New technologies for improving the accuracy of neuroimaging: Stochastic Gaussian-masked denoiser enhances subcellular structures and neural dynamics

JOURNAL/atin/04.03/02274269-202601000-00003/figure1/v/2026-02-02T122432Z/r/image-tiff Objectives: Fluorescence microscopy is inherently susceptible to acquisition noise, which obscures critical subcellular structures and crucial neural dynamics. To address this, we introduce the stochastic Gaussian-masked autoencoder (GaMA), a zero-shot denoising framework that leverages the intrinsic characteristics of microscopy noise. Methods: GaMA strategically applies stochastic Gaussian masks, statistically conjugate to the wide-band noise, to induce antipodal signal cancellation while preserving genuine biological structures. This approach exploits the physical correspondence between mask and noise distributions for effective noise nullification, eliminating the need for training data. Results: Validated on diverse imaging modalities, GaMA robustly enhances super-resolution single-molecule localization microscopy (e.g., enabling precise microtubule reconstruction) and critically, faithfully recovers subtle neural dynamics in functional Drosophila whole brain calcium imaging. Operating at > 45 frames per second, GaMA facilitates real-time denoising of dynamic neural processes otherwise compromised by noise, significantly enhancing the accuracy of downstream quantitative analysis in neuroimaging applications. Conclusion: Our proposed GaMA establishes a new physics-based paradigm for microscopic image denoising by leveraging the inherent statistical properties of microscope noise through statistical conjugation of Gaussian-masked techniques. This approach significantly enhances the accuracy of downstream quantitative analysis in neuroimaging applications.