Real-time self-supervised denoising for high-speed fluorescence neural imaging

Abstract

Self-supervised denoising methods significantly enhance the signal-to-noise ratio in fluorescence neural imaging, yet real-time solutions remain scarce in high-speed applications. Here, we present the FrAme-multiplexed SpatioTemporal learning strategy (FAST), a deep-learning framework designed for high-speed fluorescence neural imaging, including in vivo calcium, voltage, and volumetric time-lapse imaging. FAST balances spatial and temporal redundancy across neighboring pixels, preserving structural fidelity while preventing over-smoothing of rapidly evolving fluorescence signals. Utilizing an ultra-light convolutional neural network, FAST enables real-time processing at speeds exceeding 1000 frames per second, substantially surpassing the acquisition rates of most high-speed imaging systems. We also introduce an intuitive graphical user interface that integrates FAST into standard imaging workflows, providing a real-time denoising tool for recorded neural activity and enabling downstream analysis in neuroscience research that requires millisecond-scale temporal precision, particularly in closed-loop studies.