That's one of the biggest roadblocks: hospital & healthcare entities will not not have permissible mass sharing of data. When switching hospitals, its really hard to get a full copy of the record for single person. Try that for whole population. I may be wrong, but hospitals also could have incentives to tightly hold on to their proprietary data - where they spent a large sum of money. I worked in ML+Dermatology & getting data was a lot of regulatory check.

Also, DICOM is the format for maximum interoperability from different radiology modalities & manufacturers. But annotation & markup methods vastly vary between institutions. There is no commonality or agreed upon standards on information interchange.

Federated learning (with implementations such as FEDn [1]) is supposed to solve this problem; Only sharing training weights, but not sharing data.

Sure requires some coordination, but the legal parts should at least be solvable in this way.

[1] https://github.com/scaleoutsystems/fedn

I've been involved in a federated learning radiology project, the biggest issue is image technique and labelling.

Different centers practice very differently, with different imaging protocols, disease prevalences, and labelling/reporting.

This project was looking at renal masses and the only part that worked well with federated learning was image segmentation and probability that the mass is a cancer, this is a competency I expect out of a first or second year trainee.

It was horrible for predicting subtype of cancer as we couldn't get a good training set (few of these lesions are biopsied, specific MRI sequences that may help are not done the same way in every center) which is what the goal was and more of an experienced generalist/subspecialty radiologist skill.

Practically subtype doesn't make too much of a difference for the patient as if it's "probably a cancer" it'll just get cut out anyway, but highlights a challenge with federated learning.

BERT-like encoder models actually do help here, although unfortunately the dense representation is not as nice and clean as a class, it will at least map to something similar. If the hospital provides even a completely bizarre entity id, the requirement would be to provide the dictionary of descriptions and train the BERT-like model to associate the entity ids and descriptions.

> BERT-like encoder models actually do help here, although unfortunately the dense representation is not as nice and clean as a class, it will at least map to something similar.

Completely agree with this.

> If the hospital provides even a completely bizarre entity id

This is where it beings to fall apart, the hospital doesn't always provide an entity. In this specific project, some will argue "it's a big enough renal mass just cut it out, based on current evidence it's safest to remove" which while true at the moment doesn't help us in developing a model to prospectively predict the subtypes that do/do not need surgical excision.

Consequently, there is no biopsy and specific MRI sequences that are proving helpful at subtyping are missing, irrelevant for clinical care per current standard practice but hugely relevant if you're trying to change that practice as we need labelled data but we don't have the resources to have a radiologist/pathologist go over every case again and try to fill in some blanks en masse.

Im not sure it even has to be complicated and full on machine learning.

Could also just make it easy to share diagnosis to the a database stripped from personal information about the client, properly index it, share it worldwide, allow doctors to contact eachother automatically who match.

If it is a rare condition, reidentifying people from even small amounts of relevant data (age, gender, hospital visited) data is trivial. If it is a common condition, usually there is plenty of data already there.

Is that really possible? I don't understand the field exactly but at least if you use deep learning, you'll be able to recover inputs from gradients, and if you want to program it such that no other party is able to recover your inputs, that's called cryptography and homomorphic encryption is FAR from practical yet (like a million times slower than practical). Without rigid mathematical foundations, I would doubt that it's just a fancy way of gathering all the training data and doing the training and making it perceive as if everyone retained their own data.

Yeah you have to make it non-personal, as well as deal with the commercial side of it. I've seen a few initiatives, but they all seem to lack the pockets, drive and ambition of an arrogant Silicon startup. Even though it's annoying at times, it might be what can do the trick in this case.

> That's one of the biggest roadblocks: hospital & healthcare entities will not not have permissible mass sharing of data.

Maybe in some other country?

I am not certain how it works in all of Europe, but it doesn't seem possible in Sweden or UK.

In US and Japan, they will rather do the full roster of tests than transfer in radiology results from another hospital. Its not something to do with costs, but physician opinions. Every specialist has his personalized view of the patient in hand - so the corpus of information he needs is different.

As far as I know, in Sweden data about a patient is shared between every hospital... you need to sign up on a web portal, of course, but from then on, you can authorize a doctor or nurse anywhere to get your data with a click... you can go to any clinic and they have access to your full history. It's impossible to do anything properly if you don't have that. I just don't understand why in the US, it's acceptable for hospitals to be silo-ed like that to the detriment of everyone's health.

Here's a news report from 2017, when the system was being set up: https://www.bmj.com/content/357/bmj.j2069

Today, this has been working for years.

The US has multiple individual medical systems larger than Sweden with essentially the same abilities. Sweden has roughly the same number of hospitals as South Carolina.

What makes you think the Swedish system wouldn't scale to countries with 100's of millions of people??

Serious question - could it? How would that be accomplished?

What's nice is considering a global audience.

this may be a sufficiently powerful macro force to disrupt that

I think it will happen much faster in China, India or maybe Japan than in the U.S for that exact reason. It's quite possible the U.S will lag a lot in healthcare soon.

