Open TobiasKletter opened 4 years ago
Thanks!
...and would it already help to automatically determine that there is an issue and report this in the measurements?
To me this has high priority, because when it happens in the DNA channel, it hinders all downstream analysis steps
Theoretically yes, since the images analysed for the paper are pretty tidy and curated 😉 It's just sad that I have to throw out a lot of samples in my other data (which by default are quite busy) because of this. And to generalise this, I guess a lot of users would like to analyse spindles in tissue slices or organoids which all suffer from high density etc.
...and would it already help to automatically determine that there is an issue and report this in the measurements?
Yes I guess it could help when quality controlling the measurements
Ok! Then I will try to work on it this week.
I think back then we only worried about the DNA. Is it Ok if I now first think about the cases with the bright Tubulin neighbours?
I think this is the point where they touch:
So we need something that prevents the tubulin mask from growing into the neighbouring cell....
I think back then we only worried about the DNA. Is it Ok if I now first think about the cases with the bright Tubulin neighbours?
Yes let's do it step by step! In the meantime I can dig up again what we got with the CATS classifier and the DNA issue...
The variance filter seems to see the spindle edges pretty well...
Questions:
Questions:
- Would it be OK to systematically use a higher threshold for the spindle?
Maybe a tiny bit? I think we are pretty robustly close to ground-truth with the current setting, no?
Should the refined spindle poles be always inside the spindle mask?
- Currently my algorithm also allows them to be outside (...and it would help if we forbid this)
Let's test this!
I think we are pretty robustly close to ground-truth with the current setting, no?
close to ground-truth regarding which measurement?
I think we are pretty robustly close to ground-truth with the current setting, no?
close to ground-truth regarding which measurement?
Spindle volume, first and foremost. Although admittedly, there's no proper ground-truth to measure against 😭
I sent you a mail with both datasets analysed with higher threshold and forcing the spindle pole to be within the spindle mask.
What do you think?
They both look pretty good!
I'm curious! Shall I upload some more examples to test this on?
yes
https://www.dropbox.com/s/ahoupo47jf84sqt/20201006_R1E309_TubGFP_KATNB1_D5_005-2.tif?dl=0 https://www.dropbox.com/s/f1f1r09um4ekmdi/20201006_R1E309_TubGFP_KATNB1_D5_011-2.tif?dl=0 https://www.dropbox.com/s/z7ihucydgd9gsjx/20201006_R1E309_TubGFP_KATNA1_D0_016-2.tif?dl=0 https://www.dropbox.com/s/2npo4dqn8k9ge8i/20201006_R1E309_TubGFP_KATNA1_D2_013-1.tif?dl=0
and a couple of images from other datasets to check how the new thresholding looks like in general: https://www.dropbox.com/s/vpi0hz0npz0rhji/20200826_HeLa309_TubGFP_Hoechst_CDK5RAP2_PFAGA_NoMounting_002-1.tif?dl=0 https://www.dropbox.com/s/09r7dzl68q559ib/20200605_HeLa_MCB309_034-2.tif?dl=0 https://www.dropbox.com/s/27fzpdw9j82lopv/Crop_20200728-174144_HighZoom--W0000--P0001-T0079--0000.tif?dl=0
all images have the same channel order...
This one had a different channel order I think: 20190827_T0248_A.tif
The others look promising I think, I sent you a mail.
...and this one had an issue already in the DNA channel: 20201006_R1E309_TubGFP_KATNB1_D5_005-2.tif
I opened an issue for this: https://github.com/tischi/spindle3d/issues/11
This one had a different channel order I think:
20190827_T0248_A.tif
True! The accidental segmentation looks pretty interesting. The DNA segmentation applied on the tubulin signal looks really good, I think. But maybe it's because it kind of looks closer to the "old" spindle thresholding and I'm just really used to that one.
The others look promising I think, I sent you a mail.
My main concern is the following:
Since biologically we try to segment polymerized tubulin (=spindle) vs non-polymerized tubulin (rest of cell) I think we're now ignoring a considerable fraction of the former, no? Taking into account resolution and some optical blurring and so on, I would still threshold at about 1250 in this example. I'd be curious to see how the masks would look when dilated after thresholding...
I'd be curious to see how the masks would look when dilated after thresholding...
Ok, but how much should we dilate? ...always think about writing a paper or giving a talk and convincing your biophysical audience 🙂 I am not saying we should not dilate, but we should at least try to come up with something that we can defend biophysically. Any ideas?
