A new report published in the journal Nature Medicine provides a glimpse into the future of artificial intelligence-assisted medicine, demonstrating an early version of an algorithm capable of accurately identifying early lung cancer on the basis of a CT scan.
The system, developed by several medical centres in collaboration with Google, is the latest salvo in the – often successful – attempts at feeding huge amounts of medical data to artificial neural networks in hopes of training them to outperform human physicians at performing medical diagnosis.
“We have some of the biggest computers in the world,” said co-author on the paper Dr Daniel Tse from Google. “We started wanting to push the boundaries of basic science to find interesting and cool applications to work on.”
In the study, researchers fed the algorithm countless CT scans of people with lung cancer, those without it, and others with suspicious nodules which later turned malignant. A while later, the system was tested for its diagnostic power using scans which it hadn’t seen before.
“The whole experimentation process is like a student in school,” Dr. Tse said. “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.”
Results showed that in cases where no prior scan was available, the algorithm did better than radiologists, but did not outrank its human counterparts in cases where a prior scan was available, which is still mighty impressive.
According to the researchers, the idea here is not to replace human physicians, but to help them do their job even better, as artificial systems trained on massive amounts of data sometimes have a keener ‘eye’ for patterns which the human visual system simply cannot detect.
Before the system can be unleashed on the public, it will have to undergo rigorous testing, peer-review, and pilot studies of the clinical variety to weed out any potential systematic errors, which could harm a great number of people if left undetected.
Luckily, the authors claim to have already started collaborating with research bodies around the world to determine the next steps towards real-world application.