fourier ptychography offers superresolution, aberration correction, and quantitative phase imaging.
• Accurate illumination angle estimation is crucial for Fourier ptychography's forward model.

Purpose of the Study:

• To develop a robust method for illumination estimation in Fourier ptychography.
• To leverage deep learning for improved accuracy and efficiency in illumination estimation.
• To investigate the impact of raw data processing on illumination estimation performance.

Main Methods:

• Formulating illumination estimation as an object detection problem.
• Employing a faster region-based convolutional neural network (faster-RCNN) for illumination estimation.
• Comparing CNN-based method against classical approaches using raw microscopy data.

Main Results:

• The faster-RCNN approach achieved up to a 3-fold reduction in illumination estimation errors.
• The CNN-based method demonstrated superior robustness compared to traditional techniques.
• Conventional data smoothing and filtering were found to be detrimental to estimation accuracy.

Conclusions:

• Illumination estimation can be effectively treated as an object detection task for Fourier ptychography.
• Convolutional neural networks offer a powerful and efficient solution for accurate illumination estimation.
• Open-source software is provided for the developed CNN-based illumination estimation technique.