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

The halo effect arising from limited spatial coherence and system noise severely compromises the accuracy of white-light diffraction phase microscopy, hindering its broader application in high-fidelity label-free biological observations.

We propose and evaluate a physics-constrained halo-free self-supervised network (PC-HFSSN) designed to suppress the halo effect in partially coherent imaging systems without requiring experimentally inaccessible ground-truth phase labels.

Our approach embeds a differentiable forward optical model directly into the network's loss function. By integrating the physical spatial smoothness constraints of halo fields with a Gaussian illumination approximation, our lightweight architecture was trained on multi-scale synthetic data to accurately isolate and remove halo fields from the acquired images.

Trained on diverse sample data, PC-HFSSN demonstrated robust "simulation-to-reality" generalization. Experimental validation of both simulations and dynamic live cells confirmed that the proposed method effectively eliminated the halo effect while preserving high-frequency phase accuracy and maintaining high computational efficiency.

PC-HFSSN provides a robust, physically interpretable, and self-supervised solution for high-throughput quantitative phase imaging by balancing high fidelity and processing speed without the need for paired training data.