Light-field deep learning enables high-throughput, scattering-mitigated calcium imaging

Area of Science:

• Neuroscience
• Biophysics
• Computational Biology

Background:

• Light-field microscopy (LFM) offers high-throughput volumetric imaging but faces challenges with computational load and light scattering in biological tissues.
• Existing methods struggle to achieve high resolution and signal clarity in scattering environments, limiting deep brain imaging applications.

Purpose of the Study:

• To develop a novel light-field microscopy strategy for volumetric, scattering-mitigated neural circuit activity monitoring.
• To create a deep neural network capable of reconstructing high-quality neural activity volumes from scattered and blurry light-field data.

Main Methods:

• A physics-based deep neural network, 2PiLnet, was trained using two-photon microscopy volumes and one-photon light-field data.
• Light-field videos of jGCaMP8f-expressing neurons in neocortical brain slices were acquired and processed using 2PiLnet.
• The network reconstructs neural activity volumes, achieving two-photon-like contrast and source confinement from degraded light-field inputs.

Main Results:

• 2PiLnet successfully reconstructed neural volumes from scattered, blurry one-photon light fields, yielding high signal-to-noise ratios.
• The method demonstrated reduced optical crosstalk compared to conventional reconstruction techniques.
• High-speed imaging (100 volumes/sec) revealed neural activity, including putative spikes up to 10 Hz, within micron-scale volumes.

Conclusions:

• 2PiLnet provides a scattering-robust method for volumetric neural circuit imaging using light-field microscopy.
• The deep learning approach significantly reduces processing time compared to iterative methods, facilitating real-time analysis.
• This advancement supports closed-loop and adaptive experimental paradigms for studying neural dynamics in scattering tissues.