Closed selayav closed 3 weeks ago
Segmentation isn't something the model can do. I tested whole tumor detection/bounding box generation on BRaTS and got pretty good results, about what I got from nnDetection, but I didn't try to break the task down into detecting tumor components. I would guess necrotic core detection is plausible alongside detecting the whole tumor volume, but the rectangular boxes the model generates aren't going to be voxel-accurate. It may be useful for narrowing down the input volume for a downstream segmentation model, though.
I didn't get a chance to try this, but if you wanted to train on BRaTS I would try adding rotated examples like I describe doing for the abdominal dataset in the paper. My training runs suggested to me some uncaptured performance that needs additional augmentation to access.
Is segmenting brain tumors on MRI scans into edema, necrosis and enhancing tumor possible with MedYOLO?
Thanks!