Closed TobiasKletter closed 4 years ago
In this case it would be easy, because the spindle poles are the brightest points in the tubulin image. However, very often this is not the case...
For example here, other points are brighter:
Do you have a plot to see how we systematically do with spindle length? I wonder whether we should just always go back 0.5 -1 micrometer towards to the centre from the highest derivative points. I think scientifically this would make sense, too.
The overestimation is only really pronounced in cells that tend to have round spindles (HeLa, HEK):
A plot that Nils made for HEK293 cells (really round spindles...):
A quick plot from a quality control that I did on mESC spindles:
I remember we spoke about doing the spindle segmentation first and then, within the mask, look for longest extend in the rough direction of the shortest-DNA-axis?
What I tried now is to look in a circular 3D region of a radius of 1 um around the DNA axis based (derivative based) for the maximum intensity and set this to be the spindle pole. This works perfect if there is a bright spot. If there is no bright spot at the end it works not yet...need another idea here, i.e. for the example image that I posted above.
The problem with the mask is that it usually extends beyond the spindle pole, because the spindle pole can be very bright.
Sometimes the brightest spots are like such:
Here, it works better. This is the spindle where you have the ground truth and the method would in fact produce those values, i.e. 9.6 um.
Ok, I think I found some compromise which for two cases looks quite ok, and it reproduces at least your manual GT of 10.8 for your example with the very bright poles:
And here the same method for a spindle without bright poles:
Interesting! Looks like this could in both cases correspond to the actual location of the centrosomes. I will test this after the upcoming update.
Some feedback (more to come...):
Input data here are again these really tricky round HeLa spindles carrying a gTub-GFP tag, so "Groundtruth" refers to the true centrosome locations. "Manually" means "as measured in the tubulin channel by eye"
This is definitely an improvement!
@tischi More feedback:
Checked on 30+ of my samples (mESCs) and it looks really good so far.
However, in a few cases when the DNA plate is visibly non-perpendicular to the spindle axis, pole localisation is thrown off:
Raw input & output:
https://www.dropbox.com/s/qqpu3tvi5t95qko/P0001-T0280--0002.zip?dl=0
https://www.dropbox.com/s/s6ek7ys1kabwf5p/P0001-T0304.zip?dl=0
https://www.dropbox.com/s/pd7gpp4tpiamo48/P0002-T0066--0004.zip?dl=0
Thanks for the feedback! When I find some time I will look into the examples you posted!
Hi! We got some feedback concerning spindle length:
12 samples (quite homogeneous, pretty interesting non-focused poles, and a bit under-sampled in z) were tested and half of them look pretty much perfect. Here's one example:
https://www.dropbox.com/s/taxlo1tfi8y3s18/Correct-t92-crop.tif?dl=0
https://www.dropbox.com/s/x2em1zqxq2ubiaj/Correct-t92.tif-out.zip?dl=0
Here's a max projection (The little appendage on the left is in all of the samples, so I think it's something biological and it doesn't seem to perturb the analysis ๐):
The other half looks like this:
https://www.dropbox.com/s/gp7csj1gt1w05cm/Incorrect-t82-crop.tif?dl=0
https://www.dropbox.com/s/aqoc4qb8e9cnqjz/Incorrect-t82.tif-out.zip?dl=0
The user labelled these cases as "makes mistake". The main concern here being, that the poles are too far off the boundary of the volume mask. To me it doesn't look too bad, to be honest, but I'm sure for some this will be the make-it-or-break-it point for the analysis. Some of our own spindles get under-estimated in that way (see the examples from the posts above) so I'm also invested in this. ;) I remember we discussed it at some point but might is make sense to take the boundaries of the mask into consideration? Since we also do it for spindle width, this might be more cohesive? Or maybe it would be helpful to output the "current" method AND a measurement derived via the spindle mask?
@TobiasKletter Can you give me the input data for the wrong one? I can check what happens there. Because from a simple line profile I would agree that the poles are not at the correct position, even to our current criteria.
[EDIT] I have it, I will have a look! What kind of microscopy is this?
I think these were done on a laser scanning confocal.
I am confused...
In BigDataViewer the spindle pole positions look much better than in the output image:
I'll dig more into it...
Still do not get it completely...
Turns out I was looking for two hours at the wrong image in BigDataViewer ๐ฆ Well, at least I triple checked all code during this exercise... In fact, they do look the same:
I think some issues here are
I will check this, either later today or tomorrow.
@TobiasKletter Could you just post here a few example of how the spindles typically look like for you? So I can think about an approach that would work for all cases? Thanks?
Live: https://www.dropbox.com/s/v1v7rnimp64o316/20190227_HighZoom--W0000--P0001-T0007--0000_42.tif?dl=0 Fixed (with bonus centrosome staining and lots of astral microtubules): https://www.dropbox.com/s/49mwwvqac2acg0m/R1EWT_Undiffd0_aTub_568_CDK5RAP2_647_001-1.tif?dl=0
The fixed example with stained centrosomes is actually pretty informing:
The spindle volume segmentation, however, is really nice ๐
Maybe looks like we really just somehow take the spindle mask to find the poles?! Would that always work in your experience?
