search

DeepRegNet / DeepReg

Medical image registration using deep learning
Apache License 2.0
564 stars 76 forks source link

about traning_set'spacing #817

Closed YuqingCai0316 closed 2 years ago

YuqingCai0316 commented 3 years ago

Hi! Why the spcing of your training set is [1,1,1], has it been processed? What if each set of my training set has different spacing?

YipengHu commented 3 years ago

Hi @YuqingCai0316 did you mean voxel sizes/dimension? deepreg currently does not support anisotropic voxel sizes, largely due to the smoothness regulariser is somewhat sensitive to it. If you use affine transformation, then this is less relevant, otherwise (e.g. LocalNet with DDF prediction), it may be easier to re-process your data to isotropic voxel sizes (spacing). Whether it is [1,1,1] or [.5,.5,.5] is less important. Does this answer you query?

YuqingCai0316 commented 3 years ago

Hi @YipengHu Thanks for receiving your reply! My data set is generally anisotropic paired CT, such as spacing is:[0.7825,0.7825,5],The shape of each pair of my CT is also different, such as [34,512,512]and [50,512,512]... It comes from the preoperative and intraoperative scans of the patient. During the operation, there is interference from local interventions. It is also possible that the patient's twisting during the operation may be displaced or deformed. I need to register such data. At present, I use the ants framework. I want to use deep learning to further complete this project. How to deal with the different number of ct layers?Do you have any valuable comments ?Looking forward to your reply!

YipengHu commented 3 years ago

I would simply resample/crop them into the same image size and voxel size, wrt preserving regions of interest.

YuqingCai0316 commented 3 years ago

Hi @YipengHu： Thanks for your reply. If the training process is interrupted, Can I load the model and continue training？

YipengHu commented 3 years ago

Hi @YipengHu： Thanks for your reply. If the training process is interrupted, Can I load the model and continue training？

You should be able to.