High-throughput 3D super-resolution imaging in deep tissue

Abstract

Fluorescence images obtained with optical microscopes intrinsically suffer from blur and noise, which can be partially reversed by the deconvolution process. However, the deconvolution process is ill-conditioned, leading to a trade-off between detail preservation and noise suppression. Here, we develop 3D-FUDIP to fully decouple the deconvolution process into two parts: deblurring and denoising, achieving an 8-fold improvement in spatial resolution. By adopting the Poisson model, which obeys the quantum nature of photons, our 3D-FUDIP can be successfully applied to various noise conditions, especially low-light conditions where the photon number is generally extremely small. The results show that our 3D-FUDIP improves the SNR by up to 6-fold with only a one-fifth photon budget. Besides, 3D-FUDIP boosts the spatial bandwidth product (SBP) by one order of magnitude, allowing more spine details to be resolved within a larger imaging volume. By synergizing deep learning with these advances, we propose 3D-FUDIPn to further improve the imaging resolution. We demonstrate 3D-FUDIPs performance in various imaging systems, including confocal, two-photon, and light-sheet microscopes, showing compatibility and potential applications in biological science.