The article takes a while to get to the point (there are different class labels in the NIH database that - to a practicing radiologist - are visually identical, but several ML algorithms seem to distinguish them, and the author is puzzled by that) but is a very nice read for a layperson like me, because it gives nice background. The article is also very well written. Looking forward to reading the other articles on that blog.
In case anyone runs into this old thread via search: the author has a follow-up post with that conclusion, that the ChestXray14 dataset should be treated with skepticism.
The article takes a while to get to the point (there are different class labels in the NIH database that - to a practicing radiologist - are visually identical, but several ML algorithms seem to distinguish them, and the author is puzzled by that) but is a very nice read for a layperson like me, because it gives nice background. The article is also very well written. Looking forward to reading the other articles on that blog.
As far as I understand, the fact that the conditions are visually identical also means that the quality of training data is in question.
In case anyone runs into this old thread via search: the author has a follow-up post with that conclusion, that the ChestXray14 dataset should be treated with skepticism.