Temporospatial image-processing enhances salience of Ca²⁺ transients associated with neuronal activity. A. Raw image shows the wide dynamic range of static fluorescence signals that complicates detection of small amplitude Ca²⁺ transients. Boxed region is expanded on the right to highlight image changes accompanying filtering steps. B. “Derivative” image obtained by subtracting the target image (low-pass filtered using a 0.5 s moving average) from the image obtained 1.5 s earlier (low-pass filtered using a 1.0 s moving average). In this image, fluorescence changes during Ca²⁺ influx are more salient, pixels within bright regions are still quite variable, with active regions showing up as going towards white (during Ca²⁺ influx) or towards black after return to baseline (arrows). C. Spatial averaging homogenizes pixel values within regions of cellular activity, as well as in quiescent regions. D. Binary thresholding sets regions of static fluorescence to black, and regions of dynamic fluorescence to white.

Using 80 thresholds, 58 ROIs are identified in a high spatial resolution low-noise, slow optical recording data series. A. As the number of thresholds increases, the number of ROIs associated with respiration-modulated activity goes up (left); conversely, the ratio of respiration modulated ROIs to false positives decreases as the number of thresholds increases. B. View of the tissue, with ROIs under control (white outlines) and synaptic blockade (black fill, dotted white outlines) conditions, and location of recording (inset; VIIn: facial nucleus). C. Traces of respiration-modulated activity obtained from the ROIs shown in A; top left trace is rectified integrated motor output (C1). Gray bars indicate inspiratory activity; neurons active at a variety of phases, and both strongly and weakly respiration-modulated are included. Because of space constraints ΔF/F scale bars have been omitted. Traces contained in dotted rectangle in the center of B are indicated by arrows; although all these traces share respiratory modulation, they are sufficiently different to suggest that they are generated by separate units. D. Traces of spontaneous activity recorded under conditions of synaptic blockade. Under these conditions, neurons active once in 180 s, and neurons active at frequencies higher than control respiratory frequencies were observed.

Image processing sequence for ROI extraction. Boxes indicate processing steps applied to single images, the rhombus indicates processing carried out on the image stack generated by the earlier processing steps, and ovals indicate operations on traces. At the end of the ROI extraction process, mean luminance values associated with each ROI are measured from every raw image. These traces are then screened by the end-user, and good traces are saved.

Average of binary image series undergoes successive thresholding to identify ROIs. A. In the average of the binary images, regions of intense and/or regular neuronal activity are near-white, while regions of weak and/or intermittent activity are near-black. B. Intensity profile of a row of pixels indicated by dotted line in A. Using the methods described here, an ROI will be generated for points where a lower threshold intersects a peak in the intensity profile, but where the threshold above it does not. C. Superimposed binary images generated by a low threshold (grey blobs) and the threshold above it (white blobs). Only blobs retained using the lower threshold, but missed by the upper threshold are stored as ROIs (dotted boxes).

ROIs can be generated using noisy, low spatial resolution datasets. A. anatomical location of recording overlaps with the preBötC, which has been identified as lying 500 µm caudal to the VIIn (Ruangkittisakul et al., 2005). Two juxtaposed ROIs are likely generated by the same cell (arrow). In the 4 datasets analyzed, the number of respiration modulated ROI traces remained at approximately 10% of the total number of ROIs identified (inset). B. Traces sorted to match ROI numbering reveal closely matched activity in 2 adjacent ROIs (arrows), suggesting that one cell has been misidentified as 2. Despite the fact that traces 8 and 9 are likely generated from the same cell, the two traces do not match exactly (boxes). C. Effects of high- and band-pass filtering on the optical recording from one ROI (dotted box, 2^nd optical trace from the bottom, Figure 5 B). Top trace: the raw optical signal, obtained by calculating the mean of pixels within the ROI for each frame in the data series. High-pass filtering (via subtraction of the trace low-pass filtered using a 14 s moving average) eliminates slow time-course fluorescence fluctuations (middle trace); subsequent low-pass filtering (by taking a 0.6 s moving average of the signal) gives rise to a band-pass fitered signal in which high-frequency fluctuations in brightness associated with arc-lamp jitter is attenuated (bottom trace), but in which the underlying Ca²⁺ signal is retained. The effect of low-pass filtering on one Ca²⁺ peak associated with an inspiratory burst (dotted box, left) is shown on an expanded time-scale (right panel).

Simulated data were used to estimate type 1 and type 2 error under conditions of varying noise. A. i. simulated image series were based on a 20X image of indicator- loaded tissue. ii. Regions of time-varying fluorescence retained naturalistic morphology and distribution, as they were defined by thresholding the raw image to generate a binary image. iii. Pixels in black regions of the binary image were assigned random values; pixels in white regions were assigned values representing the sum of a time-varying signal and noise; so as to maintain constant dynamic range, as noise was increased the amplitude of the time-varying signal was decreased. B. Time-varying signals assigned to white regions of the binary image were obtained from fluorescence measurements from a respiration modulated neuron (i) or from simulation of sparse activity generated by splicing two peaks associated with inspiratory activity between epochs of random values scaled to 10% of burst amplitude (ii). Upper traces show traces in which signal and noise components were of equal amplitude (S/N=1); lower traces were obtained at S/N=0.33. Traces were scaled so that ΔF/F bars matched. C. Plot of type 1 (□) and type 2 (■) errors as a function of S/N for rhythmic (i) and sparse (ii) simulated datasets. Type 1 error was calculated as the ratio of ROIs that picked out “cells” to the total number of ROIs generated; type 2 error was calculated as the ratio of identified “cells” to the actual number of “cells” in the simulation. Both errors are minimized as these ratios approach 1; to convey conditions under which both types of error are minimized, (1-ratio) is plotted. As S/N decreased the lowest threshold applied to the summed image was increased so as to reduce type 1 error, and more accurately estimate type 2 error (with sufficiently high ROI numbers, all “cells” will be identified). Lowest thresholds are expressed as percent of the simulated image’s dynamic range (right axis), and are indicated by bars.