HN Reader

> “People should stop training radiologists now,” Geoffrey Hinton said, adding that it was “just completely obvious” that within five years A.I. would outperform humans in that field.

I'm starting to notice a strong trend where non-domain experts will confidently assert that X field will be replaced by AI.

So are datasets/currently available data the limitation here?

Let's say a major healthcare leak occurred, involving millions of images and associated doctor notes, diagnostics, etc... would this help advance the field or is it some algorithmic issue?

I just want to point out that the term "A.I." gets used pretty loosely in these articles, as if A.I. is a monolithic commodity that you plugin to your software to make it do chatGPT.

The example in the article is an in house developed "A.I." to help radiologists assess images. Digging a bit deeper it seems they are using mostly old CNN type architectures with a few million parameters.[1]

I think it still remains to be seen what a 1T+ parameter transformer trained specifically for radiology will do. I think anyone would be confident that a locally run CNN will not hold a candle to it.

[1]https://mayo-radiology-informatics-lab.github.io/MIDeL/index...

This reminds me of the idea that Human-Chess partnerships would be the ultimate manifestation of Chess genius. I'm not sure whether the idea is still holding on but engines are so far ahead of human play that I doubt a human in the loop can add anything these days given how devastatingly far ahead the engines are and the advent of machine learning techniques.

For a more technical dive into startups that have been chugging along with AI in pathology for ~decade, check out this:

https://www.owlposting.com/p/what-happened-to-pathology-ai-c...

> One pathology AI founder told me that it wasn’t hospitals or diagnostic labs that showed the most promise. It was R&D groups within Big Pharma. Those scientists and executives wanted new tooling. They were often sitting on massive internal datasets, had real budget allocated to experimental tech, and — critically — had a clear ROI if your model helped shave months off a study or more precisely target the right patient cohort. Most importantly, pharma didn’t care as much about the regulatory headaches, as they weren’t using your model to diagnose patients

So will it cost me less than $1.5k the next time I need an x-ray?

Glad to hear that this is one area that AI is conclusively useful.

Still not clear that the already superhuman capabilities of AI won't still fully supplement radiologist interpretive skills with every additional bit of training data that comes in.

The thing I find most interesting about ML in radiology is that a computer can observe the entire dynamic range of the sensor at once. A human will only look at a window or a compressed window.

AI is not easily going to "steal" any properly gatekept jobs that require a licensed practitioner.

Machine learning technology can be used to search for artifacts in medical diagnostic images that humans might miss. The flagged situations still have to be interpreted by a radiologist.

“Radiologists do far more than study images. They advise other doctors and surgeons, talk to patients, write reports and analyze medical records. After identifying a suspect cluster of tissue in an organ, they interpret what it might mean for an individual patient with a particular medical history, tapping years of experience.”

Now think about how much of software development is typing out the code vs talking to people, getting a clear definition of the problem, debugging, etc. (I would love an LLM that could debug problems in production — but all they can do is tell me stuff I already know). Then layer on that there are far more ideas for what should be built than you have time to actually build in every organization I’ve ever worked in.

I’m not worried about my job. I’m more worried my coworkers won’t realize what a great tool this is and my company will be left in the dust.

Reminder that according to https://openai.com/index/healthbench/, not only do current AI models provide far better recommendations than doctors, but even if doctors know the AI recommendations as a starting point, human revision does not bring any improvement.

Honestly this doesn't surprise me, considering the quality of the average doctor.

For an ML model, a sofa with a tiger pattern might be a tiger, if in its training dataset tiger stripes always means tiger.

It does not have common sense.

Working radiologist here, 20 years experience.

This article is surprisingly accurate. I fully expect to finish my career without being 'replaced' by AI.

Happy to debate/answer questions :-)

If there is one thing that NN can do amazingly well it is pattern recognition. The biggest blockers to adoption are the EHR providers, regulations and just inertia. It will happen, the only question is when and to a small degree, how.

"The staff has grown 55 percent since Dr. Hinton’s forecast of doom, to more than 400 radiologists."

Wonder what other forecasts of doom he is wrong about :|.

People get tripped up by thinking "there is a subset of what I do that only humans can do and so that means AI will not eliminate my profession entirely and my job is safe".

Let's assume for now that it's true that AI can't do a certain subset of your work. Your profession won't be eliminated from the earth, that's true. But if 80% of your work can be done by AI, 80% of your work will be done by AI. There will still be humans kept around for that remaining 20%, but fewer of them will be needed.

I formerly worked at https://corista.com, dealing specifically with integrations with AI pathology algorithm vendors like https://www.qritive.com and https://visiopharm.com. (Note: we dealt specifically with microscopy rather than radiology.)

I really think we were doing things the “right way” before I left: providing AI analyses from various vendors as overlays for slides, and being able to pre-flag slides with obvious cases of cancer or other infection. (These analyses typically being custom algorithms provided by vendors through APIs we collaborate on, not LLM output.)

With things like this simply built into an existing, established pathology platform, I’m pretty confident a team of 4-5 pathologists could do the work of 6-8, with better quality output, similar to how Copilot and other tooling speeds us up as developers. At $300-400k/yr/doctor, that’s considerable savings (and the opportunity to allocate more doctors in specialties that aren’t easily automated).

However, for lots of reasons, it seems the market doesn’t necessarily agree with me on the value potential of this approach in this field (which can certainly be a self-fulfilling prophecy).

The key to the power of the LLM is that the training process can learn effectively from vast corpora of unlabelled text. Unfortunately, there is no comparably vast database of medical images.

In order to "crack" radiology, the AI companies would need to launch an enormous data collection program involving thousands of hospitals across the world. Every time you got an MRI or X-Ray, you would sign some disclosure form that allowed your images to be anonymously submitted to the central data repository. This kind of project is very easy to describe, but very difficult to execute.

TL;DR: “Five years from now, it will be malpractice not to use A.I.,” he said. “But it will be humans and A.I. working together.”