Random blank labels in UNETR training

[btho733]

Hi,
I'm following 3D unetr tutorial (https://github.com/Project-MONAI/tutorials/blob/main/3d_segmentation/unetr_btcv_segmentation_3d.ipynb) to build a multi class segmentation pipeline. As suggested in that notebook, training patches are generated using RandCropByPosNegLabeld transform. Out of curiosity, I displayed the patches (both image and corresponding label) that are going into training loop.

As shown in the screenshot above, I am expecting patches similar to the ones in top panel(see Training 48/1000). Most of the training patches follow this. But I am observing that random patches turn out with just blank labels as in the bottom panels(see Training 49/1000). It looks like these training patches are skewing my results at the moment.

May I know why this occurs in the training ?
Any suggestion @ahatamiz , @yiheng-wang-nv , @wyli

Thanks
BT

[KumoLiu]

Hi @btho733, IMO it makes sense since you are using RandCropByPosNegLabeld if you don't want the patches with the whole blank labels, you can try to increase your patch_size. Or just use RandCropByPosNegLabeld with pos=1.
Hope it helps, thanks!