misalignment of labels after tranforms

[jakubMitura14]

Hello I had done transformations as indicated below and it leads to misalignment of labels, they are shifted to the left or to the right shift is big - multiple times bigger then the size of the change indicated by label.
transforms

            LoadImaged(keys=["t2w","hbv","adc" ,"label_name_val"]),
            EnsureChannelFirstd(keys=["t2w","hbv","adc" ,"label_name_val"]),
            EnsureTyped(keys=["t2w","hbv","adc" ,"label_name_val"]),
            AsDiscreted(keys=["label_name_val"], to_onehot=2),
            ResizeWithPadOrCropd(keys=["t2w","hbv","adc" ,"label_name_val"], spatial_size=spatial_size,),
            ConcatItemsd(["t2w","adc","hbv","adc" ], "chan3_col_name_val"),

What is my mistake?

[Nic-Ma]

Hi @jakubMitura14 ,

Could you please help share more information about problem?
Maybe give a sample image & label and simple program to reproduce it?

Thanks in advance.

[jakubMitura14]

Ok will work on it, thanks !

[ebrahimebrahim]

Are the image and label always the same size as each other to begin with? I.e. if an image has spatial size HxWxD then the corresponding label has the exact same spatial size HxWxD -- is it like this?

If not then I'd expect ResizeWithPadOrCropd to apply different operations to image versus label and potentially cause misalignment.

[jakubMitura14]

Thanks for responses @ebrahimebrahim and @Nic-Ma ! I @ebrahimebrahim seems that yoiu were right about the size. Hovewer here the problem is with the diffrence of final image size after sampling.

I perform resampling by

            Spacingd(keys=["t2w","adc","hbv","label_name_val"], pixdim=(
                1.0, 1.0, 1.0), mode=("bilinear","bilinear","bilinear","nearest") ),    

below there are printed t2w and label_name_val image sizes first before and later after changing spacing.
In all cases spacing seem to be the same (as it should). Hovewer some images after respacing get 1 voxel diffrence in size for example study 1000884

before resampling

 1000884 label size (1, 384, 384, 21)  imSize  (1, 384, 384, 21) labels_spac tensor([0.5000, 0.5000, 3.0000], dtype=torch.float64) im_spac tensor([0.5000, 0.5000, 3.0000], dtype=torch.float64) 

after resampling we see that in the dimension 1 label have now 192 and image 193

 1000884 label size (1, 193, 193, 61)  imSize  (1, 192, 193, 61) labels_spac tensor([1.0000, 1.0000, 1.0000], dtype=torch.float64) im_spac tensor([1.0000, 1.0000, 1.0000], dtype=torch.float64) 

