search

MIC-DKFZ / nnDetection

nnDetection is a self-configuring framework for 3D (volumetric) medical object detection which can be applied to new data sets without manual intervention. It includes guides for 12 data sets that were used to develop and evaluate the performance of the proposed method.
Apache License 2.0
528 stars 88 forks source link

[Question] Impact of empty GT #252

Open Thibescobar opened 3 weeks ago

Thibescobar commented 3 weeks ago

Hello,

I'm wondering what to do with datasets where a proportion of samples has nothing to be detected (eg, no lesion).

Should I include this data with empty label maps (so that the model learns the background better?)? Or should I exclude this data because we still need at least one item to detect? Or does it make no difference?

Thanks in advance for your advice.

mibaumgartner commented 2 weeks ago

Dear @Thibescobar ,

yes, I would generally recommend to include healthy/negative cases in your dataset as well. While positive cases also provide background information during training (due to the patch based sampling) there can always be other changes in not annotated structures which might differ between a negative and positive case.

In previous studies we have also observed, that detection networks are far less likely to make mistakes (FP) in negative cases rather than in positive ones.

Best, Michael

tbskbr commented 2 weeks ago

Thank you very much !