Background Removal

Suppose you open a 16-bit DAPI channel from a kidney section. Between the nuclei the pixels aren't zero. They scatter around some level, high enough that a thresholdLoading... or a mean-intensity measurement picks part of it up. The Engines Reference calls that background white noise: random values with a specific average value and finite variance.

Background Removal sits in a layer's Pre-Processing section, under Add Operations (Advanced). It generates a background model from the input grayscale imageLoading... and subtracts the model from the original. Before subtracting, it smooths the model with a filter whose size is the Background smoothing filter radius.

Object Radius tells it how large the larger objects are, so they aren't taken for background. Tick Generate background image to get the model itself as an extra output.

The reference doesn't say how the model is built. Its closest documented relative is the Top-hatLoading... operation in the Basic Operations Module, which subtracts a morphological opening, "an estimated background", and whose reference example shows an attenuation of the high background. Whether Background Removal uses an opening internally is not stated, so don't tune it as if it did.

Why background is never zero. In general, a camera adds an offset and read noiseLoading... even in darkness, and tissue adds autofluorescenceLoading... and out-of-focus glow. The sum is a floor with random scatter. In a 16-bit image, with values up to 65,535, that floor can sit hundreds of counts up. StrataQuest's own documentation describes the same thing elsewhere: the Legacy Spectral UnmixingLoading... engine notes that 16-bit fluorescence background is almost never pure black, with values in the range of 100 to 300.

Background versus shading. Image formation is often written as the specimen's signal multiplied by the illumination, with an additive floor on top. Shading is the multiplicative part, a lamp brighter in the middle of the field than at its edges. It is corrected with a correction image of the light path: StrataQuest's Shading Correction, or Illumination Correction and FL Shading Correction in TissueFAXS. Background Removal works on the additive floor, by subtraction.

The scale question. Any method that tells background from objects by size needs a definition of "large". In general, a background estimate made with a window smaller than a nucleus follows the nucleus itself, and subtracting it carves the object hollow or leaves a dark rim. That is the reason behind the reference's warning that a too-small Object Radius lets larger objects be affected by the removal.

The reference prints no default values and no recommended range. In its example the settings appear only in figures. A workable path is to set Object Radius near the radius of the largest objects you care about, tick Generate background image on a test region, and look at the model. If objects show up in it as bright blobs, Object Radius is too small.

Source: StrataQuest 8.1 Engines Reference, ch. 5 (pp. 14–17), ch. 6.8 Top-hat (pp. 281–283), ch. 39 (p. 515); User Manual §9.4 (pp. 130–134); TissueFAXS 8.1 User Manual §4.4.3 (pp. 91–96).

Take the Engines Reference's own case: a small ROI in a kidney sample, acquired at 16 bits in DAPI, FITC, Texas Red and Cy5, with a high background.

1. Add Background Removal to the layer's Pre-Processing once per channel, with that channel as Input Image.
2. Use one group of settings for DAPI and a second for FITC, Texas Red and Cy5, as the reference does. It doesn't say why; a likely reason is that the larger objects differ in size between the nuclear channel and the others.
3. Check each background model with Generate background image on a test region before running the whole sample.
4. Feed the corrected DAPI image to NucleiLoading... detection, and the corrected marker channels to Standard MeasurementsLoading... as Shades images.

Decide on purpose which image you measure. Background-subtracted intensities are lower than raw ones by the local value of the model, so use the same choice, and the same settings, for every sample you compare.