Fig. 2: d-NMF pipeline.

a Pictorial representation of the workflow. Image sequences may be filtered and/or downsampled to reduce memory requirements and improve signal-to-noise ratio. The downsampled image sequence is then split into patches. Within each patch, ROI cores are detected and their spatial footprint and temporal traces are iteratively estimated using sparse constrained NMF. Once all patches have been processed, overlapping ROIs from neighboring patches are tested to see if they can be merged. Finally, ROIs are screened for realistic activity (see panel c). b Algorithmic sketch of the ROI detection (left) and refinement (right). See Methods for further details. c Top, illustration of accepting ROIs by skewness. Left, a sample portion of a FOV containing a true ROI (orange) and false ROI (brown). Scale bar: 15 μm. Example is 1 of 11 FOVs the following analysis was performed on. Right, the activity trace for the true ROI, with a skewness value of 6.24, and the activity trace for the false ROI, with a lower skewness value of 0.58. Bottom left, receiver operative characteristic (ROC) curves for individual FOVs (light gray) overlaid with the mean ROC (thick black), parametrized on the skewness cutoff of included ROIs. The mean False Positive Rate versus True Positive Rate are plotted for four example skewness cutoff values. Bottom right, classifier performance plotted against skewness cutoff. On average, a skewness cutoff of 3.8, [3.2, 3.9] resulted in optimal performance across all tested fields of view. n = 11 FOVs.