Quantify lamellar pearlite structure


Dear image analysis specialist

I have electron microscope pictures of the microstructure of a steel wire. THe microstructure is mainly pearlite (=lamellar structure of cementite (=Fe3C) and ferrite (=+/- pure Fe)).
The microstructure has a big influence on the mechanical properties, so this is from prior importance when setting the parameters of the steel wire process.
When the wire is too long at too high temperature the lamellae become globular.

Would there be a way to quantify images (area% nice lamellae, area% bad lamellae).

Example image: left hand side is good, right hand side is bad.



From what I could gather - you want to Segment your lamellae (tube-like objects) and measure some properties of theirs to the define what you consider ‘good’ and ‘bad’ lamellae - right? What parameters would you use to define those good/bad lamellae - area? perimeter? length? orientation? This will make a difference in how you choose to set out in your analysis… So provide a bit more detail, and we can get you the help you need.

To start - you’ll need to Segment your lamellae - so here are a few helpful links regarding that:

I tried out TWS just using the default settings and got some good segmentation results:

From there you can export a probability map and then auto-threshold it and analyze particles, and measure features, etc. Just check out the Segmentation links above - especially the workshop - will be most helpful.

I hope this at least gets you started!

eta :slight_smile:


Good day Ellen,

I fear your segmentation is not what the OP was after. He wrote:

Example image: left hand side is good, right hand side is bad.

If this turns out to be literally understood, I fear that class-definition will become quite difficult:


Anyway, this task appears being a challenge for TWS and perhaps even for untrained humans.





No worries - I may have incorrectly interpreted exactly what @BasWerbrouck wants/needs for his/her analysis… but just trying to at least start a clearer dialog. :slight_smile: But from I gather - segmentation would be the first step… this is the image posted:

So I gathered that these lamellae - are more ‘elogated’ on the left side (ie - ‘nice’) and those on the right side are shorter and more circular/wide (ie - ‘bad’). But I could be off on this assessment. Perhaps some sort of filament analysis would be more appropriate? I’m not sure…

In any case - @BasWerbrouck - we will just wait on your further descriptions and know we are here to help in whatever way we can!

eta :slight_smile:


Segmenting short from long lamellae using TWS would be a tall order. But I don’t think it’s necessary in this case.

I think @etarena’s idea of segmenting all lamellae could work. Then you could do particle analysis of the identified lamellae. “Good” particles could be defined as those with high aspect ratio and/or area. “Good” images could be those with a high %area occupied by “good” particles.


Dear, thanks for all the suggestions, the problem is not the segmentation, but the quantification of picture. Segmentation could be a first step to evaluate the image.

In the ideal case is would like to segment the entire zones (black + white) where mostly good lamellea (so indeed, high aspect ratio) are present and the entire zones where globular lamellae are present. See my example picture (red =bad).

Would something like that be possible?


@BasWerbrouck, thanks for clarifying.

It sounds like you don’t need to quantify every single lamella, but would prefer to detect broad areas with high numbers of short or long lamellae.

In that case, maybe a texture-based segmentation could work, since the short lamellae tend to be clustered together. Using TWS, you could use the zones you outlined as training examples of the “bad” class.

The best thing would be to look at the links @etarena provided and play around with it. The user manual in the supplementary information of the paper is helpful too.

Hope it helps.


@BasWerbrouck - you can also look into using Extended Particle Analyzer from the BioVoxxel Toolkit - there are tons of other great tools in there you can look into.

And again - if you need any else - feel free to post again! We are here to help.




A colleague of mine (@ctrueden) also suggested you could try CellProfiler, specifically CellProfiler Analyst - apparently there are also some machine learning tools that you could try out for recognizing your regions… You can always post again on the CellProfiler Forum to see about getting some insight from them.

Again - hope this helps!

eta :slight_smile:


Good day,

your recently posted image with handdrawn red borders of “bad”-domains is helpful because it shows that the somehow unclear left/right = good/bad classification doesn’t perfectly apply.

Of course and as mentioned by others already, the task requires some kind of texture-discrimination. Basic investigations reveal that this can’t be achieved by using first and second order statistics. I also tried Haralick-measures that are available for ImageJ, but without success.

It appears possible to create a rather good binary version of your sample image that appears well-suited for processing by “Analyze Particles…” as proposed by others already. However, this approach moves the fundamental problem of domain-recognition to the evaluation and grouping of ROI-measurements …





Thank you for the support! The weka trainable segmentation give me quite good results, as can be seen at next picture.

THe globular(red) and lamellar (green) zones are quite nice separated and even the grain boundary ferrite (purple) can be taken into account.