search

CCI-Bonn / HD-GLIO

Automated deep-learning based brain tumor segmentation on MRI
Apache License 2.0
26 stars 7 forks source link

Question: Does Brats 2020 dataset needs preprocessing to run HD-GLIO? #3

Open saruarlive opened 3 years ago

saruarlive commented 3 years ago

Hi, @FabianIsensee et al, Does Brats 2020 dataset (image dimension =240,240,155) needs preprocessing (fslreorient2std, flirt registration, etc ) to run HD-GLIO-PREDICT as Brats 2020 folks have already done several processing steps? Looking forward to hearing from you.

Thanks

FabianIsensee commented 3 years ago

Hi, BraTS images do not need additional preprocessing. You can use them as they are. Note that HD-GLIO will not perform well on the very very old BraTS cases (the ones that are totally blurry). But it will work on the others. Best, Fabian

saruarlive commented 3 years ago

Thanks, @FabianIsensee for your reply. I have a follow-up query. Regarding those very old Brats subjects (blurry images), do you mean that a self-trained nnUnet (or similar methods) model trained on Brats 2020 data (369 MRI images) may perform better as compared to HD-GLIO even though HD-GLIO is trained on 3220 MRI examinations?

FabianIsensee commented 3 years ago

It's a distribution shift thing. A model trained on BraTS data will perform better on BraTS-like data, but worse on others. And the other way around. Since HD-GLIO was trained on many more cases I would recommend using that over a BraTS model when working with modern-ish MRI scanners. If you need a necrosis label (which HD-GLIO does not provide) then going with the BraTS model is the only option. We won the BraTS 2020 challenge and our inference docker is publicly available. You can use that as well (https://hub.docker.com/r/brats/isen-20/tags?page=1&ordering=last_updated, instrictuions somewhere on the BraTS homepage ^^)