Untrained deep learning-based differential phase-contrast microscopy.

Area of Science:

• Optics and Photonics
• Biomedical Imaging
• Computational Microscopy

Background:

• Quantitative differential phase-contrast (DPC) microscopy reconstructs phase images from intensity data.
• Traditional DPC relies on linearized models, limiting imaging scope and requiring aberration correction.
• Existing methods necessitate extensive measurements and complex algorithms for accurate phase reconstruction.

Purpose of the Study:

• To develop a self-calibrated DPC microscopy technique overcoming limitations of current methods.
• To incorporate nonlinear image formation models for broader object applicability.
• To simultaneously reconstruct object phase information and system aberrations without prior training data.

Main Methods:

• Implementation of a self-calibrated DPC microscope utilizing an untrained neural network (UNN).
• Integration of the nonlinear image formation model within the UNN framework.
• Validation through numerical simulations and experimental trials using an LED microscope.

Main Results:

• The UNN-DPC method successfully reconstructs complex object information and system aberrations.
• The approach alleviates restrictions on the types of objects that can be imaged.
• Demonstrated viability in both simulated environments and real-world experimental setups.

Conclusions:

• Self-calibrated UNN-DPC microscopy offers a robust and versatile phase imaging solution.
• This method eliminates the need for training datasets, simplifying the imaging process.
• The technique holds significant potential for advancing transparent object imaging in various scientific fields.