Physics-driven self-supervised learning for quantitative high-fidelity structured illumination microscopy

Abstract

Structured illumination microscopy (SIM) enables rapid, long-term super-resolution (SR) imaging of live-cell dynamics. However, although current state-of-the-art (SOTA) SIM reconstruction methods achieve high-fidelity structural SR, they consistently lack reliable intensity quantification, restricting their use in quantitative biology. Here, we develop qHiFi-SIM, a physics-driven self-supervised learning framework for quantitative high-fidelity SR-SIM imaging. By leveraging the wide-field image as a physical intensity reference, our approach enables self-supervised training without reliance on SR data with ground-truth intensity. qHiFi-SIM achieves high structural fidelity (structural similarity, SSIM > 0.95) with a twofold resolution enhancement, while maintaining excellent intensity linearity (coefficient of determination, R{superscript 2} > 0.99). It also exhibits strong transferability across diverse SIM setups and typical samples, and is compatible with SOTA SIM algorithms, enabling direct quantitative correction of their intensity deviations without retraining. We demonstrate the unique advantages of qHiFi-SIM for live-cell quantitative visualization of mitochondrial structure, intensity, and membrane potential dynamics, as well as for high-fidelity SIM-FRET (Forster resonance energy transfer) functional imaging. We anticipate that qHiFi-SIM will serve as a practical tool for SR structural visualization and quantitative functional imaging in live cells.