Closed mmlitvak closed 4 years ago
Did it fail to run, or did it fail to generate a meaningful output? If the latter is true, can you give me any details about your data?
It appears to be running. But nothing happens. No errors, no confirmation. No new files added and old files rewrites.
We have mouse brain in DICOM that has been transferred to nifti format by MRIcroGL 1.2.2
I tried segmenting that volume and the program seems to run. Have you tried using the full path for your target volume? On the other hand, I am quite sure these images look different enough that it will be difficult to obtain a meaningful output. However, if you have a small number of labeled volumes it would not be hard to retrain MU-Net to adapt it to your target volumes.
Yes. I tried full path. Unfortunately I don't know how properly running program should look like. Should it open *.nii images after it was started or save segmented images automatically? Maybe run it with FSL or MathLab?
It should look something like this:
It's really strange that we simply do not get any error in your case. I changed a few lines in the code which might help with compatibility with windows systems, could you please update the code (git pull) and try again?
I just did and nothing changed. I'll try to reinstall it under Linux
I managed to get a hold of a windows machine. Have you tried making sure your python executable is indeed python3?
It might simply be called python:
python runMU-Net.py
Thanks! Now it seems to be working! But I've found that NVIDIA card is missing on my machine. Is it crucial to have it? I've tried useGPU False, but this did not help.
Can you please update and try again?
Current version seems to work without GPU support, closing this issue.
It worked! Thanks! Unfortunately we weren't able to achieve any valuable segmentation. Maybe the problem is in our MRI machine setup. Any specific recommendations?
That is very likely, what is your MRI setup? Scanners from different vendors can produce data distributed in a significantly different way, and even choosing a different coil can significantly alter the overall visual outlook of the image, so these are all complicating factors. However, we can consider retraining the network. Depending on your data it could take a limited amount of labeled samples (even something like 5-10) and it might be a good investment if you have a large number of volumes to label.
Il giorno mar 4 ago 2020 alle ore 09:01 mmlitvak notifications@github.com ha scritto:
It worked! Thanks! Unfortunately we weren't able to achieve any valuable segmentation. Maybe the problem is in our MRI machine setup. Any specific recommendations?
— You are receiving this because you modified the open/close state. Reply to this email directly, view it on GitHub https://github.com/Hierakonpolis/MU-Net/issues/2#issuecomment-668397316, or unsubscribe https://github.com/notifications/unsubscribe-auth/AILCFDM5K5OVQM5ECACKVTTR66P4ZANCNFSM4PKHGGOA .
We would really appreciate if you will be able to train your network based our MRI setup (I have 7 mice in vivo MRI of quite good quality) or please, provide an instruction how to do so. Thanks in advance!
Is 7 the total amount of mice you want to label? Retraining the network requires data. Perhaps a small number of samples, but we would need some data to be already labeled. If you have a large number of mice to process, now or in the future, it might well be worth the effort. Before looking for a different way to adapt this, if you only have 7 scans to process and no future plans for more of them, it might make more sense to use atlas-based segmentation and manually correct some mistakes. If you plan to add more scans to your study, these could later be used to then train a network and scale it to the hundreds of mice.
On our github page you can find a link to a tutorial on how to train a similar network (on publicly available data) if you are interested, and I would be happy to work with you and provide a straightforward way to easily train this network, but let's try to make sure this is also the correct course of action for your specific situation.
Il giorno mar 4 ago 2020 alle ore 11:53 mmlitvak notifications@github.com ha scritto:
We would really appreciate if you will be able to train your network based our MRI setup (I have 7 mice MRI of quite good quality) or please, provide an instruction how to do so. Thanks in advance!
— You are receiving this because you modified the open/close state. Reply to this email directly, view it on GitHub https://github.com/Hierakonpolis/MU-Net/issues/2#issuecomment-668472013, or unsubscribe https://github.com/notifications/unsubscribe-auth/AILCFDKEMPLNTIUUFDKDH43R67D7LANCNFSM4PKHGGOA .
We are interested in using your algorithm in our research. I successfully installed all the software as instructed, but I wasn't able to achieve any result. Could you please provide a bit more detailed instructions on how to run it. Thanks in advance!