Hacker News Reader

Comment by Workaccount2 | Hacker News Reader

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...

View on HN

There are many, many papers and projects out there about tuning foundation models on various types of medical imaging data, and many organized efforts to produce large medical imaging datasets to feed that training. This stuff is well-known in the trenches and can improve on the older, smaller CNNs in some ways, but not in a way that's produced any step change in automated capabilities. People are certainly working on it!

> I think it still remains to be seen what a 1T+ parameter transformer trained specifically for radiology will do

Does image processing of something like this scale with parameters?

It makes sense that language continues to scale as the vector space is huge. Even models that generate images scale as the problem space is so large.

But here there are only so many pixels in the image and they are a lot more uniform. You likely can't have 1T images to train on, so wouldn't the model just overfit and basically memorize every image it has ever seen?

You are right, but I meant it more as a total radiologist, scans the images and can put it into medical context.

Unfortunately, VLMs don’t seem to work. Even the best ones (eg, Gemini 2.5 Pro) hallucinate to a laughable degree. I’ve been seeing them make up large-scale features in images that absolutely no human would ever mistake.

What does work is narrow AI trained to do very specific tasks (such as medical image segmentation). But I don’t think that type of AI will benefit much from massive scaling.

If the meat machine can learn it, then the silicon machine can learn it.

Medical records remain a walled garden mostly because of HIPAA. Unlike all the AI development that has managed to skirt the copyright law to train large models, training a 1T+ parameter transformer on a large enough dataset will need a lot of consumers to give up their medical records.

UK's doing it:

>Concerns raised over AI trained on 57 million NHS medical records

https://www.newscientist.com/article/2479302-concerns-raised...

Underreported competitive advantage of NHS is their unified and comprehensive data

China won't have an issue with it.

The US won't have an issue with it. You haven't realized what the US has become.

A place where it's still possible to innovate, despite the best efforts of the copyright lobby?

I've noticed a funny trend where people will bring up an ethical concern and some dweeb will immediately roll out the Chinese strawman like it justifies their incredibly dubious position.

"We should violate medical privacy laws on an enormous scale because China" Okay buddy.

I wasn't suggesting that we should, but in my lifetime we went from hospital admissions and discharges being published in the local paper, to that information being some of the most closely guarded with (theoretically) severe penalties for disclosure.

I think it was AIDS and people not wanting to be outed that changed things.

But whatever the reason, we now have HIPAA and it adds a huge amount of complexity to the management of medical records, and a huge amount of complexity to any research that needs access to a broad spectrum of medical case histories (such as training AI). Other countries don't have these concerns, and will outcompete the US on developing these capabilities.

Concerns about HIPAA as an obstacle to research are largely a red herring. Almost no clinically useful research can be accomplished by just processing large volumes of historical patient charts. In practice the researchers will have to gather additional data from study subjects so they're going to need to obtain consent anyway.

People keep claiming that other countries will out compete us in medical research and it keeps not happening.

yes but you can train relatively dumb AI on high volumes of historical data and presenting factors. This is typically just a 'diagnostic assist', but often significantly better than a human

Better how? Do you have a citation for that?

I did work on a diagnostic assist tool developed by a large pharma which diagnosed between Asthma and COPD. GPs get this right 52% time, specialists just over 60 and AI came in over 80%...using 12 relatively mundane input variables. I believe there are a lot of these situations, but not clear pathway to FDA approval

Is that worse than being worse off technologically than China and the consequences that bears

If we become worse off than China, it's because we squandered a massive decades long lead and actively moved critical industrial processes over there, not because we didn't let AI companies scrape HIPAA data.

Do you really feel life has been horrible in all these countries which have been worse off technologically than the US for the past decades?

Being first is not the same as being happy.

Because the end result is either paying China to eval scans, buy models from China that do it, or settle for inferior scan reading by humans.

The Red Scare never ended for a whole lot of Americans. And my foot is more communist than modern China is. But you know, being afraid of Communism is just as, if not in fact easier, when you have no clue what Communism actually is.

If copyright stayed at 7+7, I would agree that AI companies were "skirting" copyright law, but as I see it life + 70 violates the "limited" terminology in the Constitution by denying people the right to create derivative works.

