The image above - one of around 400,000 mammograms used by Zebra Medical Vision - shows what a breast looks like to AI.
It has been colour-coded to make it easier for a self-teaching neural network to identify breast cancer. Using this technique, Zebra Medical Vision claims to have been able to detect cancerous cells with 91 per cent accuracy. This is an improvement on the typical radiologists' rate of 88 per cent, with fewer false positives. "Right now," says Zebra Medical Vision founder Elad Benjamin, "this is better than human performance." Read more: This AI can spot skin cancer as accurately as a doctor
Launched in 2014, Tel Aviv-based Zebra Medical Vision, which consists of 25 employees, has developed image-scanning algorithms to identify bone, lung, liver and heart disease. It didn't begin to look at breast cancer, however, until it was approached in 2015 by British computer programmer Phil Teare. "His wife had died of cervical cancer at a very young age, so he had a different motivation and he really dived into it," Benjamin recalls.
Using anonymous data from 14 hospitals - including biopsy and pathology records as well as images - Teare began work on an algorithm that could identify malignant cells in mammograms. That meant finding a way to make the scans machine-readable. He approached this problem by using colour to differentiate features. "Phil separated different signals within the image and fed them into the red, blue and green channels of the network." The result is being prepared for US clinical trials in hospitals before submission to the Food and Drug Administration towards the end of 2017.
When the technique becomes commercially available, Benjamin expects it to have a big impact. "In five or seven years, radiologists won't be doing the same job they're doing today," he explains. "They're going to have analytics engines or bots like ours that will be doing 60, 70, 80 per cent of their work."
This article was originally published by WIRED UK
