Deep learning-based hologram generation using a white light source.

Area of Science:

• Microscopy and Imaging
• Computational Imaging
• Biophysics

Background:

• Digital holographic microscopy (DHM) captures 3D volumetric data but requires specific optical setups like coherent light sources and pinholes.
• Conventional bright-field (BF) microscopy is widely accessible but lacks 3D information and high contrast for certain samples.

Purpose of the Study:

• To develop a deep neural network, specifically a generative adversarial network (GAN), for transforming standard bright-field (BF) images into holographic images.
• To enable 3D volumetric information acquisition and enhanced imaging from BF microscopy using a general white light source.

Main Methods:

• A hybrid BF and hologram imaging technique was used to create a dataset of 11,050 training image pairs.
• A generative adversarial network (GAN) was trained to perform image transformation from defocused BF images to holographic images.
• The performance of the GAN was validated by comparing generated holograms with ground truth holograms of microspheres and erythrocytes.

Main Results:

• The GAN successfully generated holographic images from BF images, exhibiting significantly enhanced image contrast.
• The signal-to-noise ratio of the generated holograms was 3-5 times higher than that of ground truth holograms.
• The generated holograms provided accurate 3D positional information and revealed light scattering patterns of microscale samples.

Conclusions:

• The developed GAN-based method offers a promising approach for dynamic analysis of microscale objects using conventional BF microscopy settings.
• This technique allows precise monitoring of biological samples by providing detailed 3D positional information and scattering characteristics.
• The method overcomes the limitations of traditional DHM by eliminating the need for specialized optical components.