search

biomedia-mira / istn

Image-and-Spatial Transformer Networks
Apache License 2.0
108 stars 26 forks source link

Huge deformation in bspline transformation #10

Closed pmgautam closed 3 years ago

pmgautam commented 3 years ago

I was running the code on provided 2d data in affine and bspline mode. Affine transformation seems to preserve structure. However in Bspline, deformation is huge and the image structure is also not maintained. Training was done without changing any parameter from the repo. Original image original_image

Warped image warped_image

What could be the reason behind such deformation? I compared the bspline code with airlab and looks like diffeomorphism and Diffusion Regulariser are not added. I was hoping adding them might make the deformation look reasonable. Is there anything else that needs to be taken care of.

Thank you for the help.

bglocker commented 3 years ago

Yeah, I am not sure we ever tested the 2D B-splines on this synthetic data which was specifically constructed as a demonstration example for the concept of structure-guided registration.

You could try to increase the control point spacing, as the default setting might not be great.

My general advice would be to opt for a different transformation model than B-splines. We have recently used SVFs within learning based image registration which gives much better results.

pmgautam commented 3 years ago

Thank you very much for the response. Can you please refer me to the papers that might be useful?

bglocker commented 3 years ago

We have been recently working on an extension of the ISTN framework, and the pre-print is available here

I would also suggest to check out the airlab library and voxelmorph