data modifications

[mauricesar]

Hi, i'm trying to use Spleen 3D segmentation with MONAI as base for vessel segmentation (task08_hepaticvessel - MSD), do you have any tips for that?

[Nic-Ma]

Hi @mauricesar ,

Thanks for your interest here.
I think you can refer to this DynUNet pipeline, it can support all the Decathlon tasks:
https://github.com/Project-MONAI/tutorials/tree/master/modules/dynunet_pipeline

Thanks.

[mauricesar]

Hi @mauricesar ,

Thanks for your interest here. I think you can refer to this DynUNet pipeline, it can support all the Decathlon tasks: https://github.com/Project-MONAI/tutorials/tree/master/modules/dynunet_pipeline

Thanks.

Thank you so much for your answer! Is there a possibility for me to use that notebook that i mentioned (Spleen 3D segmentation with MONAI, i'm using google colab) to do this task (vessel segmentation)?

[Nic-Ma]

I think you can also modify this tutorial for your task, just need to change below items:

1. input data channels, output prediction channels.
2. Random crop size.

Thanks.

[mauricesar]

Hi, i'm having problems with the modification, i keep getting an error: RuntimeError: cuDNN error: CUDNN_STATUS_NOT_INITIALIZED, i tried using another version of torch (multiple people suggested that) but it didn't work. Would you mind taking a look at the code? I'm pretty sure its something really really reeeally simple but I can't fix it. I would be incredible grateful: https://colab.research.google.com/drive/1Ny-ZUa1FOmZDryOAxwrhlxSOlKBQjJ7t?usp=sharing

[mauricesar]

You could do SpatialPadd to ensure that everything is at least (96, 96, 96). Following that you should be able to use RandCropByPosNegLabeld with crop size of (96, 96, 96).

Perhaps it's just whilst you're debugging, but I notice that you're currently using RandCropByPosNegLabeld with size 37. However, your network uses strides=(2, 2, 2, 2) so your images need to be divisible by 2^4 = 16. So 96 is fine, 37 isn't.

Thanks for your reply! I resized all my images to (160,160,30), when I try to execute the training with max_epochs with a different value (2, 3, 40. The current is 1, i just need the code to work) I get an error saying - RuntimeError: Sizes of tensors must match except in dimension 4. Got 16 and 15 (The offending index is 0).

Do you guys have any ideias about this? What is confusing to me is that I was used to not be able to run this part of the code, and now I can do it but just with 1 epoch. Do you guys have any sugestions about spartial_size and the roi that I should use?

[rijobro]

It's hard to say without the full stack trace as I don't know if the problem is in the collation (dataloader) or going through the network. Either way, you can print the dimensions of your data with the transforms applied and ensure they're constant.

(You're also using set_determinism, so it's possible that the randomness is such that by chance it works for the first epoch and then fails for the second. Perhaps if you were to remove it and run it multiple times, this wouldn't always be the case.)

[mauricesar]

It's hard to say without the full stack trace as I don't know if the problem is in the collation (dataloader) or going through the network. Either way, you can print the dimensions of your data with the transforms applied and ensure they're constant.

(You're also using set_determinism, so it's possible that the randomness is such that by chance it works for the first epoch and then fails for the second. Perhaps if you were to remove it and run it multiple times, this wouldn't always be the case.)

I tried removing set_determinism, it's still facing problems, here's the full message:
(the link, if you want to see it: https://colab.research.google.com/drive/1Ny-ZUa1FOmZDryOAxwrhlxSOlKBQjJ7t?usp=sharing)

[rijobro]

The problem is that your val_transforms will produce images of different sizes, since the last crop/pad transformation is CropForegroundd, and the background will change for each image. There is then no guarantee that the image will be divisible by factor 16.

You could add a DivisiblePadd after the CropForegroundd, or for convenience, CropForegroundd has the argument k_divisible, which could also set to 16. Let me know if this doesn't work.

[mauricesar]

The problem is that your val_transforms will produce images of different sizes, since the last crop/pad transformation is CropForegroundd, and the background will change for each image. There is then no guarantee that the image will be divisible by factor 16.

You could add a DivisiblePadd after the CropForegroundd, or for convenience, CropForegroundd has the argument k_divisible, which could also set to 16. Let me know if this doesn't work.

Hi, I'm sorry I took so long to respond - yes, it worked! Thank you so much for your help!

Let me ask you this final question = do you think it is possible for me to give and image as input and then have the segmentation? I want to try other vessels (lungs vessel segmentation, for example) using images that I have (but I don't have their labels). Do you have any tips for this task?