Open jccaicedo opened 1 year ago
Yes, I do have a question about those outlines. Does DeepProfiler assume the outline is on the inside or outside of the cell? In other words, is the outline itself 'part' of the cell? I ask because the default with CellProfiler is to display outlines on the 'inside' but depending on how this is treated, we may want to ensure users select the 'outside' option. Here are the options from CellProfiler's help doc:
Specify how to mark the boundaries around an object:
I can help generate an example if we still need one. Can someone link me to a single image of a cell from the example dataset (or attach it here)?
@rsenft1 There is an example in training (export) section (Figure 3). In the perfect case scenario, we expect all mask pixels to belong to the object only (I believe from the description of CellProfiler outlining it is filled inner outline) because DeepProfiler simply multiplies image and binary mask.
Wait, sorry do you need an outline or a mask? Should the cells and nuclei just be outlined in white or should the pixels belonging to cells be all white? The process for going about getting outlines/masks in cellprofiler is somewhat different.
Oh sorry, I got confused with single-cells.
So in case of outlines (in outlines
folder ), only cell outlines are needed.
About Inner \ Outer \ Thick - it becomes a trickier question: @jccaicedo could you please help?
I believe the default option with CellProfiler should be good, as we have traditionally followed the Cell Painting pipeline. @rsenft1 DeepProfiler needs outlines as an input, but it transforms those outlines to masks. Thus, no need to save the masks. Does this make sense?
@Arkkienkeli You could select an example image from one of our datasets and display both the original (a pseudo-color composite) and the outlines. This is necessary to give people an idea of what they should aim for.
@rsenft1 if you can point to CellProfiler instructions on how to get these images, that'd be great!
Got it. Just to make sure I understand, how does DeepProfiler handle two different cell outlines that touch? Conversely, what about a cell that narrows to a point and then expands again (like the axon of a neuron might get very thin and then slightly thicker?
Usually, the way we'd deal with including information about where cells are is to use a label matix where pixels belonging to each cell are all of a given value (cell 1 pixels are all 1, cell 2 pixels are all 2 etc.) But with the outlines, it seems like there's some room for ambiguity.
@jccaicedo could I get your thoughts on ⬆️ ? I'm happy to write the section, but I'd definitely want to include some info on these edge cases as well.
This page is looking good! We need to expand its content in the following way: