Erronious activity on "Membrane Detection" classifier tool
MIB is a high-performance software package for image segmentation
Brought to you by:
ajaxel
Menu
▾
▴
4 Attachments
#17 Erronious activity on "Membrane Detection" classifier tool
New
2019-01-21
2019-01-19
No
The Membrane Detection tool works as intended, but stalls on the 8th section if the "predict dataset" function is used on a stack. Sections 1 to 7 are processed successfully (they generate a selection that can be assigned to a material). Section 8 is never completed. (as discovered below, it also seems to segment unrelated files occasionally)
I tested this with 3 different datasets (ranging 10-250 slices with different pre-processing on each stack), and 3 different classifiers. It doesnt seem to matter if the classifier is new or loaded. Images attached to illustrate the issue.
The most recent attached example used a 250 section stack with preloaded classifier. Processor usage dropped to background levels on the 8th section at the expected time (~45s after previous section), but a selection was not generated and the process did not proceed.
I also attempted to "predict shown slice" on the 8th section after the process stalled, and nothing happened.
This is surprising....
shutting down and reloading MIB apparently didn't enable me to use "predict shown slice" on the 8th section, even without attempting the entire stack.
I also tried segmenting the file with the first 50 sections removed... it still stalled at the 8th section, but now it seems to be displaying membranes from an entirely unrelated image! see "bug3b.png"
Is the classifier file uniquely tethered to the dataset?
I dont understand how it could possibly be segmenting information from an unrelated file right now.
Strange.
I'm using MIB version is V2.40 31.08.2018 deployed version
Running it on a computer that is 5-10x more valuable than my car, all drivers are up to date.
Windows 10 home edition 64bit (10.0, build 17134)
Please let me know if you need more information. I love the work you're doing, and appreciate the youtube videos. I've only been using MIB for a week, and I already want to add it to my routine workflow.
Hello Ben,
I just quickly tested on a small stack and there were no problems.
Would you be able to send me (http://www.biocenter.helsinki.fi/~ibelev/)
the first part of that dataset and the model used for training, I can test
it then?
Best regards,
Ilya
On Sat, Jan 19, 2019 at 10:47 AM Ben Padman bpadman@users.sourceforge.net
wrote:
Related
Tickets: #17
Thanks Ilya, I keep trying to post the files but Sourceforge keeps failing.
Sorry about the delay. I'll see if there is another way for me to send.
I want to send you:
The first 10 images from the stack
The classifier file generated after training
The successful FM output for slice 7 and the failed output for 8
I generated the training model in the middle of my stack, so I'll need to find a way to reduce the filesize to one slice.
I have identified a work around for the issue:
I was able to segment beyond image number 8 after isolating the data into a new folder, with its own RF_Temp directory. All images processed without an error. See attached screenshots.
I think its related to the unusual activity shown in Bug3b.png above, where membranes from an unrelated dataset were being segmented in the current dataset.
The issue is much less urgent now, but do you need any more information from me to determine why the membrane detection classifier was segmenting an unrelated image?
Hello Ben,
thank you for the datasets! I was able to load the images (tif stack) and
the classifier. In my case the prediction worked quite well. I have not
noticed any issues (please note that I was using the Matlab version of MIB).
However, I noticed a possible problem with your *.fm files. The FM-files
contain variable features used by the classifier for the prediction; these
files should roughly be the same size. In my case, all fm-files were about
448Mb. The files that you've send me were 452 and 111 Mb, so I guess there
is a problem with features in file "slice_000008.fm". Deleting of the file
should fix the problem.
I just checked the actual file and it seems that it is twice binned version
of the original file. Perhaps during the tests you've binned the dataset.
Also for this specific task I recommend to use Ilastik, which offers
possibility to select features that will be used for the classification.
Selection of relevant features will speed-up the process. It is very easy
to do from MIB: what you need to do is to save your images and the
segmentation in HDF5 format (preferably with a header). After that you will
be able to load both of these files in Ilastik for the prediction.
Let me know how it goes.
Best regards,
Ilya
On Sun, Jan 20, 2019 at 3:19 AM Ben Padman bpadman@users.sourceforge.net
wrote:
Related
Tickets: #17
Thank you for all your help!
Everything is now working as expected.
I will have a look at Ilastik as well.
All the best