Self-inspired learning to denoise for live-cell super-resolution microscopy

Abstract

Every collected photon is precious in live-cell super-resolution (SR) fluorescence microscopy for contributing to breaking the diffraction limit with the preservation of temporal resolvability. Here, to maximize the utilization of accumulated photons, we propose SN2N, a Self-inspired Noise2Noise engine with self-supervised data generation and self-constrained learning process, which is an effective and data-efficient learning-based denoising solution for high-quality SR imaging in general. Through simulations and experiments, we show that the SN2Ns performance is fully competitive to the supervised learning methods but circumventing the need for large training-set and clean ground-truth, in which a single noisy frame is feasible for training. By one-to-two orders of magnitude increased photon efficiency, the direct applications on various confocal-based SR systems highlight the versatility of SN2N for allowing fast and gentle 5D SR imaging. We also integrated SN2N into the prevailing SR reconstructions for artifacts removal, enabling efficient reconstructions from limited photons. Together, we anticipate our SN2N and its integrations could inspire further advances in the rapidly developing field of fluorescence imaging and benefit subsequent precise structure segmentation irrespective of noise conditions.