search

ldocao / owkin_breast_cancer

Owkin challenge
0 stars 0 forks source link

run inference from trained nasnet mobile #13

Closed ldocao closed 5 years ago

ldocao commented 5 years ago

current AUC score is 0.63, which is very low

in principle though, we should be able to get a very high score : https://basveeling.nl/posts/pcam/

ldocao commented 5 years ago

I guess this is a problem between PCam and camelyon16 dataset, the magnification doesn't look like to be the same. The only information I have is "The PCam dataset contains 327,680 patches extracted from Camelyon16 at a size of 96 × 96 pixels @ 10× magnification" without further details.

https://arxiv.org/pdf/1806.03962.pdf

details provided here https://github.com/basveeling/pcam#details doesn't help

ldocao commented 5 years ago

Actually, the difference in score must be biased because I took the max. I need to check the score over tile predictions, and check if the score is indeed different or not. The score should be higher for tiles-level than for patient-level. If so, it might be worth exploring the possibility of constructing a score from tile to patient, smarter than just taking the max. Indeed, if the model is wrong over a single tile, ie it predicts high probability instead of low, then the whole patient is classified as positive.