Combined learning for augmented super-resolution imaging

Structured illumination microscopy (SIM) is prone to reconstruction artifacts under low-light conditions. While deep learning-based strategies have been employed to restore super-resolved images, these methods often result in excessive smoothing, compromising the resolution of fine subcellular structures. Here, we present Combined Learning Augmented Super-resolution Imaging (CLASI), which employs a two-stage probabilistic diffusion model framework for super-resolution image restoration across different imaging modalities. CLASI integrates the high-fidelity restoration capabilities of supervised deep learning (SDL) models with the ultra-fine structure generation strengths of generative deep learning (GDL) models. CLASI first utilizes the SwinIR model for initial image restoration, followed by fine subcellular structure reconstruction leveraging the generative priors of a pre-trained stable diffusion model. We validated CLASI across various cellular structures and imaging modalities, demonstrating its ability to achieve resolution comparable to ground truth SIM (GT-SIM) images, even at fluorescence intensities 20 times lower than standard conditions, while preserving the authenticity of the recovered details. CLASI enabled long-term, super-resolution, multi-color imaging of light-sensitive processes in living cells, revealing the directional movement of lysosomes along microtubules, the hitchhiking remodeling mechanism of microtubules driven by lysosomal traction, and the rapid fission and fusion dynamics of mitochondria interacting with the endoplasmic reticulum.