full

 1000961 label size (1, 320, 320, 19)  imSize  (1, 320, 320, 19) labels_spac tensor([0.5625, 0.5625, 3.0000], dtype=torch.float64) im_spac tensor([0.5625, 0.5625, 3.0000], dtype=torch.float64) 
 1000715 label size (1, 320, 320, 19)  imSize  (1, 320, 320, 19) labels_spac tensor([0.5625, 0.5625, 3.0000], dtype=torch.float64) im_spac tensor([0.5625, 0.5625, 3.0000], dtype=torch.float64) 
 1000580 label size (1, 384, 384, 19)  imSize  (1, 384, 384, 19) labels_spac tensor([0.5000, 0.5000, 3.0000], dtype=torch.float64) im_spac tensor([0.5000, 0.5000, 3.0000], dtype=torch.float64) 
 1001301 label size (1, 320, 320, 25)  imSize  (1, 320, 320, 25) labels_spac tensor([0.6250, 0.6250, 3.5000], dtype=torch.float64) im_spac tensor([0.6250, 0.6250, 3.5000], dtype=torch.float64) 
 1000386 label size (1, 384, 384, 19)  imSize  (1, 384, 384, 19) labels_spac tensor([0.5000, 0.5000, 3.0000], dtype=torch.float64) im_spac tensor([0.5000, 0.5000, 3.0000], dtype=torch.float64) 
 1000657 label size (1, 384, 384, 19)  imSize  (1, 384, 384, 19) labels_spac tensor([0.5000, 0.5000, 3.0000], dtype=torch.float64) im_spac tensor([0.5000, 0.5000, 3.0000], dtype=torch.float64) 
 1000238 label size (1, 384, 384, 19)  imSize  (1, 384, 384, 19) labels_spac tensor([0.5000, 0.5000, 3.0000], dtype=torch.float64) im_spac tensor([0.5000, 0.5000, 3.0000], dtype=torch.float64) 
 1000107 label size (1, 384, 384, 21)  imSize  (1, 384, 384, 21) labels_spac tensor([0.5000, 0.5000, 3.0000], dtype=torch.float64) im_spac tensor([0.5000, 0.5000, 3.0000], dtype=torch.float64) 
 1000716 label size (1, 384, 384, 21)  imSize  (1, 384, 384, 21) labels_spac tensor([0.5000, 0.5000, 3.0000], dtype=torch.float64) im_spac tensor([0.5000, 0.5000, 3.0000], dtype=torch.float64) 
 1000426 label size (1, 384, 384, 21)  imSize  (1, 384, 384, 21) labels_spac tensor([0.5000, 0.5000, 3.0000], dtype=torch.float64) im_spac tensor([0.5000, 0.5000, 3.0000], dtype=torch.float64) 
 1000884 label size (1, 384, 384, 21)  imSize  (1, 384, 384, 21) labels_spac tensor([0.5000, 0.5000, 3.0000], dtype=torch.float64) im_spac tensor([0.5000, 0.5000, 3.0000], dtype=torch.float64) 
 1000345 label size (1, 384, 384, 19)  imSize  (1, 384, 384, 19) labels_spac tensor([0.5000, 0.5000, 3.0000], dtype=torch.float64) im_spac tensor([0.5000, 0.5000, 3.0000], dtype=torch.float64) 
 1000704 label size (1, 384, 384, 21)  imSize  (1, 384, 384, 21) labels_spac tensor([0.5000, 0.5000, 3.0000], dtype=torch.float64) im_spac tensor([0.5000, 0.5000, 3.0000], dtype=torch.float64) 
 1000580 label size (1, 192, 193, 55)  imSize  (1, 192, 193, 55) labels_spac tensor([1.0000, 1.0000, 1.0000], dtype=torch.float64) im_spac tensor([1., 1., 1.], dtype=torch.float64) 
 1000715 label size (1, 180, 180, 55)  imSize  (1, 180, 180, 55) labels_spac tensor([1., 1., 1.], dtype=torch.float64) im_spac tensor([1.0000, 1.0000, 1.0000], dtype=torch.float64) 
 1000618 label size (1, 180, 180, 55)  imSize  (1, 180, 180, 55) labels_spac tensor([1., 1., 1.], dtype=torch.float64) im_spac tensor([1., 1., 1.], dtype=torch.float64) 
 1000237 label size (1, 640, 640, 35)  imSize  (1, 640, 640, 35) labels_spac tensor([0.2813, 0.2813, 3.3000], dtype=torch.float64) im_spac tensor([0.2812, 0.2812, 3.3000], dtype=torch.float64) 
