I am new to scientific image analysis and now rely on it heavily for my phd project. My main task is measuring sizes of aragonite crystals (or biominerals) from SEM images. For this I am starting to build a routine by which I analyze my images.
I know a bit about segmentation, particle analysis etc. but am still super insecure if what I am doing is correct / valid / scientific. So I have a few questions which I hope you can help me with.
My current workflow is:
- Set a scale for the image
- Either threshold (by hand or automatic, I found shanbhag works best) or use weka segmentation. Weka segmentation seems to be more accurate.
My images are directional (its the shell of a bivalve growing in a certain direction, along which I want to measure variations in crystal size). So I additionally:
- Rotate the image so that the growth axis is parallel to x
- Apply watershed. I don’t know if this makes it better or worse. Some crystals are not fully segmented so I try to improve this with watershed. However it over-segments some other parts.
- Analyze particles
The result of this looks like this (crystal sizes binned into 2 bins, calculating a running mean along the x axis)
So this isn’t really optimal I think. My main questions are:
- What preprocessing steps can I take to improve my work? (Apart from sample preparation which I’m working on)
- How to measure the accuracy of my segmentation method? I read that you have to establish a “ground truth” and compare your result with it. Should I segment the image by hand and compare my results? Do I need to do this only once (if I’m reusing the weka classifier)? How do you guys handle this?
- Are there better segmentation methods in my case? Apart from the pure image, I can do surface roughness reconstructions with our SEM machine. So I could generate additional “height” data for individual pixels if that would help.
But apart from these main questions I will gladly take any suggestion. I am very new to scientific image analysis and there is not alot of expertise in our group.
Thank you very much!