Physics-embedded neural state-space module for single-shot lens-free on-chip microscopy

Area of Science:

• Optics and Photonics
• Biomedical Imaging
• Artificial Intelligence in Microscopy

Background:

• Lens-free on-chip microscopy (LFOCM) offers high-throughput imaging without bulky optics.
• Conventional LFOCM reconstruction methods amplify noise due to reliance on physics-based priors.
• There is a need for improved reconstruction techniques to enhance image fidelity and reduce artifacts.

Purpose of the Study:

• To develop a novel physics-embedded neural state-space framework for single-shot LFOCM reconstruction.
• To address the limitations of conventional methods in noise amplification and artifact generation.
• To achieve high-fidelity quantitative amplitude and phase recovery in LFOCM.

Main Methods:

• Developed a physics-embedded neural state-space framework incorporating diffraction physics.
• Utilized a dual-branch neural architecture combining local and global modeling.
• Employed an unsupervised learning approach for image reconstruction.

Main Results:

• Achieved state-of-the-art noise robustness and reconstruction fidelity.
• Demonstrated simultaneous suppression of artifacts and preservation of resolution.
• Successfully recovered high-fidelity quantitative amplitude and phase distributions.

Conclusions:

• The proposed physics-constrained framework significantly enhances LFOCM image quality.
• This approach overcomes noise amplification issues inherent in traditional reconstruction pipelines.
• The method enables advanced quantitative imaging with improved accuracy and detail.