Comparing synthetic refocusing to deconvolution for the extraction of neuronal calcium transients from light-fields

Carmel L. Howe, Peter Quicke, Pingfan Song, Herman Verinaz-Jadan, Pier Luigi Dragotti, Amanda J. FoustbioRxiv 2020 · DOI

TL;DR

Light-field microscopy (LFM) offers a fast, volumetric way to image neuronal activity. This paper compares two LFM reconstruction methods – synthetic refocusing and 3D deconvolution – finding that deconvolution provides better spatial resolution for calcium transients, though at the cost of increased computation and potential noise amplification.

What Problem Does This Paper Address?

Understanding complex neural networks requires observing the activity of thousands of neurons simultaneously in 3D. Traditional methods like confocal or multi-photon microscopy are limited by scanning speed, impacting acquisition rate and signal-to-noise ratio (SNR). While widefield imaging is fast, it lacks optical sectioning, leading to out-of-focus fluorescence that degrades image quality. Light-field microscopy (LFM) offers a solution by capturing volumetric data in a single frame, but the best way to reconstruct these 4D light-field images into usable 3D volumes for analyzing neuronal calcium transients is still an open question. Specifically, this paper investigates how two common LFM reconstruction methods, synthetic refocusing and 3D deconvolution, perform in terms of signal-to-noise ratio (SNR) and spatial confinement of calcium signals.

Methods

The researchers used two main LFM reconstruction techniques:

  • Synthetic Refocusing: This method computationally extracts individual planes from the light-field data, essentially creating a z-stack. It's described as computationally fast, but lacks optical sectioning, meaning out-of-focus light can still contaminate the signal.
  • 3D Deconvolution (Richardson-Lucy): This approach uses a more complex algorithm to reconstruct a 3D volume by deconvolving the light-field measurements with a calculated Point Spread Function (PSF) that models the LFM system. This method is expected to offer better spatial resolution due to its ability to "sharpen" the image. They also explored variations of this method, including:
  • Iterative Deconvolution: Increasing the number of iterations in the Richardson-Lucy algorithm was investigated to see its effect on signal confinement and SNR.
  • Total Variation (TV) Regularization: To combat noise amplification inherent in deconvolution, they incorporated a TV term into the Richardson-Lucy algorithm, aiming to preserve edges while reducing noise.

They acquired light-field image series from acute mouse brain slices, using both bulk-labeled neurons and single neurons loaded intracellularly with the red-emitting calcium dye CaSiR-1. They then compared the temporal SNR (tSNR) and spatial confinement of the extracted calcium signals from volumes reconstructed by these different methods, as well as from interleaved widefield images.

Datasets

The study used light-field image series acquired from acute mouse brain slices. These slices were prepared from mice aged 33 to 196 days. The calcium dye CaSiR-1 was used for labeling, either through bulk application or single-cell intracellular loading. No new datasets were introduced; the data was generated specifically for this study.

Key Insights and Ideas

  • Deconvolution for Resolution, Refocusing for Speed: The core trade-off is clear: 3D deconvolution offers superior spatial confinement (sharper signals), crucial for resolving fine neuronal structures, while synthetic refocusing is much faster.
  • Iteration Number Matters (for Deconvolution): For deconvolution, increasing the number of iterations initially boosted peak signal but eventually led to increased noise and reduced SNR. There's a sweet spot for iteration count to balance signal enhancement and noise.
  • No SNR Penalty for LFM vs. Widefield (at the focal plane): When comparing LFM signals reconstructed at the native focal plane to interleaved widefield images, there was no significant loss in tSNR for LFM. This is a key finding, as it means you get volumetric capabilities without sacrificing signal quality at the plane of interest.
  • Dendrites are Resolvable in 3D: LFM, particularly with deconvolution, successfully resolved calcium transients in neuronal dendrites across different depths, demonstrating its utility for studying signal propagation in complex neuronal structures.
  • TV Regularization Didn't Help (in this case): Surprisingly, adding Total Variation regularization to the deconvolution algorithm didn't improve signal, noise, or SNR for extracted calcium time series, suggesting that for this specific application, it might not be the optimal approach.

