Hacker News Reader

Comment by VikingCoder | Hacker News Reader

Yeah, I worked with radiologists... What they can see in x-ray, CT, MR, PET, ultrasound... it's mind-boggling.

The most mind-blowing thing is just how much those skills rely on tacit knowledge and subconscious pattern recognition.

There is really a lesson to be learned here in our age when data-based everything is en vogue. Our decision-making shouldn’t be data-driven, it should be data-informed.

I disagree. Hasn't it been shown that AI outperform radiologists on recognizing diseases from images?

Not really. On very narrow datasets, from very specific machines. When they tried to generalize (the way a radiologist does) across multiple hospitals and multiple machines from different manufacturers, results are not that good.

can you point to any reports that provides these studies and outcomes?

Anecdata but I broke two ribs and got an x-ray, a prescription and was sent home. While I was in line at the pharmacy the x-ray folks came and found me and sent me to the ER - one if their other techs looked at it and saw a punctured lung. None of the subsequent doctors/surgeons could see it on that x-ray but yeah, CT scan showed I had a punctured lung.

This one comes to mind. The discussion section is helpful:

> Despite the many interesting discoveries in pathology AI, translation to routine clinical use remains rare and there are many questions and challenges around the evidence quality, risk of bias and robustness of the medical AI tools in general

https://www.nature.com/articles/s41746-024-01106-8

And yet, as a software developer or security researcher or whatever, we can do the same thing; for example a milliseconds delay in SSH leading to the uncovering of a years-long ploy to undermine security (although the diff that introduced the issue flew under the radar), or recognising an issue from the sounds a modem or hard drive makes.

Reply on news.ycombinator.com