Artifact-free phase reconstruction for differential interference contrast microscopy based on deep learning.

Area of Science:

• Biophotonics
• Microscopy
• Image Processing

Background:

• Differential interference contrast (DIC) microscopy is a valuable label-free imaging technique for cells and tissues.
• Accurate quantitative phase reconstruction from DIC images is challenging due to non-linearity and artifacts.
• Existing numerical integration methods suffer from unknown constants and noise sensitivity.

Purpose of the Study:

• To develop a data-driven approach for artifact-free, high-precision, and fast quantitative phase reconstruction in DIC microscopy.
• To overcome the limitations of traditional phase reconstruction methods in DIC imaging.
• To enable accurate quantitative phase imaging using only a single differential phase image.

Main Methods:

• Constructed a training database of "specimen phase-differential phase" using digital holography and the DIC imaging model.
• Employed the Pix2Pix Generative Adversarial Network (GAN) model with optimized loss functions and gradient backpropagation.
• Trained the deep neural network to learn the mapping between specimen phase and differential phase.

Main Results:

• Achieved high-precision, artifact-free reconstruction of specimen phase from single-direction DIC images.
• Demonstrated effectiveness in quantitative phase imaging of polystyrene spheres and HeLa cells.
• Showcased excellent anti-noise performance and fast reconstruction capabilities.

Conclusions:

• The proposed data-driven method offers a promising solution for accurate quantitative phase reconstruction in DIC microscopy.
• This technology facilitates high spatial sensitivity detection, enhancing the analysis of biological samples.
• The Pix2Pix GAN approach provides a robust and efficient alternative to traditional phase reconstruction techniques.