search

pnlbwh / CNN-Diffusion-MRIBrain-Segmentation

CNN based brain masking
Other
14 stars 10 forks source link

[Question] Why ANTs rigid registration instead of SyN #17

Closed tashrifbillah closed 4 years ago

tashrifbillah commented 4 years ago

Hi @suheyla2 , I understand that we do ANTs rigid registration between a given b0 image and reference IIT_mean_b0.nii.gz:

https://github.com/pnlbwh/CNN-Diffusion-MRIBrain-Segmentation/blob/clean-imports/src/registration.py#L11

Can you tell me the motivation for simplifying the registration to rigid instead of SyN? A given b0 image can be quite off the space of reference and hence SyN seems more relevant to me.

tashrifbillah commented 4 years ago

Do we have any reasoning for this?

tashrifbillah commented 4 years ago

The CNN predicts mask of a warped b0 image in MNI space. If the CNN is trained well in MNI space, it should be able to predict a good mask. So, how we warp the b0 image to MNI space is not of interest to the well-trained CNN.

But we want the mask in the space of b0 image. So, after prediction is complete, we need to bring the "good" mask back to the original space. If the registration is linear (rigid), then we can surely obtain its unique inverse. On the other hand, if it is non linear (SyN), we may not obtain its inverse uniquely.

To facilitate getting back to the original space, rigid registration is the preferred choice over SyN registration.