>It's having a general understanding/view of the "baseline", aka healthy anatomy. This is something LLMs will never have
You're making the mistake of conflating AI with LLMs.
I don't think LLMs will reliably be better than a board of doctors. But an Expert System probably will (if it isn't already). That's literally what they were created for.
The biggest downside of LLMs IMO isn't the millions of Jules wasted on training models that are ultimately used to create funny images of cats with lasers. It's that all that money isn't being invested into truly helpful AI systems that will actually improve and save our lives, such as medical expert systems.
I am quite surprised that expert systems are not already used in this area (and others). As you say, this is exactly what they are meant for.
The nature of expert systems is to become experts on a system.
The reason you need a doctor, or more often, let's be honest, a good nurse, is because systems can fail in any one of 10000 as yet undiscovered ways. New nurses. New residents. New techs. And on and on and on. All the measurements you're feeding to the system are an amalgamation of the potential errors of a potentially different set of professionals each time you move a patient through the enterprise.
Full disclosure, my first startup was building PACS and RTP software back before AI reading was a thing. Current startup working across dental and medical. Rethinking the link between oral and systemic health. Partner has been in the C-suite of several hospitals over the past few decades and now runs large healthcare delivery networks.
The reason you can't hand things over to AI, is precisely because there are so many humans in the system. Each of whom are fallible. Human experts are quicker to catch it. Expert systems are not. At least not any ES or AI I've seen. And I've been going to, for instance, RSNA, for well over 25 years.
If you have an ES or AI in the system, you would naturally put the same professionals responsible for catching human screwups, in charge of catching AI and ES screw ups. Even if these AI's turn 100% accurate based on the inputs they are given, that professional would still be responsible for catching those bad inputs.
Example, it's never happened to one of my companies knock on wood, but I have seen cases of radiation therapy patients being incorrectly dosed. The doctor almost never was the one who miffed in the situation, but ultimately, s/he's responsible.
Why? Bad input should have been caught.
Another example, situations where you operate on the wrong side of the body because someone prepped the wrong leg. Surgeon didn't do the prep. Whoever did do the prep may have simply relied on the software. But the software was wrong. May have been anything. Point is, the team is good, but everyone just fell into too complacent of a pattern with each other and their tools.
Trust is good. Complacency is not.
The same will hold true for AI team members that integrate into these environments. It's just another "team member", and it better have a "monitor". If not, you're asking for trouble.
The "monitor" ultimately responsible for everything will continue to be the provider. Any change in that reality will take decades. (And in the end, they probably will not change the current system in that regard.)