I need to measure a specific inflammation area of on H&E stained tissue sections.
I came with the following way to do it in FIJI.
First I set the scale, I draw a straight line on my scale, the go to set scale and input 500µm.
Then I go to my section and with the freehand selection tool, I draw around the inflammation area that I want to quantify the surface. Then I measure and collect the Area measure (µm2, I suppose). Is is the correct way to proceed?
If you have pixel size information and the regions of interest are visually distinct enough, there are a number of classifiers you could use to detect areas automatically, with some training. See this thread for some options and or guides on Weka/Ilastik, etc.
If you add the correct pixel size to the Image->Properties… in FIJI, any area you draw you should be able to add accurate measurements to.
That is a way of doing it manually, but you have 2 serious problems.
- If you repeat the procedure several times, do you get the same result?
- If other people do it, do they get the same result?
Those are intra-operator and inter-operator variability.
Most imaging is directed to avoid those two by designing algorithm that are guided by image information and not the subjectivity of the operators.
Yes I do it manually but I would like to avoid variability and ideally find an automated process.
Here is the type of pictures I’m workin with. The inflammation is the blue part. The scale bar is 250µm.inflam.tif (6.5 MB)
How could I implement an automated system that decreases variability from user to user?
I took a quick swipe at it, and while I am sure someone who knew exactly what they are looking for could do better, it looks like some of the simple pixel classifiers might work well enough. A lot will depend on the inter-picture/sample variability, but I think it is worth pursuing, at least for a little while and based on the one image.
That looks really good!
Could you explain step by step how you did it?
Unfortunately, the method I used wouldn’t be very useful to you. It was just the pixel classifier in QuPath, which is not yet implementable at large scale (you have to create training data for each image, so it cannot provide an unbiased measurement across your samples). It really was just for testing purposes, as proof that the method was possible. You would be better off with Ilastik or Weka, both of which I am not as familiar with.
Note that including the scale bar could be problematic, unless you train it out specifically
Better to add the pixel data to the TIFFs in FIJI.