Fast, efficient, and accurate neuro-imaging denoising via supervised deep learning.

Area of Science:

• Neuroscience
• Biophysics
• Machine Learning

Background:

• Volumetric functional imaging is crucial for in vivo neuron activity recording.
• Current methods face tradeoffs between calcium trace quality, imaging speed, and laser power.
• Deep learning has shown promise in image denoising but has limitations in downstream neural analysis.

Purpose of the Study:

• To develop a supervised deep-denoising method to overcome tradeoffs in functional imaging.
• To enable high-signal-to-noise ratio (SNR) calcium trace recovery.
• To facilitate faster and long-term in vivo imaging for behavioral neuroscience.

Main Methods:

• A supervised deep-denoising framework was developed.
• The method was trained on small, non-temporally-sequential image datasets.
• The framework was applied to whole-brain, large-field-of-view, and C. elegans neurite imaging.

Main Results:

• The method effectively denoises images and recovers high-SNR calcium traces.
• Achieved fast training and inference times (50-70 ms) with a 30x smaller memory footprint.
• Demonstrated accuracy and generalizability across diverse imaging applications.

Conclusions:

• The developed deep-denoising framework circumvents traditional imaging tradeoffs.
• This approach significantly enhances the feasibility of advanced in vivo neural imaging.
• Enables new possibilities for studying neuronal mechanisms underlying complex behaviors.