How to register full volumes?

[alanqrwang]

Very interesting work. I noticed your code currently works only at the patch level, unless I'm missing something. How do you aggregate the patches to perform registration of full volumes?

[matteo-ronchetti]

Hi, tanks for appreciating this work.
Training is done on patches but inference can be done on full volumes. Since the model is fully convolutional you can feed it a volume of any size and obtain the embeddings for the whole volume.
Registration is done using this "full volume" embeddings. This repo doesn't contain the code for registration but only for training.

[alanqrwang]

Thanks for the response, that makes sense. One other question -- I see that your image patches are of size 51, and then you use the center crop of radiuses 3, 5, and 7 in computing LC2. Is there a reason you take the center crop and not just compute LC2 on the entire size-51 patch?

[huaibovip]

Hi, tanks for appreciating this work. Training is done on patches but inference can be done on full volumes. Since the model is fully convolutional you can feed it a volume of any size and obtain the embeddings for the whole volume. Registration is done using this "full volume" embeddings. This repo doesn't contain the code for registration but only for training.

Hello, can you provide a code snippet for registration to help me understand?

[matteo-ronchetti]

Sorry, but currently I only have an highly optimized implementation which is part of the ImFusion SDK and that I cannot share