The choice of color space depends on which axis lines up with the property you are measuring, and rarely on display. A few cases:
Segmenting one chromogen against another in IHC. DAB brown and hematoxylin blue absorb at different wavelengths, but their RGB values interact: a darker DAB region can look, in R, G and B, like a lighter hematoxylin region. Color SeparationLoading... is the dedicated engine for this. Conceptually, the move is to convert into a space where the chromogens separate cleanly: HSV by hue, or per-stain optical-density channels.
Quantifying stain intensity for measurement. RGB values aren't linear with stain concentration, because the Beer-Lambert relationshipLoading... is exponential. Convert RGB to OD linearizes it. Measurements on the OD image, such as mean or integrated density per cell, then quantify stain concentration directly, which is the quantity of interest.
Color-based classification with variable lighting. A ClassifierLoading... trained on RGB features performs poorly when the test images are lit differently from the training images, because the same biology shows up as different RGB values. Convert to HSV first, then train and predict on hue and saturation and leave out value, which carries the lighting variation. The classifier becomes lighting-invariant.
Computing color difference for distinguishing populations. To measure how different in color one region is from another, compute the distance in Lab, where distance corresponds to perceived difference.
Reducing to grayscale for shape analysis. When only structure matters and color carries nothing the analysis needs, Convert RGB to Grayscale drops the extra dimension. The result is the input to most filteringLoading... and labelingLoading... engines, which are designed for one-channel data.