jacobzhaoziyuan / MT-UDA

[MICCAI 2021] Official Implementation for "MT-UDA: Towards Unsupervised Cross-modality Medical Image Segmentation with Limited Source Labels"
MIT License
27 stars 12 forks source link

Checkpoint selection #5

Open msseibel opened 1 year ago

msseibel commented 1 year ago

In the "LE-UDA: Label-Efficient Unsupervised Domain Adaptation for Medical Image Segmentation" paper, it is rightly mentioned that target labels are not available for hyperparameter selection.

I just want to clarify how you perform checkpoint selection. Am I correct that you create fake target images to obtain a proxy measure for assessing the expected performance for the real images target images?

If so, did you by any chance compare this approach to supervised checkpoint selection? I.e. using labels from the target domain?