Deep-learning-based bright-field image generation from a single hologram using an unpaired dataset.

Area of Science:

• * Computational imaging
• * Machine learning for microscopy

Background:

• * Generating bright-field microscope images from hologram reconstructions is challenging.
• * Creating paired datasets for training neural networks is often impractical for certain holographic setups.

Purpose of the Study:

• * To apply an unpaired neural network training technique (CycleGAN) for generating bright-field microscope-like images from hologram reconstructions.
• * To evaluate the feasibility and performance of unpaired training in microscopy applications.

Main Methods:

• * Implementation of CycleGAN, an unpaired generative adversarial network.
• * Training the CycleGAN model on hologram reconstructions without paired bright-field images.
• * Comparison of results with traditional paired training methods.

Main Results:

• * CycleGAN successfully generated bright-field microscope-like images from hologram reconstructions.
• * The unpaired training approach yielded comparable results to paired training, even in challenging scenarios.
• * Unpaired training produced sharper and more visually realistic object reconstructions.

Conclusions:

• * Unpaired training with CycleGAN is a viable and effective method for generating microscope images from holograms.
• * This technique overcomes limitations associated with paired dataset creation in microscopy.
• * Lower metric scores in unpaired training do not necessarily indicate poorer performance but can reflect different, yet accurate, focal representations.