Pages Navigation Menu

SHOWFUN - Show & Fun & More!

AI beats doctors at visual diagnosis, observes many times more lung cancer signals

A new study from Stanford University could ruffle feathers in the medical community, as the researchers report that their newly developed machine learning algorithm can identify tissue slides exhibiting a specific type of cancer with far greater accuracy than human epidemiologists. It’s one of the first indications that computers aren’t just capable of addressing the “subjective” portions of medicine, but that they’re actually better suited to such problems than human doctors, in some cases.

In the biochemistry lab, most scientists are constantly doing favors for other scientists. If they didn’t, the whole place wouldn’t work — Josephine’s got a bunch of slides and needs to identify those showing this disease, Kevin’s got a bunch of petri dishes and wants to see which contain that type of colony. Neither can look for their own result because they have an interest in a certain outcome, so their yes-no assignments for different slides (or dishes or whatever) would be biased. It might sound absurd, but highly trained research scientists regularly collect slides just riddled with highly visible cancers, and have to walk them down the hall to ask a panel of colleagues: “Hey, see any cancer?”

movidius-machine-intelligenceBut in a busy hospital, there isn’t always time to find such a perfect human to review your results. In any case, it turns out to be quite difficult to interpret the results of many common tests, regardless of the level of interest in the outcome. Stanford professor Michael Snyder pointed out that, “Pathology as it is practiced now is very subjective… Two highly skilled pathologists assessing the same slide will agree only about 60 percent of the time.”

For instance, take a look at these stained breast cancer slides. Even with training, the distinction between trouble spots (arrows) and regular growths is slight — and things get even worse when scoring along a severity scale. Would you like your health resting on whether you score a five or a six on some scale, based on the visual evidence below?

cancer slide

The newly developed system was trained on over 2,000 slides and came to identify over 10,000 individual traits that collectively contribute to a correct diagnosis. This is compared with the human best, which incorporates only a few hundred signifiers. Better yet, the algorithm does not have any scientific or professional hubris, and will score each slide according to its individual merits alone.

cancerIt’s worth noting that when it was left to notice visual characteristics of cancer on its own, without any bias inserted by the researchers, it identified a number that were previously unknown, and could actually help humans identify cancers in the future.

What does this mean for medicine? Well, there have long been studies showing that computers are better at basic correlation finding, and that you might well be better served by having an AI doc to listen to you list your woes. But such robo-docs have always been limited in their ability to interpret test results. Sure, an AI might be able to order an X-Ray, and a nurse might be able to administer the X-Ray, but certainly we’ll always need a doctor to read and interpret the X-Ray?

Perhaps not for much longer. The wonder of machine learning is that it is a highly versatile approach, able to adapt to just about any challenge. Whether you need it to learn the visual characteristics of a broken collarbone or the spoken words associated with an ear infection, there’s not much we can’t credibly imagine an AI being able to master, given access to the right inputs.

Thus, technology seems poised to affect medicine just as it’s affecting many other previously automation-proof professions: It will first break the meta-job of doctor down into individual sub-jobs and slowly computerize everything that doesn’t involve physical labor. This will necessarily make specialists in the remaining physical tasks less elite, and de-emphasize their difficulty and lowering compensation — until the advanced robots arrive to gobble those up, as well.

It’s hard to look at the oncoming wave of automation technologies and think that we’re headed for yet another cycle of the same old historical narrative, where lost jobs are quickly replaced due to higher productivity and newly emerging demands — because, what emerging demands? Only so many people can be employed writing advanced code, or organizing rich people’s dressers.

So if even doctors aren’t safe, who is? One answer is: the frivolous. One thing that makes medicine a desirable target for tech is that the stakes are so high — literally life and death– and performance can be rated along an objective numerical scale. But less vitally important professions might hold on to the idea they’re not able to be done by a computer — though with algorithms now coming to write our news articles, that might just be wishful thinking on our parts.

Leave a Comment

Captcha image