Fig. 2. Overview of DynamicNeuronTracker.

(A) Workflow of DynamicNeuronTracker (DyNT) that takes a rigid-registered calcium video as input and produces dynamic ROIs and their calcium activity time courses. Red boxes indicate data objects and blue boxes indicate computational operations. (B) Patch matching iterations in (A) are illustrated. In the first iteration, a local mean image of a neuron firing event is matched to another firing event nearby at different time frames using spatial correlations as illustrated in (C-E). From a resulting trajectory, we obtain an updated local mean image utilized to run the second iteration. (C) An example MIP of a local mean image of a firing neuron. The 3D patch size (11 × 11 × 5 pixels) is a user input. (D) An example MIP of a neighborhood local volume that contains a neuron being tracked. Rolling spatial correlations between a local mean image (C) and its neighborhood volume (D) are computed in 3D to examine when the mean image appears again at late time points. The size of the neighborhood volume (21 × 21 × 11 pixels) is also a user input depending on the degree of jittering. (E) Maxima of rolling spatial correlations at every time frame. High correlations indicate the neuron is likely to be active at the time. (F) (Left) Raw calcium activities of an ROI are obtained by averaging fluorescence intensities within the ROI at every time frame (black). Red indicates moving medians of the raw activities computed in a rolling window of 81-frames. (Right) Normalized calcium activities, ∆F/F = (raw activity – moving median)/moving median.