Physics-informed single-frame end-to-end learning for denoising and background removal in fluorescence imaging.

Abstract

Fluorescence microscopy is often limited by low signal-to-noise ratios arising from the finite number of detected photons and by an out-of-focus background that obscures structural details. Here, we introduce a physics-informed end-to-end learning framework that enables denoising and background suppression from a single fluorescence image. By integrating a microscope-parameterized forward model, the proposed framework generates realistic training datasets entirely through simulation, eliminating the need for experimentally acquired ground-truth images. Once trained for a given imaging condition, the deep neural network can be deployed directly to previously unseen biological specimens without specimen-specific retraining, as demonstrated across the structurally distinct biological samples examined in this study. We experimentally validate the robust single-frame denoising performance in both wide-field and confocal fluorescence microscopy. The enhanced photon efficiency further enables superresolution optical fluctuation imaging using only tens of frames, substantially improving temporal resolution while preserving spatial fidelity. In addition, the proposed single-frame end-to-end learning framework can be extended to remove out-of-focus background in thick samples. These results establish physics-guided, simulation-based end-to-end learning as a general and practical strategy for rapid, data-efficient fluorescence image restoration under photon-limited conditions.