[owenthejumper]

Almost nobody here is a doctor, and it shows...over diagnosis, over treatment, those are all terms that doctors learn about in medical school.

Image segmentation is a real problem, and achieving better precision is a good goal. The "golden" standard these days is likely https://github.com/wasserth/totalsegmentator, if someone can make it even more accurate, that would be very very good. But yet again, there are infinite amounts of variations in human bodies, which means even the best models focus only on segmenting known organs, and leave anything unknown alone.

[deadeye]

This is more true than people realize. My wife and I got full body MRI scans.

Both showed "possible" medical issues. My though was "Great, I have a baseline, in two years I'll get another one and compare".

My wife on the other hand got a bit obsessed about her results and had what was probably an unnecessary procedure to biopsy something, which turned out to be benign.

I suppose you could argue that another way...better safe than sorry...but the stress that is caused by known uncertainty vs unknown uncertainty can be too much.

[gbrito]

> ...unnecessary procedure to biopsy something, which turned out to be benign.

The point here is many issues can't be resolved safely with a biopsy or minor procedures, so one ends up under serious risk of a major surgery for something that would never cause any damage.

Plenty of people die this way. If not, one might even thank his doctor for saving his life afterwards.

[andrewla]

Not a doctor, but the overdiagnosis concern is at the intersection of three phenomena:

1. Imaging is expensive, just in dollars and time, even without analysis

2. Imaging is not without impact -- CT scans, especially full body scans, expose the body to ionizing radiation

3. Imaging is time-consuming

The net result of these means that full body scans are difficult to interpret. If a doctor given a patient complaint suspects a condition that is sufficiently non-specific that a full-body scan is required, then the scan will be interpreted through the lens of the known progress of the differential diagnosis. And typically these scans must be done without a healthy baseline, so minor findings in this context might have significant diagnostic power when combined with history or other findings.

But on a healthy patient, minor findings are very likely to be noise, because we don't have a great deal of experience with scans of healthy people, for the reasons above.

This technology, if it pans out, gives a way of inverting 1, 2, and 3. If every healthy doctor visit includes one of these scans, then the medical field gets experience interpreting them, and more importantly, when new symptoms occur, previous scans can be compared to determine whether a particular finding in the current scan is new or has changed.

[nicoburns]

If something like this became commonplace (and accurate enough), then it could be fantastic for research: enabling us to map out what variations are common and which aren't in a way that hasn't previously been feasible.

[khafra]

In the dark ages of machine learning, researchers tried to fit natural language into a defined, human-curated taxonomy.

It kinda worked, for a reasonable amount of stuff; but failed quite a lot of the time, and there's an extremely long tail of things that would have been pragmatically impossible to ever address with that method--indeed, without adopting an entirely new, unsupervised model of language, continuous in places where the old way was discrete.

[saghm]

It's not unreasonable to think that the level of acceptable risk for "the language model parsed my text wrong" is in average much higher than "the medical model misdiagnosed my condition". You can probably come up with scenarios where a language model behaving unexpectedly would have drastic consequences if you imagine them hooked up to automatic systems where they have immediate control over actions that can't easily be reversed, but like, that's why it's a bad idea to use them like that, and they're the exception rather than the rule. It seems plausible that scenarios like that for medical models are a lot closer to the norm than the exception, in which case the tolerance we have for them "filling in the gaps" incorrectly would need to be much smaller.

[spongebobstoes]

this doesn't need to be a diagnostic model, just a data source for existing doctors