Hacker News Reader

Comment by Fede_V | Hacker News Reader

I am surprised they only used ~2000 samples. That's a very tiny amount of data to train a decently sized CNN, even with data augmentation.

I presume using a pretrained net probably won't help you since the features found in pathology slides are so unlike those of natural images.

View on HN

I've done a bit with deep learning and medical images (brain, lung, skin). One thing that surprised me, but kind of makes sense, is the net seemed to perform well without a ton of training data. I think it might actually be easier for the model because the images are much more similar than in a "natural image" setting where there are more degrees of freedom in the subject itself.

Each image can be multiple gigabytes so it could be computationally difficult. Also those are probably all the lung cancer slides that TCGA has (edit: there seems to be 3500+ lung cancer slides at [1] so I was wrong. Maybe they're not all H&E stained)

[1] http://cancer.digitalslidearchive.net/

You also need to retain a significant set of slides the classifier wasn't trained with so you can verify that it does work for data outside the training set.

Somewhat surprisingly, the paper doesn't actually use deep learning.

Everyone is missing this. They used simple random forests and survival models with basic k-fold cross validation. There's no breakthrough in ML here!

Pattern recognition has been getting results for decades now on various problems and for some of those problems deep learning doesn't help that much.

So, making shit up is standard for this field of biomedical imaging.

Reply on news.ycombinator.com