Enabling Real-Time Fluctuation-Based Super ResolutionImaging

Abstract

Live-cell imaging captures dynamic cellular processes, yet many structures remain beyond the diffraction limit. Fluctuation-based super-resolution techniques overcome this limit by exploiting correlations in fluorescence blinking, but they typically require hundreds of frames and computationally intensive post-processing, prohibiting real-time imaging of fast cellular events. Recent deep learning approaches aim to increase the temporal resolution; however, many rely on extensive pre-processing or large, complex models that increase training cost and inference latency, preventing real-time deployment. To address this, we employ a light-weight recurrent neural network model, which integrates sequential low-resolution frames to extract spatio-temporally correlated signals. Our method is taylored for live-cell imaging under extreme signal-to-noise ratio conditions. It significantly improves temporal resolution by reducing the required number of frames down to as few as 8 frames while doubling the spatial resolution in an inference time below 30 ms. By combining simulation based training with an efficient network architecture, we introduce RESURF, a deep-learning based real-time super-resolution fluctuation imaging framework. We demonstrate that RESURF generalizes across different biological structures and can be readily adapted to various microscope setups using transfer learning. The accompanying dataset, comprising simulations and experiments across multiple subcellular structures and labeling strategies, establishes a benchmarking platform for fluctuation-based super-resolution techniques. RESURF offers a practical, low-latency deep-learning framework for high-throughput and real-time live-cell super-resolution imaging.