Problem getting started with 2016 MSSEG dataset

[savvvyk]

Hello,

I'm modifying the spleen 3D segmentation example to work with the 2016 MSSEG dataset. The dataset consists of multiple input images (I'm using 3) and a single consensus label image. I'm currently using the pre-processed images but without the registered brain mask file. I'm currently experiencing two issues: first is my label image is shows as blank after the data loader but doesn't error (minor issue) and the second is AssertionError: ground truth has different shape (major issue, full error below). I would appreciate any help or suggestions. I've checked that onehot is set to true and I've tried remapping the labels from [1,2] to [0,1] but neither has made a difference.

epoch 1/600
/usr/local/lib/python3.7/dist-packages/monai/losses/dice.py:135: UserWarning: single channel prediction, softmax=True ignored.
warnings.warn("single channel prediction, softmax=True ignored.")
/usr/local/lib/python3.7/dist-packages/monai/losses/dice.py:144: UserWarning: single channel prediction, to_onehot_y=True ignored.
warnings.warn("single channel prediction, to_onehot_y=True ignored.")

AssertionError Traceback (most recent call last)
in ()
22 optimizer.zero_grad()
23 outputs = model(inputs)
---> 24 loss = loss_function(outputs, labels)
25 loss.backward()
26 optimizer.step()

1 frames
/usr/local/lib/python3.7/dist-packages/monai/losses/dice.py in forward(self, input, target)
155
156 if target.shape != input.shape:
--> 157 raise AssertionError(f"ground truth has different shape ({target.shape}) from input ({input.shape})")
158
159 # reducing only spatial dimensions (not batch nor channels)

AssertionError: ground truth has different shape (torch.Size([1, 128, 64, 128, 128])) from input (torch.Size([1, 1, 64, 128, 128]))

[Nic-Ma]

Hi @savvvyk ,

Thanks for your interest here.
For the first look at your error message, I feel it's related to your data shape, seems it's not correct.
Could you please help change all your AddChanneld transform to EnsureChannelFirstd transform and try again?

Thanks in advance.

[savvvyk]

@Nic-Ma, Thank you for your quick response. I tried switching AddChanneld for EnsureChannelFirstd and observed the same ground truth shape issue as before. The linked code file should be updated to reflect the latest.

[savvvyk]

I discovered that the label was not being included in the EnsureChannelFirstd command causing it to not have the channel array element. I have now been able to run the code without error (still needs much tuning however).