Fig. 2

(a) Illustration of the parallelization approach used by CaImAn batch for source extraction. The data movie is partitioned into overlapping sub-tensors, each of which is processed in an embarrassingly parallel fashion using CNMF, either on local cores or across several machines in a HPC. The results are then combined. (b) Refinement after combining the results can also be parallelized both in space and in time. Temporal traces of spatially non-overlapping components can be updated in parallel (top) and the contribution of the spatial footprints for each pixel can be computed in parallel (bottom). Parallelization in combination with memory mapping enable large scale processing with moderate computing infrastructure. (c) Quality assessment in space: The spatial footprint of each real component is correlated with the data averaged over time, after removal of all other activity. (d) Quality assessment in time: A high SNR is typically maintained over the course of a calcium transient. (e) CNN based assessment. Top: A 4-layer CNN based classifier is used to classify the spatial footprint of each component into neurons or not, see Materials and methods (Classification through CNNs) for a description. Bottom: Positive and negative examples for the CNN classifier, during training (left) and evaluation (right) phase. The CNN classifier can accurately classify shapes and generalizes across datasets from different brain areas.