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

Area of Science:

• Neuroscience
• Biophysics
• Computational Biology

Background:

• Voltage imaging is crucial for neuroscience but limited by low signal-to-noise ratios (SNR).
• Existing denoising methods struggle with the complex dynamics of voltage imaging data.
• Deep learning approaches have not fully addressed the specific challenges of this technique.

Purpose of the Study:

• Introduce CellMincer, a novel self-supervised deep learning method for denoising voltage imaging data.
• Improve SNR and enhance the detection of neuronal activity.
• Provide a robust tool for analyzing complex neural dynamics.

Main Methods:

• Developed CellMincer, a self-supervised deep learning model using masked pixel prediction.
• Conditioned the denoiser on spatiotemporal auto-correlations to capture long-range dependencies.
• Utilized a physics-based simulation framework for data generation and model optimization.

Main Results:

• CellMincer achieved a 3-fold SNR gain through spatiotemporal auto-correlation conditioning.
• Demonstrated state-of-the-art performance on simulated and real datasets, outperforming existing methods.
• Showcased significant improvements in neuronal segmentation, peak detection, and functional phenotype identification.

Conclusions:

• CellMincer offers substantial noise reduction and improved signal fidelity in voltage imaging.
• The method consistently enhances SNR by 0.5-2.9 dB and reduces variability by 17-55%.
• CellMincer represents a significant advancement for analyzing neuronal activity with voltage imaging.