High-fidelity single-frame computational super-resolution using signal-preserving denoising-enabled deconvolution.

Abstract

Computational super-resolution (SR) methods enable nanoscale imaging from single-frame wide-field or spinning-disk confocal images without hardware modifications, yet face limitations: statistical restoration suffers from noise and artifacts, while deep learning methods typically lack generalizability. We introduce 3Snet-CLID, a computational SR method which integrates a hybrid supervised/self-supervised deep learning network for signal-preserving denoising with direct Richardson-Lucy deconvolution. 3Snet-CLID's per-pixel denoising strategy suppresses noise while maintaining signal distribution, mitigating artifacts, and enhancing robustness. The method achieves more than 5-fold resolution improvement on conventional microscopes, revealing diverse structures such as the mitochondrial outer membrane, endoplasmic reticulum, and nuclear pores in live and fixed cells under standard labeling. By overcoming key computational SR bottlenecks, 3Snet-CLID offers denoising capability and an accessible platform for high-fidelity nanoscale live-cell imaging.