Subject: Re: OT: The Future and AI
He didn't say that AI will ever replace radiologists (and he thinks that will never happen), what he DID say is that a good radiologist USING AI AS A TOOL could replace 10 (or 100) radiologists.
Well that does logically mean the AI WILL replace a whole bunch of radiologists, it just won't replace ALL radiologists. If one radiologist skilfully using AI can do the job of 10 radiologists, then you could say AI will not replace that first guy, but it will replace the others.
I actually think the impact will be more modest, for several reasons. One, it will take a lot of time, so a lot of it will happen by attrition, with less new radiologists recruited into the pool. Two, a lot of radiologists are doing 'interventional radiology', not just reading films, but doing things like inserting stents in bile ducts anfd stopping brain haemorrhages and biopsying breast masses. Some day, perhaps AI will do some of that as well, but in the early stages, it will be mostly reading and interpreting images.
But third, and most importantly, there will be plenty of work for the 9 people that are no longer needed for reading films. For instance, double reading of films has been shown to vastly improve performance, but a lack of available resources has usually meant that radiologic images are read by a single, fallible person. Double reading of mammographies is standard in many countries, but not in America, but there is nothing unique about mammography that in benefiting from double reading, and it could become a norm of practice not only for mammography but for many other routine diagnostic procedures. Screening for lung cancer is a huge new field where much lower risk people could be screened, if resources were available. Annual mammographies (the norm is biannual, in most countries.) Digital colonoscopies. Low dose whole body CT scans - probably useful, but where would you find the radiologists to read them?
More generally, productivity gains usually mean more output, not less workers, in any industry.
dtb