DynUnet Transform BRATS 21

[nourmagde00]

Could any one tell me the Monai Transforms for BRATS 21 Dataset for DynUnet model ?
I already used these Transforms below and get --> RuntimeError: Sizes of tensors must match except in dimension 1. Expected size 10 but got size 9 for tensor number 1 in the list.

Thanks in Advance

[Nic-Ma]

Hi @yiheng-wang-nv ,

Could you please help share some ideas here?

Thanks in advance.

[nourmagde00]

Thank you very much for your response <3 @yiheng-wang-nv ,
I tried applying your instruction by setting roi size with the same value as the patch size (224, 224, 144) but I still recieved another error at the end of the notebook "RuntimeError: Sizes of tensors must match except in dimension 1. Expected size 10 but got size 9 for tensor number 1 in the list."

BRATS21 Jupyter Notebook.pdf

[yiheng-wang-nv]

Hi @nourmagde00 ,

If strides = [[1, 1, 1], [2, 2, 2], [2, 2, 2], [2, 2, 2], [2, 2, 2], [2, 2, 2]], the patch size should be divisible by 2 ** 5, which is 32. Therefore, size 144 can not be used. Could you please try to set patch_size = [128, 128, 128] and roi_size = [128, 128, 128]?

In addition, that's for posting this question, I submitted a PR to enhance the doc-strings of the network (see: #4102 ), and I will also enhance the DynUNet tutorial via adding the patch size check for the function get_kernels_strides.

[nourmagde00]

Mr @yiheng-wang-nv I really appreciate your responses and I'm very sorry for your effort with me ❤️❤️❤️,
I applied all of your instructions ,using roi-size =(128,128,118) crops important training parts from the brian. So, I had to add resized() in order to make the data with the shape (128,128,128), now I'm training my model
I just want to ensure that we are in the right way ,Could you check these 3 photos , it's all about MRI preprocessing for my dataset BRATS 21 so, my question : Is that preprocessing is a well preprocessing for my dataset ? I just afraid from losing time in the training and build a weak model .
I'd like to thank you again and again for your cooperation ❤️

[yiheng-wang-nv]

Hi @nourmagde00 , if Resized is used, you may need to think about how to convert the inference result back to the original shape, you may need to use Invertd transform. I suggest to use RandSpatialCropd(keys=["image", "label"], roi_size=[128, 128, 128], random_size=False), rather than using roi_size=[224, 224, 144] with the resize transform, and you can remove the crop foreground transform to make sure that the size after crop will not be changed.

You can also try to use RandCropByPosNegLabeld, which may peform better than RandSpatialCropd (it depends on the experiments, I'm not 100% sure and it is just an advice).

[nourmagde00]

Hi @yiheng-wang-nv @Nic-Ma I hope u be well,
I get a train loss equal nan when I train my model , I tried to solve it for several dayes but it does't work yet,
Could u check my notebook please ?
SEG NetJupyter Notebook.pdf

Thanks In Advance.