You know, trying different things and stuff
Description
Description
Event Timeline
Comment Actions
Multiview:
Similar to Jakobs segmentation method. 3 UNets, each trained with a different perspective. Little improvement on Hippocampus, no difference on BraTS, but worse performance on other datasets -> Max be due to anisotropy
Conclusion: Do not use this
Comment Actions
Dice reweighting:
Oversampling foreground makes the dice believe the foreground is more frequent than it actually is. Therefore, it does not punish false positives hard enough. I therefore tried to balance the dice by doing the following:
- collect class frequencies from dataset
- collect class frequencies from generated batches (where patches were sampled more frequently with foreground)
- compute corrective factors for the dice (basically reduce the influence of TP and FN in the loss function)
Result: As expected the dice now predicts less voxels as foreground, but the results are overall worse (look for reweight in description in leaderboard)
Conclusion: Do not use