Supplementary Figure 3: Limitations of thresholding-based analysis.

(a) Top row: Six example frames of ex vivo Ca²⁺ data smoothed with a Gaussian kernel of σ=0.5 after square root transformation. Second row: Baseline for each pixel is estimated with a 20-frame time window, noise level is estimated as σ_noise, and baseline is subtracted from raw data to obtain ΔF. Third row: Standard threshold is set at 3σ_noise. Many individual events are erroneously detected as one very long and large spatiotemporal component, for reasons graphically explained in (b). Fourth row: A high threshold (10σ_noise) leads to loss of many true events, and many detected events are incomplete. Each color indicates an event. Fifth row: AQuA-detection avoids the pitfalls in threshold-setting and identifies each individual event. (b) Two events are incorrectly connected after thresholding (incorrect events in yellow in each sub-panel). Intensity color bar on right, with red indicating the threshold, refers to all panels. Top: Between multiple events in the same location, even though the intensity drops a lot, not all pixels will fall below the threshold. Each event is shown with a gray bounding box. The super-voxel step in AQuA solves this problem by finding a time window for each seed, and spatially extending the windows. Middle: Neighboring events are initiated at distinct times, but are spatiotemporally connected at a later time. If two regions have very different onset times, AQuA will treat them as different events in the super-event detection step. Bottom: Two events can be separated when they appear, but meet after propagating. In the event detection step, AQuA distinguishes these events based on the single-source rule.