Bio-friendly and high-precision super-resolution imaging through self-supervised reconstruction structured illumination microscopy.

Deep-learning-based structured illumination microscopy (SIM) has demonstrated substantial potential in long-term super-resolution imaging of biostructures, enabling the study of subcellular dynamics and interactions in live cells. However, the acquisition of ground-truth (GT) data for training poses inherent challenges, limiting its universal applicability. Current approaches without using GT training data compromise reconstruction fidelity and resolution, and the lack of physical priors in end-to-end networks further limits these qualities. Here we developed self-supervised reconstruction (SSR)-SIM by combining statistical analysis of reconstruction artifacts with structured light modulation priors to eliminate the need for GT and improve reconstruction precision. We validated SSR-SIM on common biological datasets and demonstrated that SSR-SIM enabled long-term recording of dynamic events, including cytoskeletal remodeling in cell adhesion, mitochondrial cristae remodeling, interactions between viral glycoprotein and endoplasmic reticulum, endocytic recycling of transferrin receptors, vaccinia-virus-induced actin comet remodeling, and mitochondrial intercellular transfer through tunneling nanotubes.