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
551 stars 97 forks source link

low resolution scheme #18

Closed vankhoa21991 closed 3 years ago

vankhoa21991 commented 3 years ago

Hello, I have some questions about the low resolution scheme:

1/ When it is generated? 2/ How it's used later in training and validation?

Thanks for your help,

Van Khoa

mibaumgartner commented 3 years ago

Hi @vankhoa21991 ,

1) The trigger can be found here: https://github.com/MIC-DKFZ/nnDetection/blob/470ef323f90f14e120547d326bce46c5f90a9893/nndet/planning/experiment/v001.py#L186-L210 It is triggered if the 99.5 percentile of box sizes exceed the patch size.

2) The models are trained separately and the best model according to the validation results should be used for the final inference/prediction stage (currently, this is done manually by simply checking the validation score). In the current pool of data sets, this only happens for Kits19 since the tumors are quite large there. This approach also has its downsides since the low-resolution model might miss small objects while the full-resolution model can not capture large objects accurately. A better solution to this problem might be interesting for the future ;)

Best, Michael