Generative Visual Rationals

Hey there,
I stumbled upon a very fascinating paper today.

https://pubs.rsna.org/doi/10.1148/radiol.2018180887

It presents a newly way to use GANs to visualize what a neural network learns while trying to predict congestive heart failure (CHF) or not. here is a more technical paper outlining the details of the algorithm: https://arxiv.org/abs/1804.04539

CHF is very interesting to study because many image features contribute to the diagnosis CHF and the pathology is not as discrete and local as maybe lung nodules, which makes it very challenging to visualize what the algorithm learns as features.

The paper is particularly impressive because the author is a first year radiology resident and not a “classical” deep learning researcher. I wonder if he took the fastai course :smiley:

I’m currently working through it and will probably write a blog post that summarizes it and I will try to implement parts of it in fastai/pytorch. It will be a challenge because I only finished the first part course recently but with stackoverflow and github nothing is impossible :wink: