class imbalance with batch size 2, half/half or still sampling 1/3 and 2/3?

[Joeycho]

Hello all,

Nowadays, I am struggling with class imbalance issue because my own UNet pipeline predict somehow.. oversegment the lesion. I expect segmenting only upper and lower jaws bone, but it segmented the whole skull with my own pipeline, even though I mimiced nnUNet dataloader. By the way, nnUNet did not oversegment, just with some connected small pieces.. and small scattered pieces if I don't leave the largest component.

Maybe I missed something, but in 2 batch size case, nnUNet, prepare one, for sure, including the label and the other one is from random location in image, right? So, I think half/half, but another explanation in the nnUNet paper, sampling 1/3 including the label and 2/3 from random location. I am not sure what it means in 2 batch size case.

Is there anyone who has idea about it?

[Joeycho]

Plus, is there any reason for pick other ones from random location, rather than no label legion or way less including legion, by sampling the region in which the center point is not in the label? Sampling other ones from random location can still be the patches which are visible with the majority of labels. And in this case, I think we need to consider to enforce undersampling for tackling class imbalance.