Removing independent noise in systems neuroscience data using DeepInterpolation.

Area of Science:

• Neuroscience
• Biophysics
• Signal Processing

Background:

• Independent noise, such as shot and thermal noise, significantly hinders scientific progress by overwhelming physiological signals in neural activity measurement technologies.
• Techniques like calcium imaging, extracellular electrophysiology, and functional magnetic resonance imaging (fMRI) are particularly susceptible to noise, limiting the resolution and scope of data analysis.

Purpose of the Study:

• To introduce DeepInterpolation, a versatile denoising algorithm designed to mitigate independent noise in spatiotemporally structured datasets.
• To demonstrate the efficacy of DeepInterpolation across various neural recording modalities, including two-photon calcium imaging, extracellular electrophysiology, and fMRI.

Main Methods:

• DeepInterpolation employs a spatiotemporal nonlinear interpolation model trained exclusively on raw, noisy data samples.
• The algorithm does not require ground truth data for training, making it broadly applicable.

Main Results:

• In two-photon calcium imaging, DeepInterpolation increased neuronal segments by up to sixfold and achieved a 15-fold rise in single-pixel signal-to-noise ratio (SNR), revealing single-trial network dynamics.
• Extracellular electrophysiology recordings processed with DeepInterpolation yielded 25% more high-quality spiking units.
• fMRI datasets processed with DeepInterpolation showed a 1.6-fold increase in the SNR of individual voxels.

Conclusions:

• DeepInterpolation effectively denoises neural data across multiple modalities without compromising spatial or temporal resolution.
• The algorithm's ability to enhance data quality and uncover hidden dynamics holds significant promise for advancing research in neuroscience and other fields affected by independent noise.