Could use "DHW" dimension order with ITKReader?

[7oud]

Is your feature request related to a problem? Please describe.
There are different data storing order in Nibabel and ITK, WHD in Nibabel and DHW in ITK, Now MONAI uses the only dimension order WHD both with nibabel reader and ITK reader, paritcularly do data transpose in ITKReader. However, when using MONAI trained model in C++ with ITK library, addtiional transposes are needed, It's terrible and WHD order is not optimal performance.

Describe the solution you'd like
Could choose different dimension order when using Nibabel or ITKReader?

[wyli]

The LoadImage transforms also support customised readers. Would it be feasible to extend the ITK reader to handle this spatial convention?

Also, I think itk returns a view instead of actually allocates a new array for the transpose https://blog.kitware.com/convert-itk-data-structures-to-numpy-arrays/ any chance that we can make the transpose cheap by simply changing the view cc @thewtex

[wyli]

This is somewhat related to #2891 perhaps we should consolidate the discussions.

[7oud]

@wyli If using customised readers, can the transforms work correctly, such as monai.transforms.Spacing?

[wyli]

yes, it'll work as long as these properties are defined: https://github.com/Project-MONAI/MONAI/blob/dev/monai/data/image_reader.py#L233-L240

[wyli]

thanks for the comment, this pull request adds a reverse_indexing option to ITKReader. Models trained with reverse_indexing=True should be compatible with the ITK c++ spatial indexing convention.