Halo-free quantitative phase imaging via physics-constrained self-supervised network

Area of Science:

• Optical microscopy
• Biomedical imaging
• Computational imaging

Background:

• The halo effect in white-light diffraction phase microscopy limits accuracy in label-free biological observations.
• Limited spatial coherence and system noise contribute to this significant artifact.

Purpose of the Study:

• To propose and evaluate a novel physics-constrained halo-free self-supervised network (PC-HFSSN).
• To suppress the halo effect in partially coherent imaging systems.
• To achieve this without requiring inaccessible ground-truth phase labels.

Main Methods:

• Embedding a differentiable forward optical model into the network's loss function.
• Integrating physical spatial smoothness constraints and a Gaussian illumination approximation.
• Training a lightweight architecture on multi-scale synthetic data to isolate and remove halo fields.

Main Results:

• PC-HFSSN demonstrated robust simulation-to-reality generalization on diverse sample data.
• Experimental validation confirmed effective halo effect elimination in simulations and live cells.
• High-frequency phase accuracy and computational efficiency were maintained.

Conclusions:

• PC-HFSSN offers a robust, physically interpretable, and self-supervised solution for quantitative phase imaging.
• The method balances high fidelity and processing speed.
• It enables high-throughput imaging without the need for paired training data.