AN EXTENSION OF THE FRI FRAMEWORK FOR CALCIUM TRANSIENT DETECTION

Stephanie Reynolds, Caroline S. Copeland, Simon R. Schultz, Pier Luigi Dragotti

TL;DR

This paper extends a powerful signal processing framework (Finite Rate of Innovation - FRI) to more accurately detect neuronal spiking activity from calcium imaging data, particularly for newer, slower indicators like GCaMP6s. By incorporating noise reduction and robust model-order estimation, they achieve a significant improvement in spike detection accuracy, reaching 90% for electrophysiologically validated spikes.

What Problem Does This Paper Address?

Neuroscience relies heavily on imaging techniques to understand brain activity. Two-photon calcium imaging allows researchers to visualize the activity of individual neurons in real-time. However, the raw imaging data is noisy and indirect: the fluorescence signal changes reflect intracellular calcium levels, which are a consequence of neuronal "spikes" (action potentials). To truly understand neural circuits, we need to precisely infer when these spikes occurred from the noisy calcium signals. Previous methods, often based on Finite Rate of Innovation (FRI) theory, worked well for calcium indicators with fast responses. However, newer genetically encoded indicators like GCaMP6s have a slower rise time, which older algorithms struggled to handle accurately, leading to missed spikes or false positives. This paper aims to bridge this gap by adapting the FRI framework to accommodate these slower, more biologically relevant calcium transients.

Methods

The core of this work is an extension of the Finite Rate of Innovation (FRI) framework. FRI theory is well-suited for signals that can be characterized by a limited number of parameters over time, like a series of pulses.

Here's a breakdown of their approach:

  • Generalized Pulse Shape: The authors first generalize the mathematical model for the calcium transient pulse shape. Instead of assuming an instantaneous rise and exponential decay (suitable for older indicators), they model a more realistic shape with a slower, non-instantaneous rise followed by an exponential decay. This is crucial for accurately representing GCaMP6s dynamics.
  • Filtering Scheme: They derive a specific filtering scheme that transforms the problem of detecting these generalized calcium transients into a classical FRI problem of identifying the locations of a stream of "Diracs" (idealized impulses).
  • Noise Reduction (Pre-whitening): Real calcium imaging data is inherently noisy. To combat this, they implement a "pre-whitening" step. This process effectively colors the noise in a way that makes it "white" (uncorrelated), significantly improving the robustness of subsequent estimation algorithms.
  • Least Squares Model-Order Estimation: Determining the exact number of spikes within a given time window can be challenging. They employ a least squares approach to estimate the optimal "model order" (i.e., the number of spikes) by minimizing the error between the observed data and the reconstructed signal for different model orders. This is applied sequentially in a sliding window.
  • Matrix Pencil Method: This is a standard technique for estimating parameters (like time delays) from noisy data that can be represented by a structured matrix (a Toeplitz matrix in this case). Their pre-whitening step makes this method more reliable.

Datasets

The primary dataset used for evaluation is from the GENIE project, which provides simultaneous calcium imaging (GCaMP6s) and electrophysiological recordings from in vivo V1 neurons. This allows for a direct comparison between the algorithm's spike estimates and the ground truth from electrophysiology. The paper mentions using 3 traces from this dataset, totaling 678 seconds of data and containing 532 spikes.

Key Insights and Ideas

  • GCaMP6s Dynamics Matter: The paper strongly emphasizes that the specific kinetics of calcium indicators, particularly the rise time, are critical. Algorithms designed for fast indicators perform poorly on slower ones like GCaMP6s, highlighting the need for indicator-specific signal processing.
  • FRI is Adaptable: The Finite Rate of Innovation framework, known for its theoretical underpinnings in signal reconstruction, proves to be a flexible and powerful tool when its underlying signal models are adapted to match biological reality.
  • Noise Management is Key: The "pre-whitening" step is not just a minor tweak; it's presented as a crucial component for making the FRI framework robust enough to handle the inherent noise in biological imaging data.
  • Model Order Estimation Improves Reliability: Explicitly estimating the number of spikes in a window, rather than assuming a fixed number, leads to more accurate and reliable spike train reconstruction.

Key Results

The modified FRI algorithm demonstrates significantly improved performance on real GCaMP6s data compared to the original FRI algorithm:

  • ROC Curves: Figure 1a shows that the modified FRI algorithm achieves a higher true positive rate for a given false positive rate across the board.
  • Spike Detection Accuracy: On average, the algorithm detects 90% of electrophysiologically-validated spikes within a strict tolerance of 2 sample widths (0.034s) of the true spike time.
  • Comparison to Other Methods: Figure 2 (average ROC curves) shows that the modified FRI algorithm outperforms both the original FRI and a commonly used fast deconvolution algorithm (Vogelstein et al.) on this GCaMP6s dataset. The deconvolution algorithm struggles due to its assumption of uniform fluorescence change per spike, which is not true for GCaMP6s.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that the "strict target" of detecting spikes within 2 sample widths might be challenging for some applications, especially given the ~15 sample width rise time of GCaMP6s transients.
  • While robust, the algorithm's performance is still dependent on the quality of the calcium imaging data and the accuracy of the underlying noise models.
  • Future Work:
    • The authors suggest that this generalized FRI framework could be applied to other calcium indicators with different kinetic properties.
  • Further exploration of noise reduction techniques and model-order estimation strategies could potentially lead to even higher detection rates.
  • Extending the framework to handle simultaneous activity from multiple neurons and more complex network dynamics would be a natural next step.

Summary updated

Abstract

Two-photon calcium imaging of the brain allows the spatiotemporal activity of neuronal networks to be monitored at cellular resolution. In order to analyse this activity it must first be possible to detect, with high temporal resolution, spikes from the time series corresponding to single neurons. Previous work has shown that finite rate of innovation (FRI) theory can be used to reconstruct spike trains from noisy calcium imaging data. In this paper we extend the FRI framework for spike detection from calcium imaging data to encompass data generated by a larger class of calcium indicators, including the genetically encoded indicator GCaMP6s. Furthermore, we implement least squares model-order estimation and perform a noise reduction procedure ('pre-whitening') in order to increase the robustness of the algorithm. We demonstrate high spike detection performance on real data generated by GCaMP6s, detecting 90% of electrophysiologically-validated spikes.
BibTeX
@inproceedings{reynolds2016extension,
  author = {Stephanie Reynolds and Caroline S. Copeland and Simon R. Schultz and Pier Luigi Dragotti},
  title = {{AN EXTENSION OF THE FRI FRAMEWORK FOR CALCIUM TRANSIENT DETECTION}},
  booktitle = {ISBI},
  year = {2016},
  url = {https://openreview.net/forum?id=nQAO8yQf0H},
}
Figure