Is predicted images of registration method have same affine matrices as the affine matrices of fixed images

[JosianeUwineza]

To get registration pair(moving and fixed) images and masks(.nii.gz). I took fixed images and masks rotate them by 10 degrees and perform registration using monai LocalNet model. With this data I found that the dice score reached 97%, but when use Monai SaveImage where resample is set to False, the predicted image is at the same position as moving image buy visualizing it . when I checked affine of fixed image, moving image and predicted image I found that the moving and predicted image has same affine matrices. Is this mean that the monai LocalNet can't register images and mask which is slightly rotated? what I can do ?

Your helps will be much appreciated

Thanks

[wyli]

the output meta info was not updated by the network and workflow. I haven't tested it but please try something like output_image.affine = fixed.affine before saving output_image.

[JosianeUwineza]

@ wyli thanks for the suggestion, But I think if I use output_image.affine = fixed.affine before saving is like cheating, how am I going to know the registration model registered the image well? I wish there is the way that this affine matrix can be updated during the training so that the model during predict the image with correct affine

[wyli]

for the localnet, those non-linear transforms won't be precisely described by the affine, we may consider including deformation fields in metadata. all these are not implemented at the moment, once we have clearer user requirements/feature requests, we should start to design the end-to-end metadata-aware image registration workflows. currently, there are only essential components for registration tasks.

[JosianeUwineza]

I tried the GlobaNet on the 3D which I rotated by 5 degrees to get the registration pair, the dice score that I am getting is zero, does GlobaNet work on 3D data?

[wyli]

yes

[JosianeUwineza]

Any recommendation on how to train GlobaNet in a supervised manner?

[wyli]

I think if the goal is to train a supervised network to estimate the affine parameters, there's no need to use GlobalNet (and the globalnet doesn't support this kind of training). you can directly create a regressor to predict the transform, like self.fc_loc described in the pytorch tutorial https://pytorch.org/tutorials/intermediate/spatial_transformer_tutorial.html.