Invertd vs. BatchInverseTransform

[dianemarquette]

Hi!

I'm really glad that the MONAI team developed tools to invert transforms applied to the data. It's really useful for predictions! However, I'm a bit confused about how Invertd differs from BatchInverseTransform.

I've looked at the unet_inference_dict.py and the inverse_transforms_and_test_time_augmentations.ipynb tutorials. But, I still can't spot the real difference between both transforms.

Thanks in advance for the clarification :)

[Nic-Ma]

Hi @dianemarquette ,

Thanks for your interest here.

1. Currently, Invertd is based on BatchInverseTransform and provides more features: to_tensor, device, post_func, etc. And can execute for several keys together.
2. Invertd is a post transform, which can be composed with other transforms.
3. In next version, we will change Invertd to execute on channel-first Tensor data instead of a batch-first Tensor.

Thanks.

[dianemarquette]

Ok, thanks for the detailed explanation :)