A Robust and Versatile Framework to Compare Spike Detection Methods in Calcium Imaging of Neuronal Activity

Abstract

Calcium fluorescence imaging enables real-time activity monitoring of single neurons in living animals. A critical inverse problem resides in the precise inference of spike locations from noisy fluorescence traces, especially in the presence of burst spiking and non-linear fluorescence intensity. Several spike extraction algorithms have been proposed in the recent years, but a robust and objective evaluation of their performance still remains elusive due to the unsupervised nature of the problem. Here we propose a biologically-inspired mathematical framework to reproduce synthetic fluorescence traces from a time-series data of spike-trains. The idea is to create a versatile platform to objectively test the state-ofthe art spike inference methodologies over a large range of experimental parameters. Our solution appears as a complementary but more exhaustive approach to determine the robustness of existing solutions to different nature of signals, imaging artefacts, sensitivity to hyper-parameters and pre-processing steps. We benchmark state-of-the-art algorithms with the proposed simulation platform, and validate the results on an experimental dataset of the Hydra Vulgaris. We hypothesize that, in contrast to the common practice of qualitative evaluation, quantitative measure of algorithm robustness is essential in understanding the suitability of a spike inference algorithm to be used in an automated computational pipeline to decipher the neural code.

Review of data processing of functional optical microscopy for neuroscience

“…As a final note, a form of normalization that often is required in preprocessing is detrending to remove photobleaching 116 (e.g., see Refs. 117 and 118). Photobleaching is the reduction in overall fluorescence over time stemming from the fluorescent proteins becoming trapped in an intermediate quantum state and becoming inactive.…”

Statistical Coupling Between time Point-Processes

“…We use the online dataset from [13], corresponding to two-photon imaging of neuron activity in mouse visual cortex (File M1d1AS in the dataset). From calcium fluorescence traces, the exact spiking times can be obtained using spike inference techniques with variable accuracy and robustness [14]. A representative rasterplot of neuron spiking activity obtained with the constrained FOOPSI deconvolution algorithm [15] is shown in Figure 1-A (|Ω| = 5 minutes, image acquisition rate = 12.3 Hz).…”

“…It corresponds to an adaptation of state-of-the-art statistics of pointprocesses to 1-dimensional temporal case of spiking events. This method has been developed to account for false inferred spike detections and potential time-shifts that results from spike deconvolution methods in calcium imaging [14]. As our method is based on statistical characterization of the Ripley's K function and hypothesis testing, it is robust to noise (false point detections) and remains accurate even for high point density.…”