IP holders did this to themselves by extending the term beyond any reasonable interpretation of "limited".

Hell, Jack Valenti wanted to extend it to "forever minus a day"

IMO, copyright law in its current form is unconstitutional and unnatural.

Epic probably has enough breadth on their own to do it, although they don’t actually have an image display system that I know of, so perhaps they don’t have access to the raw images - only the written reports.

Right, some of that data is in Epic Cosmos.

https://cosmos.epic.com/

Epic does not have the ability to use the data as they see fit, even if they have the data available. That is why I brought up HIPAA.

When it comes to test results, does HIPAA prevent anonymization, such that all would be provided is test input & eventual result?

From a summary on HHS.gov it says "De-Identified Health Information. There are no restrictions on the use or disclosure of de-identified health information." Maybe someone with more knowledge could expand on the limitations of what counts as "De-Identified" but I think that might work. I followed the reference and nothing in the CFR jumps out at me, but I'm not a lawyer so who knows.

HHS has published clear guidelines on de-identification.

https://www.hhs.gov/hipaa/for-professionals/special-topics/d...

"The radiologist will be with you soon" also gets used pretty loosely in this article, since a radiologist is not someone you are likely to speak with. They sit in their home office in Montana clicking through images all day, and send thru their diagnosis to a different practitioner at the office you are visiting.

Which means the field has a larger issue since they can just as easily sit in the Philippines or Brazil or India.

There's a reason they can't.

Why not? Where I work scans are reported overseas. The radiologists are qualified in the country they are reporting for however.

Ok, I should have been more clear.

A radiologist practicing in the USA can live anywhere in the world, but they must be trained in the USA, licensed in the state(s) where they read from, credentialed at the hospitals or facilities they read from, and carry medical malpractice insurance.

The point being, hospitals can't throw cheap Indian or Filipino radiologists at the worklist.

Can a certified radiologist farm out work to cheap assistants in other countries, and then put their stamp on the result?

Not sure about other countries, but RAs do outsource work to RRAs(1) and then sign the final product. This is relatively new as Medicare/Medicaid didn't recognize RRAs as mid-level practitioners until 2019. I suspect the answer is yes for ARRT-certified RRAs, but it will be state-specific (similar to the answer above about RAs).

(1)https://www.arrt.org/pages/earn-arrt-credentials/credential-...

This is illegal, and you will go to jail if you do this with a Medicare or Medicaid patient and it is discovered.

Why? If they agree with it, is it that different to a software generate report that they sign?

So what you're saying is they can, they just don't at the moment for temporary, localized legal reasons.

For a while it seemed that "AI" topics were treated as niche and news articles used relatively specific terms of art like deep learning or machine learning. Somehow ChatGPT changed that and now everything is just AI.

ChatGPT changed that because it, along with other current advances, brought AI out of its niche and into widespread mainstream usage. The laboratory the source article covers (https://www.mayo.edu/research/labs/radiology-informatics/ove...) publishes some of the technical detail you're interested in, but in 2025 it's also interesting to people less familiar with theoretical computer science. The doctors quoted in the source article, using programs that can calculate kidney volume or detect blood clots, don't necessarily want or need to know the difference between "artificial intelligence" and "deep learning".

I do my best to filter ai topics here using ublock, and it's blocked about a third of the posts, and yet every day there's a new variation, whether a new technique/technology, a new product (open source or commercial), etc... that I have to add to the filter. It's really offputting. Yet apparently this is the only site besides lobste.rs that has programming related content that is of any depth that I know of. I really wish hacker news had the ability to tag posts and let me filter (and tag them generally not have to get specific on each new buzzword or product).

Why do you want to read about tech, but specially non-AI tech?

I'll speak for myself, for me it's illustrated by the spam sketch from Monty Python. It crowds out everything, including my mental space. The hype and disconnect from reality make it intriguing to my brain so I am very likely to end up reading the comments about it, but it's usually a waste of my time. Software does what it does and my opinions (or anyone else's) just don't matter. I should focus my time on learning and building instead of trying to understand hype.

Agreed

Reply on news.ycombinator.com