Confused about dimension order in MONAI and how to develop C++ inference with ITK

[7oud]

When using the LoadImage transform, both NibabelReader and ITKReader load nifti file as WHD data order, even ITKReader does a transpose for it, Although this dimension order is the same with some meta info in dicom, like spacing. But not for the best performance. Now, I use the MONAI trained model with ITK libraries in C++, it's very inconvenient. If using ITK functions, the data dimension order must be DHW, but when using model to inference, I have to tranpose the image buffer(data) to WHD, after inference, then tranpose the outputs back to DHW and write to the ITK image buffer. Specifically if using slide window inference,
the additional two tranpose and WHD order inference affect performance a lot. Is there some good methods to deal with this situation. Or can I train models with DHW data order with MONAI?

[7oud]

And MONAI is PyTorch-based, PyTorch use the DHW order, why MONAI doesn't choose the same one?

[wyli]

which version of MONAI is this related to? asking because recently there are some API enhancements in this area in both monai and itk-python

[7oud]

The lastest monai-weekly 0.7.dev2135 and monai 0.6, I dont find the method to train a DHW dim order using these 2 versions