I think the idea definitely worth the exploration! In most cases, the poles (as determined by some centrosome marker) would be just "touching the mask" from inside the spindle volume.
"Problem" with the mask is that we then will also find the little appendices:
We may be able to get rid of them by some aggressive morphological opening operations on the mask, but I am not sure that would be the point. I think here is becomes a bit unclear what one wants, biologically, isn't it?
Yes I anticipated this problem with these rather special spindles. The right pole looks pretty perfect. Since I'm not sure what these little guys are in the first place, I will ask the specialist...maybe it's even good like this. :)
@tischi Seems like we were green-lighted for this ๐
ok! So we include them? I just send you some code that excludes them... ๐
Tested the new code on some samples:
Live mESCs:
works nicely:
on Fixed with centrosome staining:
I think this is pretty good (however, the trend here is a slight over-estimation)
On zygote spindles:
It now overshoots a bit:
Could you send me the input data for the lower one? I can try to improve.
@tischi https://www.dropbox.com/s/gzxrzg1cl1xncwa/191003-cow44th-8003-M4MTBD488-H2B568-Gauss-Scene-01-t92-1.tif?dl=0 <https://www.dropbox.com/s/gzxrzg1cl1xncwa/191003-cow44th-8003-M4MTBD488-H2B568-Gauss-Scene-01-t92-1.tif?dl=0
On 16 May 2020, at 13:36, tischi notifications@github.com wrote:
Could you send me the input data for the lower one? I can try to improve.
โ You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/tischi/fiji-plugin-morphometry/issues/86#issuecomment-629632013, or unsubscribe https://github.com/notifications/unsubscribe-auth/AK5RGQY5R3NURY7TBHDGL7TRRZ3DLANCNFSM4IMGXS3Q.
@TobiasKletter Could you please try out this version? https://oc.embl.de/index.php/s/OSuM1F5qhEMkSXP
@tischi Works perfectly with live spindles of mESCs and the zygotes ๐
On the fixed samples with centrosome staining, a bit mixed:
Good one (the red is the staining):
Misses the mark a bit:
Something happened here (I suspect the fixed DNA is again the culprit):
I only tested 6 of these fixed samples with centrosome staining so far, but I am currently preparing a larger analysis with these (30+), to check if there is a pattern or not, and also if there are other issues popping up. I will finish this later today and give an update!
For the last one, I feel it already could not identify the DNA axis properly..
Here are the results for n=51 samples:
"Manual" = 3D distance of poles as judged "by eye" via the tubulin channel "Ground truth" = 3D distance as defined by centrosome staining "Plugin" = plugin output
The ones that, for one reason or another, misplaced poles entirely (n=6) are included here (probably the few short ones...). So here it seems like it's overestimating length again. However, when looking at the output images, they aren't that far off actually:
Seems it measures the "outside" edge of the dots, right? The line in the last one for example is 12 ยตm (that's what in the output table for that file). Ground truth in this case is 9.9 ยตm. I think if it was the "inner" edge or the center, it would be closer:
(Line here is 10 ยตm)
It measures the centre of the dots! If that does not agree with the number in the table, please tell me, then there is still a bug...
Yes, the number in the table always coincides with the outer edges (as seen in the screenshots above)...
Yes, the number in the table always coincides with the outer edges (as seen in the screenshots above)...
Allright, let me have a look...
@TobiasKletter Thanks a lot for spotting this. There was a bug indeed. Does this work better? https://oc.embl.de/index.php/s/yuWFDa9NGdV0pui
@tischi
I think it's corrected, but I think it's still not exactly in the center, It seems like one voxel off on each side? The lines correspond to the measurement in the table:
The general trend is still that it over-estimates spindle length by about 1 micron:
It's a bit unfortunate because for the live spindles etc. pole placement looks really convincing now. I think, to make the plot convincing, I could define as "ground truth" the outer edge of the centrosome staining...I have to think about this a bit
ok, please try this one: https://oc.embl.de/index.php/s/h5UZYTcK2ZU6wlc it works like this:
@tischi Looks really good so far! Average length is now pretty close to ground truth ๐
I will discuss the two versions from today with my colleagues and let you know asap.
After comparing and discussing more images, I favour the "second to last" version (https://oc.embl.de/index.php/s/yuWFDa9NGdV0pui).
This version really works great with live spindles. It is biologically consistent with the cow spindles too, where we wanted to place the poles at the outermost edge. Consistent with that definition, I re-measured the fixed spindles with "ground truth" staining, now instead of picking the center of the centrosome rather trying to pick the outer edge (and removing the anyway problematic ones, where the analysis is running into problems due to wonky chromatin...):
It's a bit closer and I'm more or less pleased now, but it seems like fixed samples aren't ideal for this particular benchmarking strategy (fixation really impacts spindle shape ๐ข)...I have to think of another solution.
Example with difficult-to-find poles:
https://www.dropbox.com/s/2sa36p9vvf7v6sq/20190812_HeLaKyoto_gTub-GFP_hoechst_SiRTub_005-1.tif?dl=0
10.86934796 | 9.517382256 | 13.280625 "Manually" | "Ground truth via gTub-GFP" | Plugin (v 0.5.5)