search

cwmok / C2FViT

This is the official Pytorch implementation of "Affine Medical Image Registration with Coarse-to-Fine Vision Transformer" (CVPR 2022), written by Tony C. W. Mok and Albert C. S. Chung.
MIT License
131 stars 3 forks source link

Can we use the affine augmentation params as train supervision #13

Closed Zhenghao97 closed 5 months ago

Zhenghao97 commented 9 months ago

Hi, thanks for your impressing work~

If the moving image is obtained by data augmentation, does that mean we have the groundtruth label of the affine params?

So can we leverage the supervision of affine params during training phase, such as l1 loss?

Looking forward for your reply, appreciate it~

cwmok commented 9 months ago

Hi @Zhenghao97,

If the moving image is obtained by data augmentation, does that mean we have the groundtruth label of the affine params?

Yes, it will work for the intra-subject registration task where the images are pre-aligned after image acquisition. But in our paper, we focus on inter-subject registration or atlas-based registration, in which the groundtruth affine transformation of the dataset doesn't exist. Therefore, we will need a similarity measure to supervise the training.

Thanks for your interest in our work!