Fig 1.
Signal-to-signal neural network (S2S).
(left) Block diagram of the proposed approach. (right) An illustrative example calcium signal, its corresponding spike estimate, and the ground truth spike train.
Table 1.
Overview of top-performing algorithms in spikefinder challenge.
Table 2.
Comparison of evaluation measures between S2S and state-of-the-art spikefinder baselines.
Fig 2.
Dataset-wise performance to show the difference in evaluation measures between S2S and convi6.
(a) Correlation measure, (b) Rank (non-linear correlation) measure and (c) AUC measure.
Table 3.
Dataset-wise performance of S2S on the spikefinder test set.
Table 4.
Results on paired t-test for statistical significance.
“H” represents the hypothesis that there is no difference in performance between S2S and convi6.
Table 5.
Results on scalability experiments at 25 Hz and 100 Hz sampling rates.
Fig 3.
Change in the evaluation measures with Gaussian windowing of the training targets.
(left) Bar diagram depicting the difference in Correlation, Rank and AUC when using the 100 Hz spike train as the target and when convolving this target with a Gaussian window to generate a smoothed training target. Two different windowing sizes are shown; 11 samples with 5 standard deviation and 33 samples with 11 standard deviation. (right) From top to bottom: An example calcium fluorescence signal and its corresponding training targets (spikes at 100 Hz, Gaussian targets with 5 and 11 samples each, and the ground truth at 10,000 Hz). Gaussian training targets are having an equivalent shape to the spikes target at 100 Hz. Note that all of them are approximations of the original ground truth.
Table 6.
Results on 10-fold cross-validation with and without Gaussian windowing for the training target.
GT: Ground Truth, AVG: Average and STD: Standard Deviation.
Fig 4.
Evaluation measures with changes in number of hidden layers.
(left) Bar diagram depicting the difference in correlation, rank and AUC when no hidden layer, 1 hidden layer and 3 hidden layers are used in the S2S network, respectively. (right) Illustrative example showing the improved spike estimates when 3 hidden layers are used compared to one hidden layer.
Fig 5.
Dataset-wise performance showing the generalization ability of the three hidden layer S2S network.
“GCaMP” indicates that the training was done only with GCaMP indicator dataset. Datasets 1, 2 and 4 are based on OGB indicator.
Table 7.
Generalization across indicators.
Experimental results with various number of hidden layers.
Table 8.
Results on generalization experiments.
Experiments 2b and 3b represents generalization from GCaMP to OGB and vice-versa. For every experiments, the number of datasets used for training and testing are given inside the brackets.
Fig 6.
Layer-wise output of a 3-dense layer S2S.
Note that the calcium changes which are not clearly distinguishable in the original fluorescence signal are amplified in the output of the deconvolution layer.
Fig 7.
Cumulative frequency responses of filters.
(left) Shows the frequency response of the analysis filters. (right) Shows the response of synthesis filters having larger amplitudes at equal frequency intervals. The energy at the output of S2S is purely concentrated on spikes both at the temporal and the spectral domain.
Fig 8.
Comparison of spike estimates.
(a) An example calcium input signal, spike estimates of (b) S2S and (c) convi6 methods, and (d) discrete ground truth. Observe the similarity in shape between the S2S and the ground truth, compared to the convi6 method.