 1000386 label size (1, 193, 192, 55)  imSize  (1, 192, 192, 55) labels_spac tensor([1.0000, 1.0000, 1.0000], dtype=torch.float64) im_spac tensor([1.0000, 1.0000, 1.0000], dtype=torch.float64) 
 1000961 label size (1, 180, 180, 55)  imSize  (1, 180, 180, 55) labels_spac tensor([1.0000, 1.0000, 1.0000], dtype=torch.float64) im_spac tensor([1.0000, 1.0000, 1.0000], dtype=torch.float64) 
 1000356 label size (1, 384, 384, 19)  imSize  (1, 384, 384, 19) labels_spac tensor([0.5000, 0.5000, 3.0000], dtype=torch.float64) im_spac tensor([0.5000, 0.5000, 3.0000], dtype=torch.float64) 
 1000884 label size (1, 193, 193, 61)  imSize  (1, 192, 193, 61) labels_spac tensor([1.0000, 1.0000, 1.0000], dtype=torch.float64) im_spac tensor([1.0000, 1.0000, 1.0000], dtype=torch.float64) 
 1000704 label size (1, 192, 192, 61)  imSize  (1, 192, 193, 61) labels_spac tensor([1.0000, 1.0000, 1.0000], dtype=torch.float64) im_spac tensor([1.0000, 1.0000, 1.0000], dtype=torch.float64) 
 1000426 label size (1, 192, 192, 61)  imSize  (1, 192, 192, 61) labels_spac tensor([1.0000, 1.0000, 1.0000], dtype=torch.float64) im_spac tensor([1.0000, 1.0000, 1.0000], dtype=torch.float64) 
 1000345 label size (1, 192, 192, 55)  imSize  (1, 193, 193, 55) labels_spac tensor([1.0000, 1.0000, 1.0000], dtype=torch.float64) im_spac tensor([1.0000, 1.0000, 1.0000], dtype=torch.float64) 
 1000657 label size (1, 192, 193, 55)  imSize  (1, 192, 192, 55) labels_spac tensor([1., 1., 1.], dtype=torch.float64) im_spac tensor([1.0000, 1.0000, 1.0000], dtype=torch.float64) 
 1000238 label size (1, 192, 193, 55)  imSize  (1, 192, 193, 55) labels_spac tensor([1., 1., 1.], dtype=torch.float64) im_spac tensor([1.0000, 1.0000, 1.0000], dtype=torch.float64) 
 1001447 label size (1, 384, 384, 21)  imSize  (1, 384, 384, 21) labels_spac tensor([0.5000, 0.5000, 3.0000], dtype=torch.float64) im_spac tensor([0.5000, 0.5000, 3.0000], dtype=torch.float64) 
 1000512 label size (1, 384, 384, 19)  imSize  (1, 384, 384, 19) labels_spac tensor([0.5000, 0.5000, 3.0000], dtype=torch.float64) im_spac tensor([0.5000, 0.5000, 3.0000], dtype=torch.float64) 
 1001177 label size (1, 768, 768, 31)  imSize  (1, 768, 768, 31) labels_spac tensor([0.2344, 0.2344, 3.3000], dtype=torch.float64) im_spac tensor([0.2344, 0.2344, 3.3000], dtype=torch.float64) 
 1000107 label size (1, 193, 193, 61)  imSize  (1, 192, 193, 61) labels_spac tensor([1.0000, 1.0000, 1.0000], dtype=torch.float64) im_spac tensor([1.0000, 1.0000, 1.0000], dtype=torch.float64) 
 1000716 label size (1, 192, 193, 61)  imSize  (1, 192, 192, 61) labels_spac tensor([1.0000, 1.0000, 1.0000], dtype=torch.float64) im_spac tensor([1.0000, 1.0000, 1.0000], dtype=torch.float64) 
 1000272 label size (1, 384, 384, 23)  imSize  (1, 384, 384, 23) labels_spac tensor([0.5000, 0.5000, 3.0000], dtype=torch.float64) im_spac tensor([0.5000, 0.5000, 3.0000], dtype=torch.float64) 
 1001254 label size (1, 384, 384, 19)  imSize  (1, 384, 384, 19) labels_spac tensor([0.5000, 0.5000, 3.0000], dtype=torch.float64) im_spac tensor([0.5000, 0.5000, 3.0000], dtype=torch.float64) 

[ebrahimebrahim]

It could be a bug if the spatial size after ResizeWithPadOrCropd is not exactly the size requested from ResizeWithPadOrCropd... if you get a minimal example showing this happening then you can post it as an issue. Someone please correct me if I'm wrong and this wouldn't be a bug.

[jakubMitura14]

ok thank you !