About using transforms.inverse

[leeh43]

Hi, thank you for the great work in monai. I am having a problem on using transforms.inverse in monai. As I want to resample back to the original space of the images (with the same size of original images) and output the image as nifti with its original resolution, a runtime error occurs when I used the transforms.inverse. What do you recommend for the next steps?

Here is the screenshot of the transforms and the inverse code:

Thanks a lot!

[Nic-Ma]

Hi @leeh43 ,

Thanks for your interests and experiments with MONAI.
Could you please help share your error message?

Thanks.

[leeh43]

Hi Nic,

Here is the error message about the inverse transforms. Thanks a lot!

[Nic-Ma]

I see your problem, the inverse function is not supposed to call in application directly.
You can use Invertd post processing transform or BatchInverseTransform API instead.
Refer to this tutorial if you are using MONAI v0.5.3:
https://github.com/Project-MONAI/tutorials/blob/master/3d_segmentation/torch/unet_inference_dict.py#L70

Thanks.

[leeh43]

I am currently using the latest version of the monai (0.6.0rc2). Meanwhile, I have another errror occurred after following the tutorial with the same test transformation above for input.

Here is my post transform for inverting it back to original space:

What do you recommend for the next step?

Thanks a lot!

[Nic-Ma]

Hi @leeh43 ,

MONAI v0.6 adjusted the logic of postprocessing, we are still updating all the tutorials for MONAI v0.6, so I would suggest you to use 0.5.3 so far.
Could you please help try it again?

Thanks.

[leeh43]

Thank you for the recommendation. I have created a new environment and installed 0.5.3 for monai. It seems like there is an import error for Invertd function.

[Nic-Ma]

Hi @leeh43 ,

Sorry I forgot that Invertd was not included in v0.5.3 release, you can use the BatchInverseTransform API for your case. Just follow the Batch inverse section of this tutorial:
https://github.com/Project-MONAI/tutorials/blob/master/modules/inverse_transforms_and_test_time_augmentations.ipynb

Thanks.

[leeh43]

Hi Nic,

I am so sorry for so annoying about the errors. After I change it to v0.5.3 release, I cannot import BatchInverseTransform and also cannot import it with v0.6.0rc2. Here is the error attachment:

[Nic-Ma]

Hi @leeh43 ,

Actually, I am sorry to lead you to the old version of MONAI because the examples and tutorials were not ready this morning.
But now we just released the tutorials for v0.6.0, so you can update to use MONAI v0.6.0 and refer to this example:
https://github.com/Project-MONAI/tutorials/blob/0.6.0/3d_segmentation/torch/unet_inference_dict.py

Thanks.

[leeh43]

Hi Nic,

Thanks a lot! it works now and I have one more small questions about generating multi-label segmentation map with the post_transform. For activation, I use softmax and normally we compute the max on the channel to output the multi-label segmentation, but how can I output a multi-label map with [1,2,3,4....,N] using post_transform?

Thanks a lot!

[Nic-Ma]

Hi @leeh43 ,

Sorry I don't quite understand your question, could you please help share an example program about your request?
Or @rijobro maybe you can help share some comments about the multi-label map with [1,2,3,4....,N]?

Thanks in advance.

[rijobro]

I think the question is saying that the output of the network will have the shape: B, C=N, H, W, [D]. I don't see why you can't omit the max (i.e., don't calculate the most-probable class), and keep the probabilities of all segmented classes.

[leeh43]

Hi Nic,

So currently, my goal is to output a segmentation map like this:

However, the post_transform lead to the softmax mapping with multiple channels which is like this:

Can we compute the max with the post_transform?

[Nic-Ma]

Hi @leeh43 ,

I think you can try the AsDiscreted postprocessing transform for your case:
https://github.com/Project-MONAI/MONAI/blob/0.6.0/monai/transforms/post/dictionary.py#L124.

Thanks.