HoloPhaseNet: fully automated deep-learning-based hologram reconstruction using a conditional generative adversarial...

Area of Science:

• Biophysics
• Optical Microscopy
• Computational Imaging

Background:

• Quantitative phase imaging (QPI) using off-axis digital holography offers insights into cellular morphology and intracellular content.
• Conventional QPI relies on complex numerical reconstruction, requiring precise parameter settings like reconstruction distance and phase unwrapping.
• Eliminating twin images is typically necessary for accurate phase retrieval.

Purpose of the Study:

• To demonstrate deep learning's capability to perform light propagation for QPI.
• To show that deep learning can eliminate the need for specific reconstruction parameters.
• To investigate the removal of twin-image elimination techniques in QPI.

Main Methods:

• A conditional generative adversarial network (cGAN) model was trained as an image generator.
• Holograms at the single-cell level were used as input to the trained model.
• The model was trained using holograms of size 512x512 pixels.

Main Results:

• Deep learning successfully performed the light propagation task, generating phase images.
• The method eliminated the requirement for precise reconstruction parameters.
• Twin-image elimination was found to be unnecessary for retrieving quantitative phase images.
• The model demonstrated generalization capabilities.

Conclusions:

• Deep learning offers a simplified and robust approach to quantitative phase imaging.
• AI-driven methods can overcome limitations of conventional holographic reconstruction.
• This technique enhances the accessibility and efficiency of cellular imaging analysis.