Time-lapse RESOLFT nanoscopy empowered by deep learning

Abstract

Capturing the dynamic interplay of proteins and organelles within cells requires microscopy methods with high spatial and temporal resolution. Reversible Saturable OpticaL Fluorescent Transition (RESOLFT) demonstrated great potential for live cell nanoscopy by using patterned light and reversibly switching fluorescent probes. Despite the latest improvement in both switching probes and optical schemes, the duration and frame rate of the time-lapse are still limited by photobleaching and spatial sampling steps. Here, we present two distinct strategies based on deep learning to restore low signal-to-noise-ratio (SNR) or under-sampled images pushing the experimental boundaries of RESOLFT. We demonstrate the capability of Hierarchical DivNoising (HDN) to restore low-SNR RESOLFT data while maintaining full spatial resolution, both in supervised and unsupervised settings. This allows a 10-fold reduction of light dose per frame, and up to 5 times longer imaging. Then, we present Fast-MoNaLISA, a method that reduces the acquisition time of any point-scanning systems. This allows a 4-fold increase in frame rate, and a reduction of scanning artefacts. Overall, our methods open up possibilities for dynamic data acquisition at the sub-organelle level, such as vesicles budding in and out from intact mitochondria, or the actin meshwork formation and dissociation. The authors use two deep-learning methods for RESOLFT nanoscopy: HDN denoising enables five-fold longer live-cell imaging, while spatial upsampling with temporal context restores subsampled data for fourfold higher frame rates and reduced artifacts.