[Shape Error] UNETR patch size and image size mismatch

[aalhayali]

Hi everyone!

I have a quick question regarding an error I'm facing and I'm a bit confused on how to pursue it.

I am working on sagittal MRI images of size (512,512,7) (H,W,Slices). I applied the Resized transform and resized the image to (7,256,256), along with other transformations. The resultant original image are as follows:

The sizes for the original and transformed images are (2, 1, 7, 512, 512), and (2, 1, 7, 256, 256) respectively. When building the UNETR model, I used (7,256,256) model parameter and I'm getting the following error:

ValueError: patch_size should be smaller than img_size.

I am unsure why this error is being output although I am inputting the input image size generated from the Resized transform, as per the documentation img_size (Union[Sequence[int], int]) – dimension of input image.

I would appreciate any suggestions or pointers to fix that error!

model code:
model = UNETR( in_channels=1, out_channels=2, img_size=(7,256,256), feature_size=16, norm_name='batch', spatial_dims=3, ).to(device)

[Nic-Ma]

Hi @tangy5 ,

Could you please help take a look at this problem?

Thanks in advance.

[tangy5]

@aalhayali Hi, thanks for the question. Great to see UNETR is used for broader datasets.
From the description, I found your data is a bit special, there are only 7 slices. However, the default patch size of 3D transformer in UNETR is (16, 16, 16), which means the dimension at each axis should be at least 16. I have some suggestions:

1. You could try a 2D model for training your task, using slices as inputs.
2. Upsample, in slice thickness, make the slices numbers to 32 for training a 3D model. (add Spacingd transform)
   Thanks.

[aalhayali]

Thanks!

I will try the 2D approach as a baseline model, then will try to upscale it. My images have poor quality given the number of slice, with a wide FOV. I would ideally rather avoid step 2, but I will try it out for the sake of experimentation. May I know why the slice number you chose is 32?