Deformation field show nothing in the plot

[mikami520]

Issue

Hi, I am working on the deep atlas with 128x128x96 CT scans which are normalized to [0,1]. When I used the trained registration network to make inference, the deformation field plot showed nothing. I checked the value of deformation field and its range is between [-1,1]. Also, I used the deformation field learned from the registration network to warped the moving segmentation, the result seems only to have translation which are shown below (green line shows only translation and small rotation). Do you have any idea about this issue? Thanks in advance

Result

deformation field plot

Comparison between warped segmentation, moving segmentation and target segmentation

[ebrahimebrahim]

When I used the trained registration network to make inference, the deformation field plot showed nothing

Maybe regularization is set too high?

Also, I used the deformation field learned from the registration network to warped the moving segmentation, the result seems only to have translation which are shown below

Often registration is done in two steps: first an affine or euclidean transform, then a "deformable registration" map.
The type of registration model produced by the DeepAtlas tutorial notebook is shown to work well as the deformable registration step; i.e. doing affine registration separately beforehand may be necessary to get good results.

I don't know why you're only getting this translation and small rotation. Maybe you can share how you are warping segmentations.

[mikami520]

Try smaller and smaller lambda_r until the deformation fields become reasonable to you after training converges

You mean try smaller lambda_r for the normalized Bending Loss or non-normalized one?

[ebrahimebrahim]

Either. For example if you prefer normalized, where you have tried lambda_r = 5e-4, try also lambda_r = 5e-5, then lambda_r = 5e-6, etc.

[mikami520]

Either. For example if you prefer normalized, where you have tried lambda_r = 5e-4, try also lambda_r = 5e-5, then lambda_r = 5e-6, etc.

Except decreasing the lambda_r, are there anything that I can improve to get rational displacement field?

[ebrahimebrahim]

You can try a different image similarity loss function. For example maybe a simple mean squared error loss can work for you, if maybe LNCC is causing problems for training. You can try changing how much you focus on the anatomy loss (warped segmentation dice score) versus image similarity loss. Maybe in the case of your dataset, one of those is much more helpful than the other, or maybe you will discover an error in one of them that is causing it to break training.

[mikami520]

You can try a different image similarity loss function. For example maybe a simple mean squared error loss can work for you, if maybe LNCC is causing problems for training. You can try changing how much you focus on the anatomy loss (warped segmentation dice score) versus image similarity loss. Maybe in the case of your dataset, one of those is much more helpful than the other, or maybe you will discover an error in one of them that is causing it to break training.

Gotch that. Yeap, MSE is also used in VoxelMorph. I also trained the same dataset with VoxelMorph, the deformation field was much better than the deep atlas and I used MSEfor similarity loss, but regularization is also different from what I used in the deep atlas