Threshold & Compare

Two names mislead. Threshold doesn't binarize; the reference describes it as capping values, and its output is grayscale, matching the input. And Compare tests an image against numbers, not against another image. The reference's Threshold - LT, GT example makes the capping concrete: with Threshold (LT) 80, Value (LT) 5, Threshold (GT) 100 and Value (GT) 80, pixels below 80 become 5, pixels above 100 become 80, and the rest keep their values.

For one threshold computed from all fields of view, use the Otsu Threshold (global)Loading... engine, which is added in the layer's Pre-Processing. The BOM's Otsu works on the image in front of it.

Source: StrataQuest 8.1 Engines Reference, ch. 6.11 (pp. 310–348); ch. 6.7 (pp. 206–226); ch. 33 (pp. 480–481).

Thresholding is where a continuous grayscale image becomes a discrete decision: foreground or background. It is the most basic segmentation operation, and the method and value chosen shape every result downstream. See foreground and backgroundLoading....

Global Thresholding & Otsu's Method

A global threshold applies a single value T to the whole image. It works when the histogramLoading... is bimodal, with foreground and background pixels forming two separable peaks. Otsu's method formalizes that: in the textbook form it picks the threshold that maximizes the between-class variance σ²_B = ω₀ω₁(μ₀ − μ₁)², where ω₀ and ω₁ are the class probabilities and μ₀ and μ₁ the class means. The reference states the same criterion as minimizing intra-class variance or maximizing inter-class variance.

The reference states the limits. Otsu works well when the histogram has two peaks and a deep valley. It degrades when the object area is small compared with the background, when the class variances are large compared with the difference of the means, or when noise or shading fills the valley. Its remedy is to restrict the range of values the algorithm uses, which is exactly what Otsu (with range) does.

Adaptive Thresholding

In image processing generally, adaptive thresholding sets a different cutoff at each pixel from its neighborhood, typically the local mean plus an offset, so the cutoff follows uneven illumination. StrataQuest has no adaptive threshold operation. The idea can be assembled from documented operations: an Average filterLoading... with a large Kernel radius (px) estimates the local mean, Subtract images takes it away from the original, and Compare with Greater tests the difference against an offset. The window size is the trade-off: too small and the cutoff follows individual nuclei, too large and it stops adapting.

Multi-Level & Hysteresis Thresholding

Multi-level Otsu, in general, extends the method to more than two classes by choosing several thresholds; the BOM's Otsu returns one. Hysteresis thresholding uses a low and a high threshold: pixels above the high one are foreground, and pixels between the two count only if they connect to a definite foreground pixel. StrataQuest uses hysteresis inside its Canny edge detectorLoading.... There is no general hysteresis operation, and per-pixel compares can't supply the connectivity step.

Compare Operations as Boolean Algebra

Binary masks are sets of pixels, so combining them is set algebra. StrataQuest's And, Or and Xor work bitwise on pixel values, and on 0-and-255 masks bitwise AND gives the intersection, OR the union and XOR the symmetric difference. Simple compare results build into complex regions this way: (channel A Greater T₁) And (channel B Less T₂) selects pixels bright in one marker and dim in another, a common step in multiplex immunofluorescenceLoading... analysis. See bitwise logical operationsLoading....

Compare

Use it when staining and illumination are even across the image and you know the cutoff. Check the histogram: two well-separated peaks with a clear valley mean a cutoff in the valley works. The reference's example passes every pixel Greater than 45 in a combined Hematoxylin and Ki67 image from Color SeparationLoading.... For color images, Compare (RGB) applies one test per channel and passes a pixel only if all three pass.

Otsu and Otsu (with range)

Use Otsu when the histogram is bimodal and you want the cutoff chosen for you, image by image. Use Otsu (with range) when noise, shading or a small object fraction would pull the cutoff off; set Lower limit and Upper limit around the region where the valley belongs. Use the Otsu Threshold (global) engine instead when every field of view should share one cutoff.

Interval compares

Use Compare - inside interval for a band of intensities, for example moderately stained pixels, and Compare - outside interval for both tails at once. Set Include interval limits to Yes when a pixel exactly at a limit should pass.

Threshold, Threshold - LT, GT

Use these to cap values before further processing, not to make a mask: for example, pull saturated pixels down to a ceiling so they don't dominate a later average. The result is still a grayscale image.

Comparing two images

No compare takes two images. To mark where channel A is brighter than channel B, take the difference with Subtract images, which subtracts Image 1 from Image 2, then Compare the difference with Greater against a positive value. To mark where either of two masks is set, use Or; where both are, And.