Pixel super-resolution for lens-free holographic microscopy using deep learning neural networks

Area of Science:

• Biomedical Optics
• Microscopy
• Computational Imaging

Background:

• Lens-free holographic microscopy (LFHM) offers cost-effective, large field-of-view imaging for biomedical applications.
• LFHM's spatial resolution is limited by imager pixel size due to unit optical magnification.
• Pixel super-resolution (PSR) techniques improve resolution by combining sub-pixel shifted low-resolution (LR) holograms into a high-resolution (HR) hologram.

Purpose of the Study:

• To develop a deep learning-based PSR approach to overcome the limitations of conventional iterative PSR methods in LFHM.
• To accelerate data acquisition and reconstruction for practical, fast, lens-free super-resolution imaging.

Main Methods:

• A neural network-based approach for end-to-end HR hologram reconstruction from LR holograms.
• Training the network using synthesized data from LR holograms, eliminating the need for HR ground truth.
• Utilizing a reduced number of LR holograms for consistent resolution enhancement.

Main Results:

• The deep learning PSR method significantly accelerates HR hologram reconstruction compared to conventional iterative methods.
• The approach achieves consistent resolution improvement with fewer LR holograms.
• The trained network demonstrated effectiveness and robustness across various sample types, even when trained on different datasets.

Conclusions:

• Deep learning-based PSR provides a practical solution for fast, lens-free, super-resolution imaging.
• This method enhances LFHM throughput by reducing acquisition and reconstruction times.
• The technique offers a viable alternative for high-resolution imaging in resource-limited settings.