Best practice training data different slice direction (axi/sag/cor)

[emilljungberg]

Hi all,

First, big shout-out to the nnUNet team for fantastic software. So easy to get started and it works like a charm right out the box. I find it rare for that to be the case for research software, thanks!

I'm working on a segmentation task where I have data that is acquired in different orientations with thick slices (1.6 x 1.6 x 5 mm). I have ground truth segmentation for an equal distribution of axial, sagittal, and coronal training data.

If I want to use my network to predict segmentation on any orientation (axi,sag,cor), what is the recommended best practice? Should I resample training and inference data to say 1.6 mm isotropic resolution and use the 3d_fullres network? If I use the 2d network, will it slice the data in the "correct" direction automatically? Any advice/suggestion is much appreciated.

Thanks a lot, and again amazing software!