Subpixel-registered dual-mode imaging enables label-free inference of mitochondria in living cells

Abstract

Fluorescence microscopy remains indispensable for specific organelle imaging but suffers from photobleaching and phototoxicity. Here, we introduce a strong physics-constrained deep learning strategy to generate virtual fluorescence images of mitochondria directly from quantitative phase imaging. By constructing a dual-mode system that captures quantitative phase and fluorescence images in situ with subpixel registration, we impose a strong physical prior that ensures native subpixel alignment and data fidelity. This native spatial constraint significantly reduces the burden on the neural network, enabling high-confidence, label-free identification of mitochondria from phase data alone. Once trained, the model bypasses the need for fluorescent labeling, eliminating photodamage and facilitating long-term dynamic studies. The trained network exhibits remarkable generalization capability, making it highly practical for routine use and paving the way for truly nondestructive, high-content imaging of subcellular structures in living cells.