I'm a relative beginner in deep-learning (understand by that, I'm doing my first kaggle competition right now, and I have loads to learn still) and I was just wondering something. Let's say you have pathology/biopsy tissue images from patients dying from a disease A and patients died from other causes (whatever causes actually but not related to disease A).
To date, I think we can say that nobody actually really knows what causes at the level of a biopsy the disease A. My idea, as my group could have actually a lot of these biopsies for both groups, would be to use them to fuel a neural network. Why would I do that? Biopsies images are rather complex images, and maybe some fine details are hard to guess for a human being, or maybe the sum of some details is actually important to tell whether disease A kills the patient or not. But again, I don't think anybody could come and say: on those tissue biopsy, the sign(s) for disease A are x, y, z.
My question then becomes a bit more theoretical: given the fact that you have enough data to actually give chances to the algorithm to find differences, is it a good idea to train a neural network without having actually any idea of what could differentiate the two groups? Do you know examples of such a strategy? How hard is it afterwards - in the case of a rather good accuracy - to understand what makes it so recognisable?