Exactly: what if my radiologist doesn’t know my particular rare condition?

I would feel comfortable removing the human diagnostician. Let’s have actual human doctors acting as researchers working to improve the AI diagnostics.

My perspective as both a radiologist and CS/AI researcher so exactly what you're proposing:

1. We don't practice in imaginary vacuums, it's easy* to identify that something looks abnormal and then refer to clinical resources/other physicians for rare diseases with specific questions in mind (i.e. recognizing the nuanced imaging finding and referring to a resource to assist).

As a rare disease example, tumors of the eyeball/orbit are very rare, but detecting them is not. If I open the case and see one I can refer to StatDx to help me narrow my differential knowing what imaging features I'm looking for. This reduces my misdiagnosis rate (which as an aside is ~2-5% for radiologists, "major" clinically significant that impact morbidity or mortality ~2-10% of those depending on the study).

2. Rare disease are hard to diagnose, and would likely also be hard for an AI. Imaging appearances are not unique to the vast majority of diseases, especially what we cal the "weird and wonderful".

Pelvic tuberculosis, endometriosis, advanced cervical cancer and advanced rectal cancer can look identical/nearly identical on MRI and the clinical portion as well as additional testing helps us get to the diagnosis.

We don't have to diagnose everything based on a single imaging test, nor should we given:

3. Diagnosis is a tradeoff of sensitivity and specificity. You can't have both.

Let's consider adrenal gland tumors. Statistically these are going to be benign, there is no specific imaging feature to tell a small adrenocortical carcinoma (ACC) ~1 in 1 million incidence from an adrenal adenoma (99% of adrenal lesions).

We also can't tell them apart with a biopsy under a microscope.

If you're unlucky enough to get an ACC you're basically shit out of luck as the only options we have are to recommend adrenal surgery (and their complications which can be death) to optimize sensitivity, or assume that it's benign and optimize for specificity considering disease prevalence and risk of overdiagnosis.

In practice, we just use a cutoff of 4cm. I'm not sure how an AI would solve this, especially as there isn't a large enough training set. MD Anderson has the most experience of any center and they've had ~600 cases in 40+ years which as you can imagine encompasses a very heterogeneous imaging set (we didn't have multidetector CT or 3T abdominal MRI 20+ years ago).

Overall, AI can and should help radiologists and as someone involved in this field I can't envision a world where we can safely remove the human diagnostician element from the mix, given that it's a spectrum of grey not black/white labelling as it is for object detection.

We've had attempts with mammography and stroke AI and it's still horrendously inaccurate compared to what I expect out of a resident radiologist let alone an experience staff physician.

A great post. I can attest the same about information extraction from semi-structured documents. The situation is far from full autonomy. Can't do anything without human in the loop, not even with the latest AI.

I am seeing this trend - everyone can explain at length why AI is "not quite up there" in their own field, but believes it's "near AGI" in other fields. We find it hard to imagine future difficulties AI will have to face in general, we can only do that in our own field where we have learned from direct experience.

* * *

NER, Entity linking, and relationship extraction definitely seem to be 'low hanging fruit' due to LLM improvements, but one of the big problems is that they really need a completely different architecture to limit the decoder vocab if using a decoder transformer for producing the set of sequences in relation extraction with specific entity ids. A Longformer with full attention to input sequence, and sliding window attention to a large dictionary could be a decent way to find tune a system like this, but there are few that try it. Unfortunately there's a lot of stupidity going around right now in thinking the answer is just to 'pRoOoMpT tHe LLm RiGhT', but that will always be exceedingly wasteful such that processing terabytes of files will be prohibitively expensive, and there's no guarantee the system will always restrict to the specific vocab and structure desired.

The images in radiology definitely make these types of things harder, and the sparsity is an enormous issue. However, working with some projects in this area, I don't think it's as impenetrable as a lot of radiologists in AI suggest. The main thing needed in the field is adoption of better techniques and architectures to deal with these problems.

I agree it's not impenetrable, that's why I'm working on this problem. What I disagree with is the "this is trivial" statements.

> A Longformer with full attention to input sequence, and sliding window attention to a large dictionary could be a decent way to find tune a system like this, but there are few that try it.

> Unfortunately there's a lot of stupidity going around right now in thinking the answer is just to 'pRoOoMpT tHe LLm RiGhT'

I agree with that this is not the right approach despite all the media hype (and that fine-tuning is), my research has been (more or less) attempting what you've proposed.

The problem right now is that there is no good dictionary (ontology). RadLex and UMLS have inadequate coverage. Working on a better one as we speak. Context length isn't an issue but Longformer/BigBird are good ideas, it's what I've tried as well.

I previously worked for a Radiology PACS and it's hard to get funding/interest to even tackle the problem. With how lucrative it could be, I would think that a corporation would be very interested in putting resources into it, but this has not been my experience.

No PACS that I know of even wants to tackle digital pathology in a significant way, which last I heard had about 5% adoption versus glass slides.

