Robust self-supervised denoising of voltage imaging data using CellMincer.

Area of Science:

• Neuroscience
• Computational Biology
• Signal Processing

Background:

• Voltage imaging offers high-throughput neuronal activity analysis but suffers from low signal-to-noise ratio (SNR).
• Existing denoising methods, including deep learning, fail to fully capture temporal dynamics and spatiotemporal dependencies in voltage imaging data.
• Limitations of conventional and current deep learning denoising techniques necessitate novel approaches.

Approach:

• Introduced CellMincer, a self-supervised deep learning method for denoising voltage imaging data.
• Employed a masking and prediction strategy across sparse pixels within short temporal windows.
• Conditioned the denoiser on precomputed spatiotemporal auto-correlations to model long-range dependencies efficiently.

Key Points:

• CellMincer achieves a 3-fold noise reduction by conditioning on spatiotemporal auto-correlations.
• Demonstrated state-of-the-art performance on simulated and real voltage imaging datasets.
• Significantly improved detection of subthreshold events and cross-correlation with electrophysiology ground truth.

Conclusions:

• CellMincer enhances neuronal segmentation, peak detection, and functional phenotype separation in voltage imaging analysis.
• The method effectively models complex spatiotemporal dependencies without requiring large temporal contexts.
• Provides a powerful tool for advancing high-throughput neuronal activity investigation.