5ttgen output: vis file and quality control


#1

Dear experts,

I may have quality issues with 5tt generation / segmentation. Providing the two screenshots: both processed through the same pipelines, both are the subjects from the same study (i.e., same acquisition). Interestingly, even with such drastic difference in 5tt images, the output from dwi2fod and tckglobal both look very well. I wanted to resolve my doubts before I proceed to more time and space demanding operations for further analysis. The question is - is it normal NOT to see clearly different types of tissue in the 5tt vis image, as in one of the examples I provide? Or will this error be reflected further?

Thank you!


#2

Hi Olga,

Interestingly, even with such drastic difference in 5tt images, the output from dwi2fod and tckglobal both look very well.

Neither dwi2fod nor tckglobal use a 5TT image as input. Did you mean tckgen?

The question is - is it normal NOT to see clearly different types of tissue in the 5tt vis image, as in one of the examples I provide?

I wouldn’t normally expect to see this kind of variation; but then again, it’s not entirely clear exactly what it is I’m looking at. The second image looks more like I would expect from 5tt2vis given the default tissue -> greyscale mapping. The first looks unusual, but could be achieved either: by playing with the intensity value options in 5tt2vis; or by putting the five tissue volumes into the incorrect order in the image. Ultimately the output image from 5tt2vis is too far downstream to be debugging: you need to be looking at the raw volumes within the 5TT image itself, ensuring that the image conforms to the format.

Rob


#3

Thanks, Rob!

Please bear with me, since I’m new to this. Just to make sure I understand, couple more questions:

Right, but I use msmt_5tt algorithm for the response function estimation, which makes me think that the integrity of 5tt generation will impact virtually every consequent step of the process. Am I correct?

It looks to be in the correct format, with brain tissue sum equal 1 and the non-brain 0. However, I noticed that the subcortical structures (CP, thalamus) were considered “non-brain”, giving me a value of 0 in the deep brain regions. Here is the 5tt image:

If that is the reason, is there a way to resolve this without 5ttedit and manual delineation of the subcortical structures to input into 5ttedit? To generate 5tt, I am using a high resolution T1, after I perform linear registration of it onto my B0. Freesurfer way is not desirable as we are interested in the structures the freesurfer likes to omit from its segmentation (brainstem, midbrain).

Thanks a lot!
Olga


#4

Right, but I use msmt_5tt algorithm for the response function estimation, which makes me think that the integrity of 5tt generation will impact virtually every consequent step of the process. Am I correct?

Sure. But those have additional degrees of separation between what’s going wrong with 5TT generation, how that might hypothetically affect response function estimation, then how that might hypothetically affect FOD estimation / global tracking. So there’s not a whole lot to be gained in looking at the outputs of dwi2fod / tckglobal, even if they are subjectively OK: For the sake of bug-squashing it’s best to focus on the 5TT generation step.

It looks to be in the correct format, with brain tissue sum equal 1 and the non-brain 0. However, I noticed that the subcortical structures (CP, thalamus) were considered “non-brain”, giving me a value of 0 in the deep brain regions.

That would be highly unusual… and a different result to what you first presented… But are you actually checking the sum of tissue volume fractions (i.e. mrmath 5TT.mif sum - -axis 3 | mrview -)?

Here is the 5tt image:

That screenshot only shows the first of the five volumes; in order to properly assess what’s going wrong, I’d need to be able to see all five volumes.


However, I’m going to take a stab in the dark anyway. It looks to me like for whatever reason, FAST is not providing the expected tissue classifications based on a T1-weighted image for your data.

In the first image you provided:

  • The darkest tissue is CSF, as expected;
  • The middle tissue (normally GM) contains both GM and WM together;
  • The brightest tissue (normally WM) has instead segmented a bright band outside the skull.

I can’t properly assess the second image, since I can only see one of the five volumes and not the greyscale visualisation image, but you can clearly see that this tissue component (volume index 0, so should be cortical GM) contains primarily WM, with a bit of cortical GM. This would seem consistent with what happened in the first image you posted.

So, things you can try:

  • Make doubly sure it is in fact a T1-weighted image that you’re providing to 5ttgen.

  • Try obtaining a better brain masking (make sure that bright band outside the brain is masked out), then re-run 5ttgen with the -mask option. It’s possible that whatever is providing the high intensities out there (which isn’t being removed by bet when it’s called within 5ttgen) is so severely skewing the image intensity histogram that it’s correspondingly invalidating the classification of fast's Gaussian mixture model components into brain tissues.

Rob


Labelsgmfix left thalamus not labelled
#5

Thank you so much! Masking did help a LOT. Here is the much improved VIS:


As well as opening a 5tt file in its own mrview (I could not understand why I can’t see the different volumes while opening 5tt in overlay :flushed:). Thanks so much, again, your wandering in the dark cleared up a few other things I was unsure of.

Yes, that’s what I meant - that despite somewhat okay looking further outcomes, I cannot trust those until I get to the bottom of my 5tt issues.

Olga