2D segmentation runtime error

[witness97]

Hi MONAI team and thanks for providing such amazing package. I try to run 2D_segmentation Unet training code on my own dataset, but it got the error

RuntimeError: Caught RuntimeError in DataLoader worker process 0.
Original Traceback (most recent call last):
File "/usr/local/lib/python3.7/dist-packages/monai/transforms/transform.py", line 48, in apply_transform
return transform(data)
File "/usr/local/lib/python3.7/dist-packages/monai/transforms/croppad/array.py", line 327, in call
self.randomize(img.shape[1:])
File "/usr/local/lib/python3.7/dist-packages/monai/transforms/croppad/array.py", line 315, in randomize
self._size = fall_back_tuple(self.roi_size, img_size)
File "/usr/local/lib/python3.7/dist-packages/monai/utils/misc.py", line 174, in fall_back_tuple
user = ensure_tuple_rep(user_provided, ndim)
File "/usr/local/lib/python3.7/dist-packages/monai/utils/misc.py", line 131, in ensure_tuple_rep
raise ValueError(f"Sequence must have length {dim}, got {len(tup)}.")
ValueError: Sequence must have length 3, got 2.

The shape of image in my dataset is (512, 512, 3), I just wonder why this error would happen and how to turn my data to fit the code. Thanks.

[Nic-Ma]

Hi @witness97 ,

Thanks for your interest here.
The difference between the example data and your data is that: your data is channel-last and has 3 channels, the example data has no channel dim.
So you can change the AddChanneld transform to EnsureChannelFirstd transform and try again.

Thanks.

[Nic-Ma]

Hi @witness97 ,

We have 2 sets of transforms: dictionary-based and array-based.
The name of dictionary-based transforms ends with a "d", like EnsureChannelFirstd.
You should not compose the dictionary-based transforms and array-based transforms together.
I would suggest to make your data as a dictionary and use all dictionary-based transforms.
You can refer to: https://github.com/Project-MONAI/tutorials/blob/master/3d_segmentation/spleen_segmentation_3d.ipynb

Thanks.

[witness97]

Hi @Nic-Ma ,

Really thanks for your quick reply which helps me a lot. I can successfully run the 2D segmentation in dictionary-based version. However, I face the same problem as Q&A #1455 that the mean dice I get from the Unet is always larger than 1. I tried the approach in #1455 to turn the crop areas but it seems not work well. My dataset are images with multi-classes and size (512,512,3), I just wonder if this incorrect mean dice is caused by the dataset or something else?

Thanks.

[Nic-Ma]

Hi @rijobro , according to your training experience, what do you think the possible reason that mean dice > 1?

Thanks.

[rijobro]

Hi @witness97 , I'll try to put together a multi-class segmentation tutorial, I think this will be useful for quite a few users. I'll let you know once it's done. see here: Project-MONAI/tutorials#247.

[rijobro]

Sorry, as @Nic-Ma pointed out, this the brats segmentation tutorial is already multi-class.
I would suggest checking where your code differs from that tutorial, especially in the post-processing, and the calculation of the Dice score.

If you still can't get it working, feel free to send a minimum working example (using a public dataset), and we can have a look.