Google Researchers Trained an Algorithm to Detect Lung Cancer Better Than Radiologists

Gif: Google

Algorithms have screwed up in horrifying, hilarious, and unfortunate ways, so it’s nice when one with the potential to save lives nearly nails it. On Monday, Google AI researchers along with healthcare researchers published research showing that they’ve successfully trained a deep learning algorithm to detect lung cancer with a 94.4 percent success rate.

Advertisement

The findings were published in the journal Nature Medicine on Monday, which indicated that aside from just a high accuracy rate, the algorithm outclassed radiologists under certain circumstances. According to the study, the system achieved that success rate on 6,716 cases from the National Lung Cancer Screening Trial with similar accuracy on 1,139 independent clinical cases.

The researchers conducted two studies—one in which a prior scan was available, and one in which it wasn’t. In the former scenario, the deep learning algorithm—which was trained on computed tomography scans of people with lung cancer, without it, and with nodules turned cancerous, the New York Times reported—had a higher identification rate than six radiologists, and in the latter, the humans and machine were even.

Advertisement

“The whole experimentation process is like a student in school,” Dr. Daniel Tse, a project manager at Google, told the New York Times. “We’re using a large data set for training, giving it lessons and pop quizzes so it can begin to learn for itself what is cancer, and what will or will not be cancer in the future. We gave it a final exam on data it’s never seen after we spent a lot of time training, and the result we saw on final exam — it got an A.”

But the results of this study are very much the baby’s first steps version of algorithmic identification. It’s still far from having proven itself as accurate enough to be deployed pervasively across healthcare institutions providing cancer screenings. It does signal some promise when it comes to automating a process that struggles with false positives and negatives. “Lung CT for smokers, it’s so bad that it’s hard to make it worse,” Dr. Eric Topol, director of the Scripps Research Translational Institute in California, told the New York Times.

A lot of tech companies, including Google, are already leaning on algorithms as a tool for detection on its platforms in a way that’s large-scale, mostly for moderation. And these automated systems are still deeply flawed, leading to mistaken censorship or, worse, a failure to identify violent and hateful content running rampant. But the researchers working on the lung cancer detection technology acknowledge the dangers and risks of unleashing such a system without first more comprehensively confirming that it’s effective—and also ensuring there are checks and balances in place to continuously regulate it and protect it from bad actors.

“We are collaborating with institutions around the world to get a sense of how the technology can be implemented into clinical practice in a productive way,” Dr. Tse told the New York Times. “We don’t want to get ahead of ourselves.”

Advertisement

Share This Story

Get our newsletter

DISCUSSION

This is pretty cool. Radiology involves a lot of pattern recognition, so it’s probably a good place to work in AI. I doubt computers will replace radiologists any time soon - a good radiologist will be looking at so much more than just nodules!

I think there’s room for AI as a tool for radiologists to use, but I think people will have to avoid the urge to farm out more complex tasks to cheaper computers. I’m not a radiologists, but order a lot of MRIs to quantify iron overload in patients who get chronic transfusions. The algorithms to do this were developed based on correlation with the gold standard of liver biopsy. I’ve seen too many cases of an inexperienced radiologist just running the software and getting an iron level that makes no sense clinically. When we show the same scans to a more experience radiologist running the same algorithms, they can tell us that the curve fitting is wrong and make some tweaks to get a result that makes more sense given other markers.

TL/DR: medicine is complex. computers will help, but you still need humans!