Physics-guided deep neural network for quantitative phase reconstruction from Zernike phase contrast microscopy

Area of Science:

• Biomedical Optics
• Microscopy
• Computational Imaging

Background:

• Zernike phase-contrast microscopy (ZPCM) is a common label-free imaging technique but provides only qualitative data.
• ZPCM suffers from artifacts like halo and shade-off, limiting its quantitative applications.
• Quantitative phase microscopy (QPM) offers precise measurements but requires complex and costly setups.

Purpose of the Study:

• To develop a method for reconstructing quantitative phase microscopy (QPM) images from standard Zernike phase-contrast microscopy (ZPCM) intensity images.
• To overcome the limitations of ZPCM, such as qualitative output and artifacts.
• To enable quantitative phase imaging using simpler, more accessible ZPCM instrumentation.

Main Methods:

• A physics-guided deep neural network was trained using a differentiable ZPCM forward model.
• The network learned to explicitly model ZPCM image formation from quantitative phase distributions.
• Supervised training enforced physical consistency, enabling reconstruction from single ZPCM images.

Main Results:

• The deep learning network successfully reconstructed QPM images from ZPCM intensity images.
• Artifacts such as halo and shade-off were significantly suppressed.
• Accurate phase recovery was achieved, with a Structural Similarity Index (SSIM) greater than 0.97.
• The reconstruction process is rapid, taking less than 10 ms per image on a standard GPU.

Conclusions:

• The proposed physics-guided deep neural network effectively converts ZPCM images to QPM images.
• This method offers a pathway to quantitative phase imaging with reduced artifacts using standard ZPCM.
• The approach is computationally efficient, supporting real-time quantitative microscopy applications.