DeepVID v2: self-supervised denoising with decoupled spatiotemporal enhancement for low-photon voltage imaging

Area of Science:

• Neuroscience
• Biophysics
• Computer Science

Background:

• Voltage imaging is crucial for studying neuronal activity but suffers from severe Poisson noise in low-photon conditions.
• Existing self-supervised deep learning methods for denoising voltage imaging data often face a trade-off between spatial and temporal performance.

Purpose of the Study:

• To introduce DeepVID v2, a self-supervised denoising framework designed to overcome the limitations of low-photon voltage imaging.
• To enhance both spatial and temporal resolution in voltage imaging data without requiring ground-truth information.

Main Methods:

• DeepVID v2 builds upon the original DeepVID framework, using sequences of frames for denoising and incorporating blind pixels for local spatial information.
• Introduces a novel spatial prior extraction branch for learning high-resolution spatial details.
• Offers two variants: an online version for real-time inference and an offline version for optimal performance using the full dataset.

Main Results:

• DeepVID v2 successfully overcomes the spatial-temporal performance trade-off, achieving superior denoising.
• Demonstrates enhanced capability in resolving high-resolution spatial structures and rapid temporal neuronal activities.
• Shows generalization across various signal-to-noise ratios and extreme low-photon conditions.

Conclusions:

• DeepVID v2 is a powerful tool for improving voltage imaging quality.
• The framework shows potential for application to other low-photon imaging modalities.
• Facilitates advanced studies of neuronal activity in the brain.