Indetyfying pigment networks for Malenoma disease

Dear Cellprofiler users,

I was wondering if there is any exmple for dettion of pigment network in Malenoma images

I have attached a sample images of Malenoma (sfu.ca/~msa68/research_PN_detection.htm)

I wonder any help or hints how to do this with cellprofiler as I did not find any specific example.

Regards,

Hi,

It sounds like a nice project, and thanks for providing the background. No, we don’t have a specific melanoma or other skin pigmentation example, per se (speak up anybody else out there who might!). Though if you want to upload a raw image rather than the composite examples you provided, we could try and provide you with a pipeline seed to start from. CellProfiler does have some color deconvolution (UnmixColors), edge detectors (like Laplacian of Gaussian), and object segmentation built-in, but your graph analysis would have to be done post hoc. For that, CellProfiler does have MeasureObjectNeighbors to approximate a connected neighborhood, and even our MeasureNeurons module has a limited graph node output, but they would likely not be sufficient for your purposes.

As you know I’m sure, non-fluorescent images, be they brightfield, histopathology, or skin pigmentation, etc., are notoriously much harder to analyze in a controlled way. So variations in illumination as well as the color mixing can cause headaches, however we can certainly give it a try if you provide a couple raw images (one with network Present, and one Absent).

Thanks,
David

Dear David,

Thanks for the answer

I have attached a sample of images …the some image have a typical pigment (3,8,9,142,143,144,146,147) and the rest(2,4,6,417,418,419,420,421) have an atypical network.
Typical.zip (7.99 MB)
ATypical.zip (8.24 MB)

Hi,

I looked at your images, and attached is a pipeline that you can try and get started with. The biggest issue is “What do you judge is a pigmented/non-pigmented area that you call a node in the cyclic graph?”. This judgement is hard and best done by you, the expert.

So there are two approaches I suggest.
(1) Standard image analysis pipeline, as you have done already. To run the pipeline, use File > Import Pipeline. Then you need to drag/drop your images
NOTE:the Metadata module expects the folder name to be the “Condition” like “Atypical”.

ALSO NOTE: Many modules have comments in the Module Notes at the top that I wrote, so please refer to my thoughts there.
The first module (ClassifyPixels) is disabled (click on the green checkmark to enable/disable. This is used in the Approach #2 below.

(2) Use ilastik, a pixel-based machine learning classifier, that is included in the CellProfiler download, as a ‘seed’ for CellProfiler. The benefit is that the classifier would hopefully learn your pixel regions as you see them intuitively (a tough image analysis problem!), and might help to be robust against lighting and local intensity variations in background, etc. I zipped my attempt at an ilastik classifier (.ilp project file) as well as its exported file which gets imported into CellProfiler.

You would need to run ilastik and train the pixel classes, as shown in this screenshot: cl.ly/image/1g2k0p2K1w19 Then you would import use the ClassifyPixels module in CellProfiler (just enable it in the attached pipeline). Then you could add IdentifyPrimaryObjects to identify the objects on the classified images.

Hope that helps!
David
ilastik_files.zip (4.51 MB)
DL_melanoma_pipe_ilastik.cppipe (11.7 KB)