A doctor’s work isn’t all done in examination rooms. Many specialists spend lots of time alone with the lights out, examining photographs that reveal their patients’ internal workings.
That might soon change. A paper by Google published in the Journal of the American Medical Association details an algorithm that can detect when someone has developed blindness as a result of diabetes, trained and tested by board-certified ophthalmologists. It shows algorithms can, at least in the case of this particular affliction, make a diagnosis with an accuracy on-par with medical professionals,
A key difference between this research and previous papers on medical imaging by large tech companies is its publication and defense by a respected medical journal like JAMA.
Concurrent with Google’s paper, JAMA also published an article translating the finding for medical professionals and urging the community that this is a good thing—algorithms can let doctors spend more time with patients, rather than reading scans.
It seems likely that these algorithms will reshape specific aspects of these specialties as more algorithms are developed to address a wider range of medical imaging tasks. Because these algorithms are by their nature standardized, repeatable, and scalable, they can be deployed to analyze a large number of images in hospitals around the world once an algorithm has been developed and validated, enabling clinicians to focus on other aspects of their practice.
After writing a lengthy blog post on Nov. 27 detailing how machines can’t yet beat doctors, radiologist Luke Oakden-Rayner saw the Google research and was forced to concede his point.
“I remain convinced that we have yet to see a machine outperform a doctor in any task that is relevant to actual medical practice,” Oakden-Rayner wrote. Sentences later, he continues, “While I was writing this, literally this last paragraph, it became untrue.” (Emphasis his.)
The reason why Google’s work is so good, he writes, is because they paid to create a good set of data—a panel of certified ophthalmologists hand-graded nearly 130,000 images—to teach the algorithm.
Now that the method has been shown effective, others can copy the technique in different domains—radiology seems a likely target—and start the machinations towards eventual clinical adoption. The algorithm itself was not built on any proprietary information— it was first trained on an open-source image set, then refined on the medical images.
There’s still has a long way to go before algorithms are making real diagnoses in hospitals—they’ll have to clear clinical trials. Google says they’re now working with the US Food and Drug Administration (the governing body in the US for this kind of test) and other regulatory agencies to work on building trials. The search giant’s sister company, Alphabet-owned DeepMind, is working closely with the UK’s National Health Service on similar retinal image analysis, and AI’s application in cancer treatment.