The issue with that is that misclassifying something in an image can have serious consequences. So unless ML is 100% accurate, you'll always want an alternative way to check if it's really an artifact or something serious, so, why bother making an accurate model when you'll still double check the results?

I don't follow. Why would anything need to be 100% accurate? Humans aren't 100% accurate, and there's no reason ML would need to be 100% accurate either.

The is a gigantic difference between a doctor looking at something in a scan and deciding it's an image artifact when it's not, subsequently missing a diagnosis, and an ML algorithm looking at that same image, deciding something is an artifact and removing it from the image. At that point you're not looking at an actual scan, you're looking at something different.

It would be interesting then to give the human a side by side comparison with the artifacts removed (and maybe shown in color), so that the doctor can more easily see whether he agrees with that.

The difference is that a doctor can try to explain his findings and collaborate with others. This includes not just looking at the scan data but also reading other test results and patient history. ML is not capable of this as far as I know.

>This includes not just looking at the scan data but also reading other test results and patient history. ML is not capable of this as far as I know.

ML is capable of doing that if you train it to do that.

Correct, the target should be instead better than Xth percentile humans across the majority of cases, with the failure modes categorized.

I don't get this perfection requirement that gets put on so many systems here. No system is perfect, the question is what failure rate you accept.

You're looking at this in isolation instead of in the context of a diagnostic process. Thinking of medical ML models in terms of benchmarks does not make much sense, you're interested in whether they actually change the process.

So yeah, maybe you get a model with 5% false negatives where humans have 10% false negatives. That's good. However, when you go to apply that model in reality, what happens to those 5% where model and doctor disagrees? First, we should think about the kind of disagreement. In most of those cases I'd bet that the doctors see something wrong but can't say what, not that the model says "this is bad" and the doctor "it's not".

Second, we need to think about what happens when doctor and model disagree. As the model is not 100% accurate, the doctor doesn't know if the model is mistaken or if they're the ones mistaken, so they'll probably order more tests anyways. If it can be something serious, it's worth it to do an extra test to make sure. They'd probably ordered those tests anyways if they weren't sure of what was happening, model or not.

So what did the model for a single test change? Did it really change the diagnostic outcomes? What's the actual benefit of the model? How much it's worth to get from 10% to 5% false negatives with a model for a single test if just adding more tests (say, with the same 10% false negatives) to the mix can give you a 1%, 0.1% false negative rate?

That's my point. Unless accuracy is really high, ML models are not going to remove uncertainty in diagnostics, few diagnostics consist of just a single test. Benchmarking models against human performance in a single test is not a metric that can drive implementation in the real world.

I agree, that it would not need to be 100% accurate. However, automation with low accuracy would induce other kinds of errors (e.g. complacency or loss of skill in doctors). Therefore, automation shouldbe significantly better than humans to justify their use.

We were talking not about about clear signs but about small signs that could be artifacts or actual issues. In that case you are going to have to do some complimentary or repeated tests to verify, so the ML output saying something is an artifact will have no effect on those procedures unless it's 100% accurate.

But in general, these are matters of life and death, literally. You're still going to have someone looking at the images and verifying the output. That limits a lot the potential cost benefits of these applications.

I still don’t understand why decisions about life or death have to involve 100% certainty or 100% correctness. I would think that it’s the opposite, that when the stakes are high, you want anything that improves your odds. If your life were on the line, you’d accept treatments or diagnoses with much less than 100% certainty.

See my comment below for the same discussion: https://news.ycombinator.com/item?id=27433368

My point isn't that you need 100% accuracy. It's that a diagnosis is a process, not a single test. If your model applies to a single step of the process, and if it doesn't remove enough uncertainty about that step, you're still going to continue the diagnostic process and the model is not going to change anything really.

You said ML will have “no effect on those procedures unless it's 100% accurate.” I disagree with that, and think it’s a bit absurd—the 100% figure is not possible anyways, so you might as well stick some other impossible predicate in there and the statement is equally insightful.

E.g., “it will have no effect on outcomes until pigs fly.”

There will always be uncertainty, and uncertainty isn’t the only relevant parameter.

> There will always be uncertainty, and uncertainty isn’t the only relevant parameter.

But it's a really important one and a lot of medical ML research doesn't seem to address in the proper sense. A very simple example: a ML model that classifies lung nodules as benign or malign, with 95% accuracy vs 70% accuracy of regular radiologists. Very good, right? But for actual, real world results, you need to see how the patient outcomes change. If the patients where the model and doctors disagree were going to have extra tests or followup regardless of what the model says, the model is not actually offering anything new despite the increase in accuracy.

So no, it's not absurd to say that models that work for a single type of test need to be very, very accurate to actually bring changes to the procedures that are worth the investment. What matters is whether they actually change patient outcomes, and sometimes it seems like the ML researchers barely consider it.

Where I live it's not uncommon for doctors to consult the images with their colleagues whenever they come across something unusual, even if they are "pretty sure" of their diagnosis, because the consequences of an error can be severe - we're talking literally about life and death here. So the ML system would have to be as precise as 2-3 humans together in order to be considered "good enough".