Radiology has held the most potential for non-human analysis. Granted, if it gets too good, the insurance lobby will stop it.

Most of medicine though, not happening.

I'll hold my breath until it's revealed it's actually just test defeating and not really detecting anything.

Radiologist used computer aids almost as long as they've read digital scans, including various iterations of machine learning based ones.

This is huge, but "outperforming" is a misleading title. In practice, this this would be tool on a radiologist's monitor. As they're flipping through scans, it would flag this with a blinking notification, prompt them to squint at the spot, and likely confirm its warning (like existing tools already do in, say, highly detailed breast cancer scans).

And that's what machine learning should be. Continuously finetuned, laser-focused tools that are uncannily good at their specific job, checked by humans (so the job is faster and less tedious).

"Replacements" only in the sense that maybe you need fewer radiologists.

And don't let literal sociopaths in the tech world tell you otherwise.

Can it distinguish pancreatitis from actual malignancy? Pancreatitis is a huge risk factor for later development of pancreatic cancer, but you wouldn't want to approach every detected change as cancer.