Abstract

Progress in nearly every scientific discipline is hindered by the presence of independent noise in spatiotemporally structured datasets. Three widespread technologies for measuring neural activity--calcium imaging, extracellular electrophysiology, and fMRI--all operate in domains in which shot noise and/or thermal noise deteriorate the quality of measured physiological signals. Current denoising approaches sacrifice spatial and/or temporal resolution to increase the Signal-to-Noise Ratio of weak neuronal events, leading to missed opportunities for scientific discovery. Here, we introduce DeepInterpolation, a general-purpose denoising algorithm that trains a spatio-temporal nonlinear interpolation model using only noisy samples from the original raw data. Applying DeepInterpolation to in vivo two-photon Ca2+ imaging yields up to 6 times more segmented neuronal segments with a 15 fold increase in single pixel SNR, uncovering network dynamics at the single-trial level. In extracellular electrophysiology recordings, DeepInterpolation recovered 25% more high-quality spiking units compared to a standard data analysis pipeline. On fMRI datasets, DeepInterpolation increased the SNR of individual voxels 1.6-fold. All these improvements were attained without sacrificing spatial or temporal resolution. DeepInterpolation could well have a similar impact in other domains for which independent noise is present in experimental data.