Fig. 4

(a) Comparison of CaImAn batch (top) and CaImAn online (bottom) when benchmarked against consensus annotation for dataset K53. For a portion of the FOV, correlation image overlaid with matches (left panels, red: consensus, yellow: CaImAn) and mismatches (right panels, red: false negatives, yellow: false positives). (b) Performance of CaImAn batch, and CaImAn online vs average human performance (blue). For each algorithm the results with both the same parameters for each dataset and with the optimized per dataset parameters are shown. CaImAn batch and CaImAn online reach near-human accuracy for neuron detection. Complete results with precision and recall for each dataset are given in Table 1. (c–e) Performance of CaImAn batch increases with peak SNR. (c) Example of scatter plot between SNRs of matched traces between CaImAnbatch and consensus annotation for dataset K53. False negative/positive pairs are plotted in green along the x- and y-axes respectively, perturbed as a point cloud to illustrate the density. Most false positive/negative predictions occur at low SNR values. Shaded areas represent thresholds above which components are considered for matching (blue for CaImAn batch and red for consensus selected components) (d) F1 score and upper/lower bounds of CaImAn batch for all datasets as a function of various peak SNR thresholds. Performance of CaImAn batch increases significantly for neurons with high peak SNR traces (see text for definition of metrics and the bounds). (e) Precision and recall of CaImAn batch as a function of peak SNR for all datasets. The same trend is observed for both precision and recall.