Neighbor2Mean: a self-supervised denoising method via local spatial redundancy for live-cell fluorescence microscopy

Live-cell fluorescence microscopy typically requires low excitation doses and short exposure times to minimize phototoxicity, resulting in photon-limited images with low signal-to-noise ratio (SNR). Supervised deep learning improves image quality but requires paired high-SNR ground truth data that are difficult to obtain in dynamic live-cell experiments. Existing self-supervised methods can be broadly classified into temporal-correlation-based and spatial-correlation-based methods. Temporal methods fail in dynamic samples, and spatial methods exclude available information or introduce estimation bias during training, which may limit the recovery of fine subcellular structures. We present Neighbor2Mean (N2M), a self-supervised denoising method that exploits local spatial redundancy to construct low-variance pseudo-targets via local mean aggregation from single noisy acquisitions. Compared with state-of-the-art self-supervised baselines, N2M achieves a 5.59 dB PSNR gain and a 0.16 SSIM improvement over SN2N.