Radiologists are not even trying. They treat their methods like FDA-approved medical devices. Even basic image segmentation to help with 3D structure recognition is off-limits. The benefits of neural networks in diagnostic radiology will not occur in one shot, and I don’t think it will happen at all in the United States until people start sending their data elsewhere. But good luck getting it. I just got my CT scan from a total mis-diagnosis that resulted in an unnecessary surgery. It came on a CD that won’t read in the only drive I have access to. And even that is just images. It’s not possible to get the actual data. This is not a failure of DNN. This is active AMA hostility toward technology, and not just in radiology. Just watch, people are going to start going elsewhere for medical care. They will do everything they can to get insurance requirements, subsidies, and laws against it, but they will lose. They are a dishonest Luddite cartel, and they’re hurting people.
Radiologists most definitely are trying. Our institute's entire medical imaging research arm is driven by several very motivated practicing radiologists. You just misunderstand what it is that they do, fundamentally. Diddling with some pics and publishing papers is just not in the same league as making medical diagnoses. A lot is riding on their understanding every little artifact of the algorithm/approach that gives them a modified image to interpret. They will never accept black-box automagic, and they will always evaluate the benefits of novel algorithms together with the drawbacks of having to get used to their quirks and opaque artifacts, possibly with outcomes impacted and/or lives lost in the process. Where the risk/benefit analysis is clear, they do adopt plenty of common-sense automation tools for a very simple reason - they get paid per scan read, so their time is (lots of) money, to them.
I don't think the blame falls on practicing radiologists, but the OP is absolutely correct that medical data is way too inaccessible. It is often impossible to get your own raw data, and even worse it is sometimes impossible to share that data with another doctor. Two large hospitals in major US cities apparently can't share EEG data because they use different software to read it. Guess who wins when all your prior data gets essentially thrown out? It's not the insurance companies, and it's certainly not you - it's the new hospital.
How realistic it is to have ML involved in reading radiology results in theory I don't know, but the larger point is that in practice it is sure as hell not going to happen until patients have real access to do what they please with their own data. Not only am I pissed I can't have my own EEG data, but I also would gladly contribute it to a database for development of new tools, or any other research study that asked. But there is essentially no way to even do that, at least at either institution I've asked. Just think of all the data that is being utterly wasted right now!
The number of patients that are interested in viewing or accessing their own data has to be negligible. Last time I got an Xray they actually gave me a DVD of the imaging itself. I remember looking at it, I thought it was neat, but ultimately there was little use in there for me. I dont know what % of patients have bothered to look at it.
Viewing their own raw data may be negligible, but sharing between medical professionals is a relatively common and necessary practice. Currently it is extremely difficult to get one doctor to share medical information with another, and it shouldn't be.
Provider organizations are understandably reluctant to accept removable media from unknown sources due to the risk of malware. Many of the computers that doctors use don't have DVD drives or they're disabled for security.
It's not about the patient reviewing their own data as much as it is about the patient having easy access to their data and can easily share that data with other consumers of it (i.e. some AI based interpretation service)
The P in HIPAA stands for portability. It should be a HIPAA violation for them to not give me the original data sets for my healthcare when I request them in person.
I am aware there isn't, what I'm saying is there should be - particularly for dense datatypes like EEG that we probably aren't fully leveraging at the moment.
That's false. Both Epic and Cerner, the largest EMR companies, have databases of millions and millions of patients. They're used for research, and stuff.
Yeah but all the EMR data I've seen have just been doctor notes, some prescription info, standard clinical scales, etc. Not that this can't be interesting, but there's not really a major database for richer data types like imaging or EEG (AFAIK)
I left radiotherapy for this reason: they haven't the faintest idea what they are doing, and are constrained by whatever manufacturers managed to get the FDA to approve, which is 20 year old tech at best. Not that radiotherapists/logists know how to deal with modern tech... A small hint: our software repo was migrated TO svn in 2014.
When people do 'AI' in radiology, they mean making little with scripts in Tensorflow. Sure, it's a beginning, but I've seen entire institutes being unable to get to grips and move past that stage. You wouldn't be able to tell from their slides, of course.
> they haven't the faintest idea what they are doing, and are constrained by whatever manufacturers managed to get the FDA to approve (...)
If you take this layer of proprietary magic nobody understands, and add DNNs to mix, you'll get... two layers of proprietary magic which nobody understands (and possibly owned by two different parties).
I've left the field for greener pastures ;) Looked at your website, interesting, but curious how you deal with all the legacy software. Java was considered quite recent at one of the places I worked, they were rolling Delphi in 2019 and were not planning to switch!
It is an interesting market, for sure. The fact that these vendors are reluctant to upgrade tech stacks actually creates a bunch of opportunities. (Although we wish they'd just upgrade their Windows version:)
Also in the US, I had a similar 'awakening' if you will after being at the side of a loved one for a little over two years of intensive medical intervention. I've been left quite bitter and ultimately distrustful about where things stand today.
That said I do recognize that I have the advantage of not making life and death decisions and have no idea what its like to weigh the advantages of innovation against the risk of untimely death or significant impairment/expense that comes with advancing the frontiers of medicine.
Couple of points: The 21st Century Cures Act has recently expanded rules for information portability, which will make it much easier to get access to your data in the future. The challenge here has nothing to do with radiologists hoarding your data. The lack of interoperability typically stems from limitations of electronic health information systems. Most radiologists would love to be able to look at your scans from multiple previous hospitals where you were imaged previously, but technical barriers currently make that difficult.
I don't blame the practicing radiologist, but I also don't buy this is purely a technical issue. The hospital is quite literally incentivized to have you repeat tests. I highly doubt they care to make data accessibility/portability a priority. Hopefully these new rules will force their hand.
You are right, to an extent. Health systems and EHR vendors both have historically had an economic disincentive to share data. Think “ecosystem lock-in”. My impression is that things are gradually changing for the better.
There is an IHE-Standard Profile that specifies how data has to be to laid out in portable media(USB,CD). All PACS systems follow it and it's basically a no-problem today.
I don't doubt about your anecdote, but I don't think that it's a very common occurrence.
You have the legal right to obtain a copy of your medical data. Providers can require you to pay reasonable administrative fees for making data copies but they do have to give it to you. If they don't comply then you can file a formal complaint.
That's complicated. I work at a very large health care org that employs ~1000 radiologists. We definitely would love to have a good solution, but there just aren't good enough vendor solutions and even working with vendors it's hard to get things to an acceptable state. In that sense I have more hope in derm.
The problem is with the money and power. Doctors will never let go their income for a data scientist.
This type of invention will never start in the US, it will start in a communist country where leaders can move mountains or in Africa where there is a big lack of Doctors.
US based companies have been shipping AI/ML tools for nearly 30 years at this point, which undermines your argument.
The biggest problems are data access (big) and data quality/labelling quality (bigger).
Medical conservatism is a real issue, but nowhere near as big as those. There isn't a big cabal trying to keep AI out, it just hasn't worked very well so far.
I don't buy this, doctors are not one united body. Doctors with an AI tool will be more efficient than doctors without it. If the tool has a measurable positive impact on patients (outside of cost reduction) then it will become necessary to have the AI in order to get the patients.
