A database and deep learning toolbox for noise-optimized, generalized spike inference from calcium imaging

Abstract

Inference of action potentials (‘spikes’) from neuronal calcium signals is complicated by the scarcity of simultaneous measurements of action potentials and calcium signals (‘ground truth’). We compiled a large, diverse ground-truth database from publicly available and newly performed recordings in zebrafish and mice covering a broad range of calcium indicators, cell types, and signal-to-noise ratios, comprising a total of >35 recording hours from 298 neurons. We developed an algorithm for spike inference (CASCADE) that is based on supervised deep networks, takes advantage of the ground-truth database, infers absolute spike rates, and outperforms existing model-based algorithms. To optimize performance for unseen imaging data, CASCADE retrains itself by resampling ground-truth data to match the respective sampling rate and noise level; therefore, no parameters need to be adjusted by the user. In addition, we developed systematic performance assessments for unseen data, openly release a resource toolbox, and provide a user-friendly cloud-based implementation.

Effective and Efficient Neural Networks for Spike Inference fromIn VivoCalcium Imaging

“…All the inference systems were then benchmarked under this same sampling rate. We also tested our system under 7.5Hz as suggested by CASCADE in 26 . The impact of sampling rates on inference results is discussed in this work.…”

“…In practice, considering the typically slow dynamics of calcium signals and relatively low firing rates of neurons imaged, this estimation method operates in linear time in proportional to the duration of recordings. We have validated this spike-events estimation algorithm with the Monte-Carlo importance sampling based algorithm proposed in CASCADE 26 .…”

“…Secondly, we use the van Rossum distance (vRD) 42 for the evaluation of spike rates prediction where the time constant 𝜏 is the normalizing factor (smoothing window size) for smoothing spike-events into spike-rates (e.g., 𝜏 = 0.025 s for our proposed system). Moreover, Error and Bias proposed in 26 are also used to evaluate spike-rates On the other hand, for measuring spike-event prediction, we adopt the Victor-Purpura distance (VPD) 44 . It is defined as the minimal cost to transform the PD spike-events to the GT spike-events.…”

“…We have validated this spike-events estimation algorithm with the Monte-Carlo importance sampling based algorithm proposed in CASCADE 26 .…”

“…For real world applications, a calibration-free inference system that could generalize on un-seen recordings is always preferred. Recently, with the increased availability of paired data, a data-driven method CASCADE was proposed 26 . It was reported that top-ranking performance on un-seen testing datasets was achieved by the neural networks, which were deliberately trained with paired data whose noise-levels matched the testing dataset.…”