Neural imaging at high speeds produces noisy data. Self-supervised denoising helps, but existing methods are too slow for real-time use. A study in Nature Communications introduces FAST - a framework that cleans neural recordings faster than most cameras can capture them.
The Noise Problem
Fluorescence imaging of neural activity faces a fundamental tradeoff: faster imaging means less light per frame, which means more noise. Calcium imaging, voltage imaging, and volumetric time-lapse imaging all suffer from this constraint.
Deep learning-based denoising can dramatically improve signal quality. But if you want to use denoised data for real-time applications - like closed-loop experiments - the processing can't lag behind the acquisition.
Self-Supervised and Ultra-Fast
FAST (Frame-multiplexed SpatioTemporal learning) uses self-supervised learning, meaning it trains on the noisy data itself without requiring clean ground truth images.
The key innovation is how it balances spatial and temporal information. By carefully leveraging redundancy across both dimensions, FAST preserves structural details while avoiding the over-smoothing that plagues many denoising approaches - particularly important for rapidly changing fluorescence signals.
1,000+ Frames Per Second
Using an ultra-light convolutional neural network, FAST achieves real-time processing speeds exceeding 1,000 frames per second. This is faster than most high-speed imaging systems actually acquire data.
The practical implication: denoising is no longer a bottleneck. It can happen as fast as data comes in.
Immediate Applications
The team provides a graphical user interface that integrates FAST into standard imaging workflows. For closed-loop neuroscience experiments - where neural activity triggers interventions in real time - this removes a key limitation.
Millisecond-scale precision becomes achievable not just for recording, but for responding to neural activity as it happens.
Reference: Li X, et al. (2025). Real-time self-supervised denoising for high-speed fluorescence neural imaging. Nature Communications. doi: 10.1038/s41467-025-64681-8 | PMID: 41136418
Disclaimer: The image accompanying this article is for illustrative purposes only and does not depict actual experimental results, data, or biological mechanisms.