Self-supervised neural network for phase retrieval in QDPC microscopy.

Area of Science:

• Biomedical Optics
• Microscopy
• Computational Imaging

Background:

• Quantitative differential phase contrast (QDPC) microscopy offers high-resolution, label-free imaging for transparent biological samples.
• Traditional QDPC methods rely on the weak phase assumption and manual Tikhonov regularization, limiting their applicability and convenience.
• The weak phase assumption restricts analysis to thin specimens, and manual parameter tuning is cumbersome.

Purpose of the Study:

• To develop a self-supervised learning method for accurate quantitative phase retrieval in QDPC microscopy.
• To overcome the limitations of the weak phase assumption and manual regularization in phase reconstruction.
• To enable precise phase prediction directly from intensity measurements without ground truth data.

Main Methods:

• A deep image prior (DIP) based self-supervised learning framework was employed for phase retrieval.
• The DIP model was trained using synthesized intensity measurements derived from predicted phase images.
• A physical layer was integrated to generate intensity measurements, guiding the DIP model's reconstruction process by minimizing intensity differences.

Main Results:

• The proposed method successfully reconstructed phase information from intensity measurements in phantom studies.
• Phase values for micro-lens arrays and standard phase targets showed deviations of less than 10% compared to theoretical values.
• The method demonstrated high accuracy in predicting quantitative phase without requiring ground truth phase data.

Conclusions:

• The self-supervised DIP method is feasible for accurate quantitative phase prediction in QDPC microscopy.
• This approach eliminates the need for ground truth phase data, simplifying the imaging workflow.
• The findings highlight the potential of deep learning for advancing quantitative phase imaging in biomedical research.