Open timqqt opened 3 years ago
Usually public datasets are rather small. Maxpooling MIL works best with larger datasets. How many samples you need depends on the task. From my experience, at least several hundreds are necessary to have reasonable performance. The main issue though is that with such small datasets it is not possible to show generalization performance to real life clinical practice.
I looked through the paper. I am wondering why there is no experiments of MIL training on public dataset and testing on public dataset? Thanks.