Semi Supervised Learning and nnUNet

[FrancoisPorcher]

Hi nnUNet team, thank you for your amazing work!

I am currently trying to improve the performance of the nnUNet by adding some pseudo-labels, but the performance does not actually increase (in some cases it even worsens). For the supervised images I am using the Mindboggle dataset. with 61 images, and for the unlabelled dataset I tried to to add ~1000 images from the Aomic dataset.

Until now I generated the pseudo-labels, and added them with the real labels to the student model.
Any idea why the performance does not improve? Do you have advice to maximise the performance boost with unlabelled samples? Maybe add more data augmentations? Did you try training the model on the pseudo labels first and then fine tuning on the real labels?

Thanks!!