There are several corporations, some of which have raised > 100M, tackling this problem. AIdoc being one of the more notable ones.

> In practice, we just use a cutoff of 4cm.

In other words, we extract a single feature, and apply a single nonlinear activation function to that feature to decide whether or not to activate the 'treat' signal. We've replaced all that vaunted human judgement and mental modeling of the body with a heuristic that has equivalent power to a single neuron neural network.

I appreciate that there's a dearth of training data, and it varies in quality. But this is precisely the kind of thing where sufficiently powerful ML could do better than the simple heuristics we can come up with.

The heterogeneity of imaging types doesn't have to be a problem. Train the model on all the data, and all the different kinds of scan, all the anatomical knowledge.

Look at how LLMs are able to do stuff like write code that has comments written in pirate speak. Do you think they learned how to do that by studying a large body of code with pirate-speak comments in? No. They picked up examples of pirate speak in one context, and code in another context, and they're able to combine them together in ways that make sense.

ML models looking at diagnoses with small training sets that are largely in obsolete scan formats could still, in theory, learn how to spot those diagnoses in more modern scan images, because they have learnt from other, much larger datasets how things in the newer format correlate with features in the old form.

You're missing the point, it takes me < 5 seconds to clear the adrenals. The example is intended to illustrate that there is no feature to extract that would make a model BETTER than a human for the things that we care about (rare and challenging diagnoses).

No one is arguing that ML can't segment and measure a structure, this is the lowest hanging fruit. ML can't diagnose an adrenocortical carcinoma (an example of a rare disease) because medicine doesn't know how to.

> In other words, we extract a single feature, and apply a single nonlinear activation function to that feature to decide whether or not to activate the 'treat' signal.

Now do this for the > 1000 other possible diagnoses on a CT abdomen, and have it be as fast as a human with equal or better ROC curve in under 5 seconds and cheaper than the $70 a radiologist bills for this. Unless you can eliminate having someone like me read this scan a ML model to measure the adrenal glands is worth $0.

I'm aware of the literature in this space. Your hypothetical is not novel and has been attempted. Everything from logistic regression to GANs and ViLMs has been tried and SOTA can barely beat human performance on an x-ray in idealized situations.

A genuine offer, my group has funding to support a FTE on this project so if you (or any reader) think you can solve this on a test set please reach out, my e-mail is in my profile.

> Look at how LLMs are able to do stuff like write code that has comments written in pirate speak.

You mean the things that hallucinate ~40% of the time? LLMs aren't even ready for a production CRUD app let alone human life.

> ML models looking at diagnoses with small training sets that are largely in obsolete scan formats could stil

It's not obsolete. It's a completely different image type. This is akin to saying a ML model trained on black and white sketches can paint the Mona Lisa in 3D.

> all the anatomical knowledge.

A misunderstanding of the problem. The anatomy is easy. The pathology is updated every 1-5 years so there is no historical dataset.

You need to pair a system like this with regulation and changes in the medical field to upskill nurses so they have a larger extended skill set and essentially have a path to becoming doctors via practical experience.

There's a lot of information on medical conditions. I remember hearing about IPhone apps that can get a 90% hit rate on correct diagnosis for general practitioner style visits. Taking that into account, and a large scale medical database, you'd think we'd be pushing that technology down so we can get better productivity out of primary care, so most of the time Dr's were using their big brains for big complex problems.

90% hit rate is pretty easy to accomplish for a standard doctor's visit, a big percentage of the doctor's visit are standard. You can probably train most humans to deal with that quickly. They often use students if the pressure is high here. It's about the other 10% + unforeseen issues.

Speaking to radiology, ~90% of imaging studies are normal so I can sign my default template without looking at the images and be correct the vast majority of the time.

Also a rad (and heavily in the AI field). I think you mean 90% have inconsequential findings and not that they are "normal". Can you really apply untouched templates to 90% of your cases? I find that very hard to believe. How many templates do you have and how can they possibly describe the variety of possible findings seen on each type of study?

90% hit rate could probably be achieved with "take some panadol and come back in a couple of days if you still feel bad."

> have a path to becoming doctors via practical experience.

Lol no

> I remember hearing about IPhone apps that can get a 90% hit rate on correct diagnosis for general practitioner style visits.

Was that in marketing materials?

> so most of the time Dr's were using their big brains for big complex problems.

I have been misdiagnosed multiple times to my significant detriment. The issue each time was doctors failing to pay attention and exercise their faculties and knowledge, because they didn't care. These people didn't seem to have "big brains" and my subsequent experience taught me they weren't interested in responsibility or accountability either. I've heard similar stories from (for example) women who struggle with conditions like PMDD.

I'm not on board with giving these people tools that they can use to justify paying even less attention.

I honestly don't think GP's need big brains, just slightly above average and a good memory.

Skills required for being a good GP are probably empathy, communication, decent memory, reasonable analytical skills, not to big of an ego. Very different from a researcher.

You could make the opposite argument that doctors don't pay attention so we should replace them with AI.