U-Net, artefact detection, some history and some perspectives - January-March, 2021
Third archival post, covering a diversity of topics.
In early 2021, I went in a few different directions with the blog, after I had gotten the “pedagogical” content out of the way.
In “Fifty shades of U-Net”, I criticized (perhaps a bit harshly) conference poster sessions, and the mass of publications that seemed to boil down to “train U-Net, get state-of-the-art results, publish”, and — a point that would be significantly expanded in my thesis — how results that didn’t take into account the imperfect nature of digital pathology datasets should be treated with a lot more caution.
I expanded on the idea of imperfect annotations in “Artefact detection in Digital Pathology”, explaining in the process the results we got in our first published article, from the CloudTech'18 conference.
I then went in a bit of a different direction in “AI: better than your doctor?”, where I expressed my opinion that this statement, which can sometimes be found in science journalism, is often meaningless given that the standard by which we judge AI algorithms is, in most cases, human opinions.
Finally, as I was at the time writing a chapter of my thesis that looked at the origins of computer-aided pathology, I presented in “Computer in pathology: a long history” the Cytoanalyzer, a 1950s-era (failed) attempt at automated cytology.