Physics-driven universal twin-image removal network for digital in-line holographic microscopy

Area of Science:

• Biomedical Optics
• Computational Imaging
• Machine Learning Applications

Background:

• Digital in-line holographic microscopy (DIHM) offers cost-effective quantitative phase imaging for cell motility and bio-microfluidics.
• Twin-image noise significantly degrades DIHM reconstruction quality, limiting its applications.
• Existing noise reduction methods often require complex hardware or are computationally intensive with limited efficacy.

Purpose of the Study:

• To develop a novel deep learning solution, UTIRnet, for efficient and universal twin-image suppression in DIHM.
• To provide a fast, robust, and physics-informed deep learning approach for enhancing DIHM image quality.
• To enable widespread adoption of advanced DIHM techniques by minimizing the need for experimental training data.

Main Methods:

• Development of UTIRnet, a deep learning network trained exclusively on numerically generated holographic datasets.
• Implementation of a physics-based constraint to ensure consistency between reconstructed images and input holograms.
• Open-source code release to facilitate integration into diverse DIHM systems.

Main Results:

• UTIRnet demonstrates fast and robust suppression of twin-image noise in DIHM reconstructions.
• The network achieves high performance without requiring extensive experimental training data.
• Reconstruction results show improved reliability and consistency due to the physics-informed approach.

Conclusions:

• UTIRnet offers a significant advancement in DIHM image processing, overcoming limitations of conventional noise reduction techniques.
• The method is broadly applicable across various DIHM setups, accelerating research in cell migration and neurodegenerative diseases.
• The open-source availability and physics-based foundation of UTIRnet promote its practical implementation and reliability.