fastMRI sample in SSIM loss

[Khoa-NT]

Hi, thank you for adding the SSIM loss.
Can I ask one question for this term:

A fastMRI sample should use the 2D format with C being the number of slices.

In my knownledge, you mean reshape the 3D volume from (B,C,W,H,D) to (B,C*D,W,H) ?
However, the current code only support C=1, therefore, the loss can't run with (B,C*D,W,H).

Can someone explain this to me? @mersad95zd

[mersad95zd]

Hi @Khoa-NT, thanks for your question.

A fastMRI sample should use the 2D format with C being the number of slices.

In my knownledge, you mean reshape the 3D volume from (B,C,W,H,D) to (B,C*D,W,H) ?

If you're working on the fastMRI dataset, then no need to reshape because fastMRI's default shape is (B,num_slices,W,H).

However, the current code only support C=1, therefore, the loss can't run with (B,C*D,W,H).

Good point, since there are datasets where C varies across samples, the current implementation works with C=1 and the loss function should be called for each channel, separately. We'll correct this in the docstring.

Does this address your issue?

[a-parida12]

I have a 2D image of batch size =128 and channel 3. What is the correct way to use the SSIMLOss

from monai.losses.ssim_loss import SSIMLoss
criterion = SSIMLoss(spatial_dims=2)
loss = criterion(images, pred, data_range=images.max().unsqueeze(0))
# images.shape == pred.shape == torch.Size([128, 3, 64, 64])

In the above code, I get a RuntimeError.

RuntimeError: shape '[1]' is invalid for input of size 3