...and could we remove this guy from the test set? I think there is weird stuff going on:
Any ideas?
No I'm actually no longer convinced myself, because in this way we would just make it larger for the sake of making it larger and not because we necessarily would capture the pixels we want to.
A bit more defendable would be the opening operation after thresholding with a lower threshold that we discussed, no? Alternatively, is 3D watershedding an option? Probably also difficult because of the two halves of the spindle...
...and could we remove this guy from the test set? I think there is weird stuff going on
Haha yes this one doesn't look to healthy
I am trying more stuff today.... In case I make some progress: time for a zoom this afternoon or tomorrow morning?
I am trying more stuff today.... In case I make some progress: time for a zoom this afternoon or tomorrow morning?
Yes!
Just pasting here some images for us to think about it. Given this image, what is a biophysically sensible way to define the spindle threshold?
What I did experiment with here is:
Tubulin intensity cytoplasm based threshold = ( medianCytoplasm + 5 * madCytoplasm )
In principle I like this a lot as it just says: "significantly brighter than in the cytoplasm".
From a biophysical point of view maybe the best we can do.
Obviously the 5
is arbitrary. We could also just make it much higher...
The values in this image are:
Tubulin intensity cytoplasm (median +/- mad): 741.0 +/- 46.0
Tubulin intensity cytoplasm (mean +/- sdev): 755 +/- 76; numPixels: 5345
Tubulin intensity cytoplasm based threshold (median + 2 * mad): 833
Tubulin intensity cytoplasm based threshold (median + 3 * mad): 879
Tubulin intensity cytoplasm based threshold (median + 4 * mad): 925
Tubulin intensity cytoplasm based threshold (median + 5 * mad): 971
Tubulin intensity spindle (mean +/- sdev): 1284 +/- 428; numPixels: 2913
Spindle threshold = median + 5 * mad: 971.0
In order to avoid the merging with the other cell we would have to go as high as 1200 with the threshold..
...maybe you could play a bit yourself with above image in ImageJ to see whether you can come up with a suggestion?
I added a description of the new Otsu based method: https://github.com/tischi/spindle3d/blob/master/README.md#methods
I think it works nicely! Feels coherent with the Otsu-based DNA volume segmentation Before finally committing to it I would like to test it on a larger testing set, if that's fine, to rule out that there are unexpected effects happening.
Feels coherent with the Otsu-based DNA volume segmentation
Yes, that is what I thought as well.
I uploaded the new version to Fiji. Happy testing.
So far, the images I tested with the new Otsu thresholding look good. 👍
A couple of thoughts and suggestions, though:
With the old thresholding this wasn't as evident, but there are now more cases where I get the impression there is a proper void in the center of the spindle:
I think this is a valid segmentation of the "polymerized tubulin mass" in the cell and we would want to measure the volume of this mask, currently named "Spindle_Volume_um3". However, I think biologically it would be interesting to know, whether it's this mass that is scaling with e.g. DNA or cell volumes, or it's rather the geometrical 3D extent of the entire spindle structure, regardless of the distribution and density of microtubule mass within this structure.
In the most extreme case, there might be a situation where all microtubules are concentrated "on the edge". They still could generate a big enough structure to successfully distribute the chromosomes, but we would measure a pretty small volume (A, blue):
With the generally lower threshold in the previous versions, this was not really an issue but now I think we would learn the most if we would quantify both. So if possible, I would suggest we additionally measure the convex hull volume like in (B) based on the thresholded mask in (A). What do you think?
(A follow-up to #9 )
Bright Tubulin signal in neighbourhood: https://www.dropbox.com/s/jgzjs17xy4xwarj/20201006_R1E309_TubGFP_KATNA1_D0_016-3.tif?dl=0
https://www.dropbox.com/s/qzdte49u7531zds/20201006_R1E309_TubGFP_KATNA1_D1_023-1.tif?dl=0
Bright DNA (and tub) signal in neighbourhood https://www.dropbox.com/s/obo2v4yzaahl1oq/20201006_R1E309_TubGFP_KATNA1_D4_021-1.tif?dl=0
I can usually solve these issues by extensive cropping around the cell of interest but maybe there is a more elegant solution?
Maybe some morphological filtering can already help in these cases?
Also, in general, how valid is background subtraction here? It seems the algorithm has an easier time when the background is subtracted and I have had some success with it in difficult images, but my feeling is one shouldn't mess around with the pixel values?