an approximate solution to compute mutual information about voxelmorph HOT 24 CLOSED

@soanduong and everyone,

I've added an experimental NMI loss to losses.py, please take a look.

from voxelmorph.

hi @argman , C+P from my answer just now on the other issue:

Yes, thank you for the code -- we're just in the middle of deadline season (MICCAI is on tuesday) and most of us are working on that. We'll be happy to get back to you after this!

from voxelmorph.

Hi @argman, is this implementation of MI correct? I tried to train VM with this loss and its not learning at all.
Hi @adalca, have you ever tried to train VM with MI loss? If yes, what is the result. It would be helpful for us if you share experience on MI loss with us.
Kind regards

from voxelmorph.

Hello @nazib, I have tested the code by @argman, but it requires much GPU memory to compute MI for bins > 10. It shows less value of MI when using same tensor as input nmi_gaussian(R, R, win=10, eps=1e-5)result = -0.19 than nmi_gaussian(R, T, win=10, eps=1e-5) with result = -0.42.

from voxelmorph.

@nazib @farhanone we have a implementation we've been sending around for NMI, feel free to email me and I will send it to you. We'll have it in the official doc once everything is reorganized.

@nazib yes we've tried the NMI loss for voxelmorph, and it worked well for registering different modalities within MR.

from voxelmorph.

@adalca Thanks. I have already sent you an email for the code...

from voxelmorph.

@farhanone you mean before or just recently? Anyway, if I missed it, just re-email please. It could have slipped through the cracks.

from voxelmorph.

@adalca I just sent you the email again. If there is some problem getting the email then, please send me at
[email protected]

from voxelmorph.

@farhanone done.

from voxelmorph.

@adalca many thanks

from voxelmorph.

Hi @adalca
Thanks a lot I will send you email now.
Hi @farhanone
Thanks for sharing your experience on MI loss. I also tried with this MI code and It was not good. Lets see how Adrian solved this issue. I will try to share my experience after retraining my model.

from voxelmorph.

Hi @adalca,
Could you please check your email.

from voxelmorph.

@nazib done

from voxelmorph.

Hi @adalca
Thank you so much for sharing the VoxelMorph source code. I think this is great work!
Could you please share the NMI implementation with me? I have been searching for it for weeks.
My email address is [email protected].
Many thanks!

from voxelmorph.

@junyuchen245 done!

from voxelmorph.

Hi @adalca, Thanks for sharing your great work!
I am also suffering in NMI implementation. Could you please share the implementation at
[email protected]
Thank you!

from voxelmorph.

Hi @adalca, thanks for making this open-source, I am wondering if you could also send me your implementation of NMI whenever you are free? It will save me lot of time.
[email protected]
Thanks in advance

from voxelmorph.

done. I'll try to add this to the losses.py since this is not a sustainable model :)

from voxelmorph.

Hi @adalca ,thanks for sharing the VM, I am training it with my dataset. I am wondering if you could send me your implementation of NMI if possible? I want to try it. My email is [email protected]. Thank you!

from voxelmorph.

Hi, @adalca , thanks for sharing your work. I am working on cross-modality registration, could you send me the MI loss when you are free, many thanks, my email address is
[email protected]
thanks again.

from voxelmorph.

Hi @adalca, thanks for publishing your work.

Would you mind to send the MI loss? My email is [email protected].

Thanks again.

from voxelmorph.

@adalca thank you very much for the sharing.

from voxelmorph.

Hi @adalca , thanks a lot for your work and for making it open source.

Could you please send the MI loss? My email is [email protected]

Thanks again.

from voxelmorph.

@lucasestini95 the mutual information is available directly in the github code.

We have one implemented in voxelmorph:
https://github.com/voxelmorph/voxelmorph/blob/dev/voxelmorph/tf/losses.py

We also have a newer, cleaner one implemented in neurite which will eventually take over:
https://github.com/adalca/neurite/blob/dev/neurite/tf/losses.py

from voxelmorph.