High-resolution lensless holographic microscopy using a physics-aware deep network.

Area of Science:

• Quantitative phase imaging
• Computational microscopy
• Deep learning for imaging

Background:

• Lensless digital inline holographic microscopy (LDIHM) is an emerging technique for quantitative phase imaging.
• Existing deep learning methods for LDIHM require extensive training data or lack robustness for complex biological samples.
• Physics-aware deep networks improve reconstruction without prior training but struggle with data fidelity.

Purpose of the Study:

• To develop a hybrid deep learning framework combining trained and untrained models for high-resolution phase recovery in LDIHM.
• To overcome the limitations of separate trained and physics-aware untrained deep networks.
• To achieve high-fidelity phase reconstruction from single low-resolution holograms.

Main Methods:

• Proposed a hybrid deep framework (HDPhysNet) integrating pre-trained high-definition generative adversarial networks (HDGAN) and physics-aware untrained deep networks.
• Utilized a plug-and-play approach where HDGAN generates high-resolution phase, which then regularizes the physics-aware network's reconstruction.
• Integrated physics of holography within the loss function for robust phase recovery.

Main Results:

• HDPhysNet demonstrated improved performance with higher Structural Similarity Index Measure (SSIM) and phase Signal-to-Noise Ratio (SNR) compared to purely trained or untrained deep networks.
• Achieved significant improvements in phase-SNR (8.2-9.8 dB) on experimental biological cells (cervical and red blood cells).
• Showed enhanced robustness against perturbations in imaging parameters like propagation distance and wavelength.

Conclusions:

• HDPhysNet effectively combines the strengths of trained and untrained deep learning models for superior phase recovery in LDIHM.
• The proposed method offers improved accuracy and robustness, particularly for complex biological imaging.
• LDIHM integrated with HDPhysNet presents a promising, portable microscopy solution for point-of-care cytology.