Google AI Is Better At Identifying Breast Cancer Than Human Doctors
Neural networks have helped machines learn some skills previously limited to humans like recognizing objects in a photo or absolutely demolishing you when you fire up StarCraft II. Google is one of the leading developers of neural network AI, and it has developed a new system that uses the power of AI to identify breast cancer in mammograms that doctors might miss.
Currently, mammograms are the best way to diagnose breast cancer, but it’s far from perfect. Trained radiologists can use these tests to catch most cancers, but about one in five slip through the cracks. Additionally, about half of women get at least one false-positive result over any 10-year period. While the new study doesn’t suggest doctors should be cut out of the diagnosis process, it does show that AI can help medical professionals to make the right call more often.
Like all neural networks, Google’s new system needed a lot of training data. It got that in the form of more than 100,000 mammogram images — 76,000 from the UK and 15,000 from the US. All the training data was labeled, allowing the system to learn all the ways tumors can present. With the network established, Google engineers tested its accuracy on a new set of 30,000 mammograms from the US and UK. For each image, it pitted the AI against a human radiologist to see which was more accurate.
When compared with US screening, the AI produced 5.7 percent fewer false positives. That would save women from unnecessary worry and expense to cope with a tumor that’s not present. More importantly, the AI had 9.4 percent fewer false negatives — those are tumors that the radiologist missed (see above). The AI didn’t beat the UK system by as much because the NIH always has two doctors checking each mammogram. There, the AI reduced false positives by 1.2 percent and false negatives by 2.7 percent.
Google also set up a competition between a team of six US radiologists and the AI with a sample of 500 randomly selected mammogram scans. Again, the AI outperformed the humans, identifying several cancers that all six doctors initially missed. However, there was at least one case in which the AI missed a cancer that all the humans spotted.
The study suggests that Google’s AI could help to reduce inaccurate mammogram readings, but the technology needs a lot more testing first. The study primarily used mammogram images captured on a single manufacturer’s devices, and there could be unexpected variation with other equipment.