Computational interference microscopy enabled by deep learning.

Area of Science:

• Biomedical Optics
• Quantitative Phase Imaging
• Computational Microscopy

Background:

• Quantitative phase imaging (QPI) is crucial for cell and tissue characterization.
• Spatial light interference microscopy (SLIM) offers high sensitivity but slow acquisition.
• Diffraction phase microscopy (DPM) provides single-shot imaging but suffers from noise.

Purpose of the Study:

• To develop a deep learning method for generating SLIM-quality phase maps from single-shot DPM images.
• To overcome the limitations of DPM's noise and SLIM's slow acquisition rate.
• To enable real-time, high-sensitivity phase imaging using computational methods.

Main Methods:

• A U-net based deep learning model was trained on over 1000 paired DPM and SLIM images.
• The DPM and SLIM modules were integrated onto an inverted microscope for simultaneous data acquisition.
• Neural network inference was implemented into live acquisition software for real-time processing.

Main Results:

• The deep learning model successfully removed speckles and background noise from DPM images.
• Quantitative metrics showed high fidelity: SNR of 29.97, PCC of 0.79, and SSIM of 0.82.
• Real-time, low-noise phase images were generated from DPM, enhancing image quality significantly.

Conclusions:

• Deep learning effectively translates noisy single-shot DPM images into high-quality, SLIM-equivalent phase maps.
• This computational approach significantly improves phase imaging capabilities for biological samples.
• The method offers a promising pathway for real-time, noise-free quantitative phase imaging in microscopy.