Cutoffs

A cutoff lives on a diagram, not in an engine. You build a scattergram or histogram from the measurements of a layer, perhaps a ratio made with Derived MeasurementsLoading..., and place a cutoff on an axis. Events above it are positive for that measurement and events below it negative. On a scattergram, cutoffs on both axes make four quadrants: upper-right double positive, lower-left double negative, upper-left positive for y only, lower-right positive for x only.

Placing a cutoff. Open the dropdown of the Cutoff button in the diagram toolbar and choose Manual or Automatic. Manual means clicking in the diagram where the line should go. Deviation from Mean, with Negative and Positive Deviation options, puts the line at the population mean minus or plus 1 to 4 standard deviations; on histograms only the x-axis options exist.

A percentile cutoff is also offered, and the diagram settings hold a Cutoff Percentile Value, the percent of the events population you give. Keyboard shortcuts set the deviation cutoffs directly: Ctrl+1 to Ctrl+4 for mean minus 1 to 4 SD and Ctrl+5 to Ctrl+8 for mean plus 1 to 4 SD, adding Alt for the x axis only or Shift for the y axis only.

Adjusting it. Drag the line, or double-click its value and type a new one. You can move the vertical or the horizontal cutoff, but not both at once. The right-click menu sets a cutoff as Local, for one region (its value shows in italics), or Global, propagated to all diagrams in the sample links.

What comes out. Quadrant statistics report the percent of events, by count or by area, in each quadrant, for the whole diagram and for each gateLoading.... Each quadrant can be opened as Backward Data on the images, chosen as an input gate for another diagram (the list holds gates, quadrants and phenotypesLoading...), turned into a new layer of events, or added as a column in the Statistics Report. The manual describes no per-event measurement written by a cutoff; its results travel as quadrants.

Source: StrataQuest 8.1 User Manual §12.6 (pp. 177–209), §12.7 (pp. 216–219), §13.1 (pp. 231–234), §16.1 (p. 271).

Where to put the line. In statistics generally, a threshold between two overlapping populations trades one error for another. Move it down and you catch more true positives along with more false ones; move it up and the reverse happens. Where the best compromise lies depends on how the two populations are spread and how common each is. For two Gaussian populations, textbooks give the minimum-error threshold in closed form from their means, variances and proportions.

What the deviation cutoffs assume. Mean + 2 SD flags something narrower: the events that are unusual for this population. The manual notes that these thresholds are statistically predefined while their absolute values come from the population at hand, so the same rule lands at a different value on a population with a different mean or spread. If the population is mostly negative cells and roughly Gaussian, about 2.3% of the negatives fall above mean + 2 SD, and the line flags cells that stand out from the background. If half the cells are positive, they inflate both the mean and the SD, and the same rule falls somewhere less meaningful.

Choosing from outcomes. In translational research, a cutpointLoading... is often chosen from outcome data. For each candidate value, you compute sensitivity and specificity against a recorded outcome such as response or survival, and then read the tradeoff off an ROC curveLoading.... One common choice maximizes the Youden index, sensitivity + specificity − 1. The User Manual describes no outcome-based optimizer, so that analysis happens on exported data, and you then type the chosen value into the cutoff.

Cutoffs on events, thresholds on pixels. A cutoff thresholds events in a diagram. The Otsu ThresholdLoading... engine and the threshold enginesLoading... threshold pixels in an image to make a mask. The idea is the same, applied to different things: one determines which cells count as positive, the other which pixels count as foreground.

Here is a cutoff doing its own job, with a sample-level score built on top of it.

1. Detect nuclei, measure each cell with Standard MeasurementsLoading... on the CK and PD-L1 channels, and plot CK mean intensity (x) against PD-L1 mean intensity (y) on a scattergram.
2. Set the x cutoff where CK-positive tumor cells separate from the rest, and the y cutoff where PD-L1 staining begins. The upper-right quadrant now holds PD-L1-positive tumor cells and the lower-right PD-L1-negative tumor cells.
3. Check the placement with View Backward Data for the UR quadrant, which outlines those cells in red on the images.
4. In the Statistics Report, add the events count of each quadrant as a Diagram Statistic. Combine them with Derived Statistics, which apply one arithmetic operation to two statistics, or to a statistic and a number: upper-right divided by (upper-right plus lower-right), times 100.

The result approximates a Tumor Proportion Score for the sample. The published TPS convention counts membrane staining of any intensity, so a mean-intensity cutoff is a stand-in you would validate against manual scoring of the same slides. A sample-level threshold, such as the 50% level in the published TPS convention, is read off the report by you. It is not a diagram cutoff, because a diagram cutoff divides individual cells and TPS describes the whole sample.