Key Results

  • Spatial Confinement: 3-iteration Richardson-Lucy deconvolution significantly improved axial signal confinement compared to synthetic refocusing and widefield stacks. Lateral confinement also improved with deconvolution, plateauing around 7 iterations.
  • SNR: For single-cell experiments, 3-iteration RL deconvolution yielded significantly higher SNR than synthetic refocusing or widefield. However, for bulk-labeled slices, the SNR was comparable between synthetic refocusing and deconvolution.
  • Peak Signal: Deconvolution generally led to a higher peak signal, especially in single-cell trials, with improvements up to 2x compared to synthetic refocusing.
  • Noise: While deconvolution increased peak signal, it also amplified noise, particularly at higher iteration numbers.
  • Computational Cost: Reconstructing a volume with synthetic refocusing was significantly faster (40 seconds/frame) than 3-iteration RL deconvolution (20 seconds/iteration/frame). For a typical time series, synthetic refocusing took ~2 hours, while 3-iteration RL took ~3.5 hours. Higher iteration deconvolution could take up to 22 hours.

Limitations and Future Work

  • Scattering Limitations: Both reconstruction methods rely on ballistic photons, limiting their application in highly scattering brain tissue to depths of about 50 microns. The authors suggest using red-shifted dyes (like CaSiR-1) to mitigate scattering.
  • Computational Cost of Deconvolution: While deconvolution offers superior spatial confinement, its computational intensity is a significant drawback, especially for applications requiring real-time analysis. The authors note that deep learning solutions are being explored to speed up reconstruction.
  • TV Regularization Ineffectiveness: The TV regularization did not yield benefits in this study, suggesting further investigation into regularization techniques or alternative approaches might be needed.
  • Future Directions: The authors highlight the importance of choosing a reconstruction method based on specific experimental goals, balancing SNR, spatial confinement, and computational cost. They suggest that applications needing online analysis might favor faster methods like synthetic refocusing or low-iteration deconvolution, while those prioritizing spatial detail might opt for higher-iteration deconvolution despite the computational burden. They also point to the potential of LFM for studying dendritic calcium signals and its broader applicability to genetically encoded indicators.

Summary updated

Abstract

Abstract Significance Light-field microscopy (LFM) enables fast, light-efficient, volumetric imaging of neuronal activity with calcium indicators. Calcium transients differ in temporal signal-to-noise ratio (tSNR) and spatial confinement when extracted from volumes reconstructed by different algorithms. Aim We evaluated the capabilities and limitations of two light-field reconstruction algorithms for calcium fluorescence imaging. Approach We acquired light-field image series from neurons either bulk-labeled or filled intracellularly with the red-emitting calcium dye CaSiR-1 in acute mouse brain slices. We compared the tSNR and spatial confinement of calcium signals extracted from volumes reconstructed with synthetic refocusing and Richardson-Lucy 3D deconvolution with and without total variation regularization. Results Both synthetic refocusing and Richardson-Lucy deconvolution resolved calcium signals from single cells and neuronal dendrites in three dimensions. Increasing deconvolution iteration number improved spatial confinement but reduced tSNR compared to synthetic refocusing. Volumetric light-field imaging did not decrease calcium signal tSNR compared to interleaved, widefield image series acquired in matched planes. Conclusions LFM enables high-volume rate, volumetric imaging of calcium transients in single cells (bulk-labeled), somata and dendrites (intracellular loaded). The trade-offs identified for tSNR, spatial confinement, and computational cost indicate which of synthetic refocusing or deconvolution can better realize the scientific requirements of future LFM calcium imaging applications.
BibTeX
@misc{howe2020comparing,
  author = {Carmel L. Howe and Peter Quicke and Pingfan Song and Herman Verinaz-Jadan and Pier Luigi Dragotti and Amanda J. Foust},
  title = {{Comparing synthetic refocusing to deconvolution for the extraction of neuronal calcium transients from light-fields}},
  year = {2020},
  doi = {10.1101/2020.09.07.285585},
  howpublished = {bioRxiv},
}
Figure