Enhancing ultra-high density single-molecule localization with deep spatiotemporal networks.

Single-molecule localization microscopy (SMLM) is a powerful imaging technique that surpasses the diffraction limit of light by computationally localizing individual fluorescent molecules. However, achieving sufficient spatial resolution in SMLM requires extensive frame acquisition, limiting temporal resolution. Increasing the density of fluorescent molecules is a common strategy to enhance temporal resolution, but this often results in overlapping point spread functions and computational challenges in distinguishing adjacent molecules. We developed a deep learning-driven approach, termed super-resolution spatiotemporal information integration (SRST), for ultra-high-density molecules' precise three-dimensional (3D) localization. SRST leveraged temporal information from adjacent frames and the blinking mechanism to enhance localization accuracy, demonstrating a 10% increase in the Jaccard index and a 14 nm reduction in localization error compared with the state-of-the-art methods in low signal-to-noise ratio conditions. SRST exhibited broad applicability and maintains accurate reconstruction in ultra-high-density scenarios, enhancing structural detail in 3D imaging of subcellular structures such as mitochondria and microtubules while reducing imaging artifacts and improving structural smoothness. SRST will hold substantial promise for detailed structural analysis of cellular components, providing high-resolution imaging with enhanced localization accuracy.