Algorithm can predict pancreatic cancer from CT scans well before diagnosis AI algorithms can predict whether a patient will develop pancreatic cancer years before an official diagnosis, or so this research suggests. Tens of thousands of people in the US are diagnosed with pancreatic ductal adenocarcinoma – the most common type of pancreatic cancer – every year. Less than 10 percent of patients live …
"the team doesn't really know what it is analyzing when it makes its predictions."
This is a big problem and common to so much of this ML stuff. It would probably be more useful in the long run to understand exactly what the significant features are in biological terms. That way there might be scope for preventative measure.
It's not helped by the fact that the fact that the sample is small.
So finding a diagnoses a lot earlier is a big problem? I dont think so. It would be really good to find out what is triggering the diagnosis but not knowing is not really a problem though I am really looking forward to them finding out the answer. It may be however, that ML is actually a lot smarter than humans in certain situations.
So you're happy to have a total pancreatectomy because the ML says that it's right 86% of the time?
Interesting patient: I don't think I'd want to treat you and face the inevitable litigation when the Pathology report comes back 'No cancer found.'
I agree that it would be good for these doctors to work with programmers to enhance the ML system to show the "why" of it's conclusion. But on the other hand, pancreatic cancer is devastating. Although the article says 9 out of 10 people are dead within five years, it's much worse than that. Pancreatic cancer develops to Stage 4 with NO symptoms. When you finally do have symptoms, you are 30 days or less away from death. The doctor can only tell you to wrap up your affairs and say goodbye to your loved ones.
The thing is, IF some random accident causes your pancreatic cancer to be found early, it's actually a well treatable cancer. It doesn't have to be a life sentence - if it's found early. But it almost never is.
I'm glad these doctors are working on this problem, and hope they can further their work well. Sure, I'd like ML systems to be developed to where they can point out the differences between noise and information that caused the conclusion; but even without that, I welcome anything that produces early detection of pancreatic cancer.
This. My father was taken by pancreatic cancer and, as you say, once discovered there is pretty much no hope. 1 to 6 months later, you are gone, and doctors can only prolong the inevitable for only a short while.
This is huge. I have been awaiting a hope for pancreatic cancer considering how much attention is put towards breast cancer and prostate cancer, both of which are 'highly survivable' compared to the viciousness of pancreatic. I hope that this is a real breakthrough.
The algorithm is not meant to research pancreatic cancer causes, just to detect possible cancers in a CT scan.
Nobody has yet found a cause for most cancers. We believe that cell damage is involved but have no more than theories as to what parts of the cell need to be affected or how the damage is incurred.
The idea that this algorithm, if put into medical practice, would result in unnecessary surgeries is absurd. Instead it would allow doctors to target biopsies to suspected tumor areas. It is designed to identify asymptomatic patients before surgery or, worse, palliative care become the only options.
As far as preventive measures, they are the same as all diseases that we don't fully understand: eat healthy, exercise, and don't smoke.
I'm reminded of reading about a similar study, about lung cancer I think. The ML model matched the doctors' predictions with a stunningly high percentage rate - until given scans that the doctors hadn't examined yet, then it was below 50%.
Turns out it wasn't identifying the actual issue by looking at the scan, but was picking up on the marks the doctors put on the scan, metadata about the patient printed in the margins, etc. It couldn't actually identify anything itself.
While I'd love for ML to be advanced enough to pull this kind of amazing prediction, I strongly suspect we'll find out it's not anywhere near as accurate as currently, er, predicted.
"Although the AI algorithm seems promising, the team doesn't really know what it is analyzing when it makes its predictions."
"The researchers reckon the model..."
Hmmm. Back in the day we had something called "debug statements" which could be used to tell the programmer exactly what his program was doing at any given point. I guess that advanced technology has been lost in the mists of time.
OTOH, at least they aren't wanting to use this software, that they wrote but apparently don't understand, on anything important. Oh, wait...
One great thing about this study is that it should really help with finding the cause of the cancer, because there's clearly signs of it before it's obvious to humans.
If those markers can be identified, and there's no reason why they can't be, then there should be better cues as to the root cause(s)
So?
In medicine, many decisions are probabilistic. I chop out lots of funny looking lesions from peoples' skin. 3/4 turn out to be cancer, the rest not.
Small scar for a 75% major life benefit. Considered a good cost/benefit ratio by everyone.
In this AI case, a much bigger scar, with post operative pain and other issues, for an 86% good result. 86% may be cured of a bloody horrible disease and potential shitty death, the rest suffer from an unneeded major operation and its sequelae, but can enjoy their lives.
Is it a good risk/benefit ratio? Does chopping out decent chunks of pancreas, with all the long term potential downsides, win against premature horrible death?
In the article, the folk running this study see this as a potential great winner, but see the need for more study.
To me, they have done a stunning job, with a proper amount of caution.
I know some maths, so I don't gamble. In a couple years if there is more confirmatory data, I would risk having an operation based on that AI guess.
The dataset was split for training and testing purposes; 66 CT scans from 44 patients
I am not a statistician, but that sounds like a small sample.
The thing that worries the hell out of me is that cost-cutting politicians/administrators/whoever start treating the machine as an infallible source of diagnoses. Healthy people are people who actually need treatment will suffer. And unlike a human, you can't go to a big box of ML and say "what were you thinking?" because it wasn't thinking.
I agree; there is a serious risk of p-hacking if you refine your training versus a small control data-set. But this work should be the motivation to start a bigger screening study, which would resolve the issue. So it sounds like a very promising first step and all we have to do is wait another 5 years until the new data has arrived :).
It's a naive Bayes classifier, not a deep learning model or something which can be hard to interpret. If they don't understand how to work out what features it is using, they shouldn't be using it or any other ML tool. Anyway, 4000 features on such a small dataset with no real external validation? Overfitting.
And not in a good way.
Others have highlighted several significant problems, and this is another one:
"The statistical Student’s t
-test was performed on the extracted radiomic features to identify those that are significantly different between the healthy and pre-diagnostic groups. About 4.5% of the total number of extracted radiomic features showed significance at a p-value of 0.05."
Which is no better than you'd expect at random and is the subject of an XKCD cartoon. https://xkcd.com/882/
This is actually an astonishingly bad example of ML and is likely just a random result. Why are El Reg publicising this crap?
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
