How to train a SwinUNETR model on my own MRI dataset

[laizhui]

I want to train a SwinUNETR model on my own MRI dataset, using the code and model.pt file given at:
https://githubfast.com/Project-MONAI/research-contributions/tree/main/SwinUNETR/BRATS21
the shape of the MRI picture is [256,256,176], in_channels=1, out_channels=4(without the background)
How can I get started with the task and what should I do with the code?
Thank you very much!

[KumoLiu]

Hi @laizhui, thanks for your interest here.
You could refer to these tutorials to train a model using MONAI.
https://github.com/Project-MONAI/tutorials/blob/main/3d_segmentation/swin_unetr_brats21_segmentation_3d.ipynb
https://github.com/Project-MONAI/tutorials/blob/main/3d_segmentation/swin_unetr_btcv_segmentation_3d.ipynb

BTW, you can find more tutorials in this repo.

Hope it helps, thanks.

[laizhui]

Which one should I choose to process the MRI images?
And how can I change in_channels and out_channels?
Thank you very much!

[laizhui]

I try to use BTCV model to train, however, I got errors as follows:
Traceback (most recent call last):
File "main.py", line 243, in
main()
File "main.py", line 98, in main
main_worker(gpu=0, args=args)
File "main.py", line 228, in main_worker
accuracy = run_training(model=model,
File "/home/zwlai/SwinUNETR/trainer.py", line 169, in run_training
train_loss = train_epoch(model,
File "/home/zwlai/SwinUNETR/trainer.py", line 43, in train_epoch
logits = model(data)
File "/opt/Anaconda/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl
return forward_call(*input, **kwargs)
File "/home/zwlai/.local/lib/python3.8/site-packages/monai/networks/nets/swin_unetr.py", line 286, in forward
dec4 = self.encoder10(hidden_states_out[4])
File "/opt/Anaconda/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl
return forward_call(*input, **kwargs)
File "/home/zwlai/.local/lib/python3.8/site-packages/monai/networks/blocks/unetr_block.py", line 258, in forward
return self.layer(inp)
File "/opt/Anaconda/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl
return forward_call(*input, **kwargs)
File "/home/zwlai/.local/lib/python3.8/site-packages/monai/networks/blocks/dynunet_block.py", line 81, in forward
out = self.norm1(out)
File "/opt/Anaconda/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl
return forward_call(*input, **kwargs)
File "/opt/Anaconda/lib/python3.8/site-packages/torch/nn/modules/instancenorm.py", line 57, in forward
return F.instance_norm(
File "/opt/Anaconda/lib/python3.8/site-packages/torch/nn/functional.py", line 2326, in instance_norm
_verify_spatial_size(input.size())
File "/opt/Anaconda/lib/python3.8/site-packages/torch/nn/functional.py", line 2293, in _verify_spatial_size
raise ValueError("Expected more than 1 spatial element when training, got input size {}".format(size))
ValueError: Expected more than 1 spatial element when training, got input size torch.Size([4, 768, 1, 1, 1])

[KumoLiu]

Hi @laizhui, as you can see from the error message, the spatial size of your data is only 1, you should make sure that it is larger than 1, and you can check the size of the data you passed into the model.