Adding evidence of performance improvement by using test-time augmentation (TTA) in the tutorial notebook

[dongyang0122]

Current tutorial notebook about test-time augmentation (TTA) simply shows how to apply it for brain MRI segmentation in 2D planes. However, it is unclear the if the TTA method would improve the segmentation quality in practice.
https://github.com/Project-MONAI/tutorials/blob/master/modules/inverse_transforms_and_test_time_augmentations.ipynb

It is better to show the accuracy and quality (visual) improvement of segmentation masks and let users know how to properly apply the TTA method in practice boosting the performance.

[Nic-Ma]

Hi @rijobro ,

Could you please help take a look when you are back?

Thanks.

[rijobro]

Hi @dongyang0122 are you saying you'd like a comparison between the inference of a single image versus the mode of multiple random instances computed with TTA?

[dongyang0122]

My point is to justify the usage of TTA in practice. It would be nice to show the segmentation improvement in terms of accuracy and appearance.

[ahatamiz]

Hi @rijobro

I tried to use this TTA implementation for a segmentation application. After reading the source code, it turns out it expects raw input data dictionary as opposed to a PyTorch tensors ? ideally this is a post-transform which should also accept PyTorch tensors and then apply additional transforms.

I also agree with @dongyang0122. The tutorial can be improved by including a 3D use-case.