Automated phase unwrapping in digital holography with deep learning.

Area of Science:

• Biomedical Optics
• Computational Imaging
• Machine Learning for Biology

Background:

• Digital holography provides quantitative phase images crucial for analyzing biological sample morphology and content.
• Numerical reconstruction in digital holography results in phase values limited to -π to π, causing discontinuities due to the modulo 2π operation.
• Accurate phase unwrapping is essential for reliable quantitative phase imaging but is challenging with abrupt phase changes.

Purpose of the Study:

• To develop an automated deep learning model for reconstructing unwrapped focused-phase images from digital holography data.
• To improve the accuracy and speed of phase unwrapping compared to traditional numerical methods.
• To enable real-time observation of biological cell morphology and movement.

Main Methods:

• Integration of digital holography with a Pix2Pix generative adversarial network (GAN) for image-to-image translation.
• Development of a deep learning model for automatic phase unwrapping.
• Comparative analysis against numerical phase unwrapping techniques and U-net models.

Main Results:

• The proposed GAN model successfully reconstructs unwrapped focused-phase images, overcoming limitations of modulo 2π discontinuity.
• The method achieves phase unwrapping at twice the rate of conventional numerical approaches.
• The model demonstrates robust generalization across different cell types and superior performance compared to U-net models.

Conclusions:

• The developed deep learning approach offers a significant advancement in quantitative phase imaging using digital holography.
• This method provides a faster and more accurate solution for phase unwrapping, particularly in the presence of complex phase variations.
• The technique holds promise for real-time monitoring of cellular dynamics and morphology in biological research.