Color Separation

Color Separation sits in a layer's Pre-Processing section, under Add Operations (basic), and works only in brightfieldLoading... projects. Its Input Image is the Original Image by default, or another color image in the layer. You add one marker per stain with Add Marker, give each a reference color, and choose a Method.

Single Reference Shades performs an unmixing of each pixel's color against the markers' reference colors. Suppose you pick a blue shade for Hematoxylin and set its Intensity to 200. In the Hematoxylin output, a pixel with as much marker as the spot you picked reads 200, a lighter blue reads less and a darker blue more. If the spot you picked was itself a mix, enable Is mixture of, name the other marker and enter its Mixture Percent; Pure Color then shows the corrected reference.

Multiple Reference Shades is a classification instead. Each marker gets a set of shades, and the engine estimates, for every pixel, the probability that it belongs to each shade from the distance between the colors in the chosen Color Space. The closer the color, the brighter the output pixel. Each shade has its own Measure, Gain, Sigma, Gauss and Color Space settings.

Either way, the result is one grayscale imageLoading... per marker, plus a color image per marker if you tick Generate Color Outputs.

Source: StrataQuest 8.1 Engines Reference, ch. 10 (pp. 382–395), ch. 6.2 (pp. 98–99), ch. 6.11 (p. 311); User Manual §11.1 (p. 154).

Why absorbing stains are harder than fluorescence. In general, a fluorophore adds light and a stain removes it. Transmitted light falls off exponentially with the amount of stain (the Beer-Lambert lawLoading...), so two stains in one spot multiply their transmissions rather than add them. Taking the negative logarithm turns transmission into optical densityLoading..., where stain amounts add. That is the space in which textbook color deconvolution (Ruifrok and Johnston, 2001) separates stains with a matrix inverse.

What StrataQuest documents. The Engines Reference's formula for Single Reference Shades is an image missing from the text, so whether the engine works in optical density is not known from the documentation. StrataQuest does have Convert RGB to Optical Density (OD) in the Basic Operations Module, described as lossless and based on the Beer-Lambert law, so you can build an OD image yourself.

How many stains fit in RGB. In the textbook linear model, an RGB camera gives three numbers per pixel, so at most three stains can be solved for exactly. The Engines Reference states no such limit for Color Separation. Only Autodetect Colors is restricted, to two markers.

Color spaces for clustering. Multiple Reference Shades lets you pick the color spaceLoading... and the distance, Euclidian or Manhattan. RGB is not recommended, because lighter shades of different colors may group together. HSV/HSL group shades of different intensities, and LAB/LUV are recommended because they are closer to human perception. In general, LAB was designed so that equal distances look roughly equally different to the eye, which is why it suits color clustering.

The Engines Reference example is an 8-bit RGB image of a Ki67 IHC stain on colon tissue, separated into Hematoxylin and Ki67.

1. Add Color Separation, choose Single Reference Shades, and define two markers, Hematoxylin and Ki67.
2. Press Autodetect Colors, or pick each reference with the Color Picker. Autodetect works on a 1000 × 1000 pixel area; if it assigns an unexpected color, change Outliers (%) to leave out extreme shades.
3. Check each marker with View Gray Image before moving on.
4. Use the Hematoxylin image for NucleiLoading... detection and the Ki67 image as the Shades input of Standard MeasurementsLoading.... The engine's Quick Action Add Nuclei Segmentation and measurements sets up that pair for the first marker.

Several Threshold and Compare examples in the Engines Reference start from these same two images, combined into one grayscale image and turned into a maskLoading... with a threshold.