Hacker News Reader

Comment by TuringNYC | Hacker News Reader

No, Im not an MD, but my co-founder was.

Globally, hundreds of thousands of radiologists have been trained over the years have have collectively achieved generally consistent practices. Radiology is pattern matching and a set of very complex decision trees. They arent magic, because we consistently churn out more practitioners who achieve the same consistent outputs given inputs.

Anyone trying to improve things is like every other scientist, they aren't trying to figure out the entire decision tree or every single thing, they are trying to chip away on complex problems little by little.

I also strongly disagree with "radiology is more art than science" because if it was, radiologists wouldn't be able to agree on diagnoses.

View on HN

Trivial radiology is pattern matching. Radiology for which you need a radiologist for is not. Yes, you could replace part of radiology and some radiologists with algorithms today. Unfortunately, those are also mostly the cases where we don't need radiologists.

I'm sorry, but in many cases and even many important diagnoses that impact the patient, 2 radiologists that aren't allowed to communicate will sometimes reach widely different conclusions. Additionally, if you've interacted with radiologists you were certainly able to see that their analytical and logical thinking capabilities are often far below that of a CS major.

A complex problem in radiology is not: "this patient has TB because [chest X-ray]". It's more like: "take this patient that had 30+ surgeries in the past 5y and tell me if his problem is more likely to be diagnosis 1 or diagnosis 2". And the problem for those cases is that you will never reach the critical mass of training data required (within the capabilities of a single startup) because those cases are too rare. You'll need the collaboration of the health system as a whole for that, _i.e._ state policies. So you can't really chip away at the complex problems on your own.

The reason a doctor is educated in general anatomy is so that when their specialty hits edge cases, they are able to reason about what could be the underlaying cause. This reasoning in a sparse information environment is exactly the scenario AI/ML is worst at.

If two radiologists can consistently reach wildly different conclusions then it really can only mean two things. Either or both radiologist are incompetent. There is not enough information available in the diagnoses tools we have to make a diagnoses.

In both cases I would say this counts against radiologist and for an AI solution. I would rather have a continuously improving AI then a coin tossing radiologist.

Reply on news.ycombinator.com