Towards ultrafast quantitative phase imaging via differentiable microscopy [Invited]

Area of Science:

• Biomedical Optics
• Microscopy Techniques
• Computational Imaging

Background:

• Quantitative phase microscopy (QPM) is a label-free imaging technique with broad applications in metabolomics and histopathology.
• Current QPM throughput is limited by image sensor pixel rates, hindering real-time analysis and high-volume studies.
• Advancements in sensors and deep learning have improved QPM but not overcome the fundamental hardware bottleneck.

Purpose of the Study:

• To develop a novel framework for compressed image acquisition in QPM to overcome hardware limitations.
• To introduce a learnable optical compression-decompression system for enhanced QPM throughput.
• To demonstrate significant throughput improvements in QPM through content-specific feature learning.

Main Methods:

• Numerical simulation of a differentiable quantitative phase microscopy (∂-QPM) system.
• Implementation of learnable optical processors for image compression at the hardware level.
• Utilizing a reconstruction neural network for post-acquisition image decompression.

Main Results:

• Achieved a compression factor of ×64 in numerical experiments.
• Maintained high image fidelity with a Structural Similarity Index (SSIM) of ∼0.90.
• Preserved image quality with a Peak Signal-to-Noise Ratio (PSNR) of ∼30 dB for cellular imaging.

Conclusions:

• The proposed ∂-QPM framework offers a new pathway to significantly enhance QPM system throughput.
• Learned optical compression effectively bypasses sensor pixel-rate limitations.
• This approach promises unprecedented improvements for label-free, high-throughput microscopy.