apply_inverse_transform function inverts transformed image to different size instead of original size?

[ruby0101]

Hi, I have tried to invert a transformed image to the original size with apply_inverse_transform function. But after I applied the function, size is different than the original size as shown the below screenshot. How do I apply the inverse of all applied transforms and obtain my original image size?

[fepegar]

You cannot. At the moment, Resample is not invertible. Why do you want to apply the inverse transform?

[ruby0101]

I want to change a predicted label image to original image size. The predicted image size is (1, 48, 60, 48), original image size is (1, 512, 512, 86). Could I just apply resize transformation to it with no loss? Or how can I put back the predicted image to original size?

[fepegar]

You can use tio.Resample(original_image)(prediction).

However, I recommend you train your images in the original resolution. Images in the tutorial are downsampled so it can be run faster. You can use patch-based training for that.

[ruby0101]

Thank you so much, you are right, data types are different. Is there any available function in torchio to convert data types? How could I change data types of NIfTI images (int16 to float32)?

[fepegar]

You can use the data property and set_data method of the Image class to modify the stored tensor data, e.g. image.set_data(image.data.byte()).

[ruby0101]

Thanks a lot, I have another issue. I have tried patch-based training but prediction result as follows:

What is the problem with my code? Could you tell me please what transformation should be applied for patch-based training? I applied the following transformation.

[fepegar]

There are many potential causes of poor results. I suggest you follow Andrej Karpathy's advice: http://karpathy.github.io/2019/04/25/recipe/

I can help you with questions related to TorchIO, but I'm unfortunately not available to supervise all steps of your project, sorry!

[ruby0101]

Thank you so much for your all support. You can close this Q&A as answered status. Thanks.