Adding New Data to nnUNet Training Set for Active Learning

[JiaGUO1997]

Hello,

I am currently working on integrating nnUNet for active learning. As part of this process, I need to add new data to the training set each time I update the segmentation. However, I’ve noticed that the preprocessing step alters the size of the imageTr/labelTr data.

I would appreciate your guidance on how to correctly add new data and their corresponding segmentation files to the training set, specifically to the imageTr and labelTr folders.

Additionally, I am uncertain about how to perform preprocessing before training using the nnUNetPlans.json file which defines all the things. Could you provide some insights or point me towards relevant documentation or best practices?

Thank you for your support!
Best regards,

[ykirchhoff]

Hi @JiaGUO1997,

I think doing it right, requires some coding, as nnUNet per default isn't optimally suited for active learning. Something you could do is using separate datasets for each stage, preprocessing them and then training on the specific datasets. This wouldn't require any coding, but isn't the best way imo to achieve this. If you want to preprocess more data using a previously generated nnUNetPlans.json, you can run nnUNetv2_preprocess ... -plans_name nnUNetPlans_active.json on a separate dataset and copy the plans file in there, but that then generates a completely separate preprocessed dataset, which is just preprocessed with the same plan.
Let me know if you have further questions or need some guidance how you can change specific parts of the code.

Best,
Yannick