The American company Deep Health has developed an AI that is better at finding early signs of breast cancer than radiologists . When the AI had to examine mammography images, it could identify cancer one or even two years before the radiologists discovered the cancer.
The AI is also good at sorting out images where there is definitely no cancer. In this way, the radiologists do not have to plow through lots of images of healthy breasts.
The idea is now that the AI should be able to function as a fast and skilled assistant to radiologists and give them more time to analyze difficult-to-interpret cases.
To test the AI, the researchers had it and five radiologists examine the same images. The AI was then better at detecting cancer than all the radiologists. In addition, as I said earlier, it could detect abnormalities that later turned out to be cancer, but which the radiologists initially missed.
A challenge for the researchers at Deep Health was to find material that the AI could train on. Pictures are needed where someone has noted if the picture shows someone who has or will get cancer. It was difficult to obtain enough such images to be able to use the usual methods of machine learning.
So the researchers developed a method where the AI manages with fewer training images.
- We have developed a method that imitates the human way of learning by letting the AI train on materials that become increasingly difficult over time. By utilizing information from previous learning, this strategy results in the AI being able to detect cancer with high reliability while at the same time needing less training data, says Bill Lotter CTO at Deep Health.
Deep Health has submitted an application to the FDA to get the AI approved and if it goes as it should, their AI can start helping radiologists prioritize mammography images already this year.