PhaseStain: the digital staining of label-free quantitative phase microscopy images using deep learning

Area of Science:

• Biomedical Imaging
• Computational Pathology
• Artificial Intelligence in Medicine

Background:

• Quantitative Phase Imaging (QPI) provides label-free imaging of biological tissues.
• Histological staining is crucial for visualizing tissue morphology in pathology but is time-consuming and costly.
• Deep learning offers novel approaches for image analysis and transformation.

Purpose of the Study:

• To develop and validate a digital staining technique (PhaseStain) using deep neural networks.
• To transform label-free Quantitative Phase Images (QPI) into images resembling traditional histological stains.
• To reduce the reliance on physical staining in pathology and biomedical research.

Main Methods:

• A deep neural network, specifically a generative adversarial network (GAN), was trained.
• The GAN was trained using paired image data: label-free QPI and corresponding brightfield microscopy images of stained tissue sections.
• The technique was validated on human skin, kidney, and liver tissue samples.

Main Results:

• PhaseStain successfully transformed QPI into images equivalent to Hematoxylin and Eosin (H&E), Jones' stain, and Masson's trichrome stained images.
• The virtual staining accurately reproduced the morphological details typically visualized with conventional histological stains.
• The method demonstrated effectiveness across different human tissue types.

Conclusions:

• PhaseStain offers a viable digital staining solution for label-free QPI, enhancing its utility in pathology.
• This AI-powered virtual staining eliminates the need for physical histological staining, reducing costs and preparation time.
• The study highlights the potential of deep learning for data-driven image transformations in biomedical research.